Advances on Broad-Band Wireless Computing, Communication and Applications: Proceedings of the 15th International Conference on Broad-Band and Wireless Computing, Communication and Applications (BWCCA-2020) [1st ed.] 9783030611071, 9783030611088

Table of contents :
Front Matter ....Pages i-xxvii
Performance Evaluation of a Message Relaying Method for Resilient Disaster Networks (Yoshiki Tada, Makoto Ikeda, Leonard Barolli)....Pages 1-10
A Comparison Study of Constriction and Random Inertia Weight Router Replacement Methods for WMNs by WMN-PSOSA-DGA Hybrid Simulation System Considering Chi-square Distribution of Mesh Clients (Admir Barolli, Shinji Sakamoto, Phudit Ampririt, Seiji Ohara, Leonard Barolli, Makoto Takizawa)....Pages 11-21
Multi-source and Multi-target Node Selection in Energy-Efficient Fog Computing Model (Yinzhe Guo, Takumi Saito, Shigenari Nakamura, Tomoya Enokido, Lei Li, Makoto Takizawa)....Pages 22-33
Epidemic and Topic-Based Data Transmission Protocol in a Mobile Fog Computing Model (Takumi Saito, Shigenari Nakamura, Tomoya Enokido, Makoto Takizawa)....Pages 34-43
The Energy-Efficient Object Replication by Excluding Meaningless Methods in Virtual Machine Environments (Tomoya Enokido, Makoto Takizawa)....Pages 44-54
Experiences with a Single-Page Application for Learning Programming (Minoru Uehara)....Pages 55-66
Approach of a Word2Vec Based Tourist Spot Collection Method Considering COVID-19 (Yuki Nagai, Nobuki Saito, Aoto Hirata, Tetsuya Oda, Masaharu Hirota, Kengo Katayama)....Pages 67-75
Detecting Distracted Driving from Images by Processing Relative Locations of Objects of Interest Inside Vehicles (Arup Kanti Dey, Bharti Goel, Sriram Chellappan)....Pages 76-86
Cost and Performance Analysis of Cuckoo Search Based File Replication in MANET (Takeru Kurokawa, Naohiro Hayashibara)....Pages 87-96
A New DTN Relay Method Reducing Number of Transmissions Under Existence of Obstacles by Large-Scale Disaster (Qiang Gao, Tetsuya Shigeyasu)....Pages 97-107
Performance Comparison of Multi-class SVM with Oversampling Methods for Imbalanced Data Classification (Seunghyun Park, Hyunhee Park)....Pages 108-119
Message Transmission Scheduling for Multi-hop Wireless Sensor Network with T-Shaped Topology (Linh Vu Nguyen, Masahiro Shibata, Masato Tsuru)....Pages 120-130
Performance Evaluation of Improved V2X Wireless Communication Based on Gigabit WLAN (Akira Sakuraba, Goshi Sato, Noriki Uchida, Yoshitaka Shibata)....Pages 131-142
Improvement of Dental Treatment Training System Using a Haptic Device (Masaki Nomi, Yoshihiro Okada)....Pages 143-153
A Proposal of Air-Conditioning Guidance System Using Discomfort Index (Samsul Huda, Nobuo Funabiki, Minoru Kuribayashi, Rahardhita Widyatra Sudibyo, Nobuya Ishihara, Wen-Chun Kao)....Pages 154-165
An Efficient Content Sharing Using Dynamic Fog Considering Transition of Number of Mobile Terminals in a City (Takuya Itokazu, Shinji Sugawara)....Pages 166-175
Oversampling for Detection of Malicious JavaScript in Realistic Environment (Phung Minh Ngoc, Mamoru Mimura)....Pages 176-187
DTN Routing Protocol Using Reinforcement Learning (Kenta Henmi, Akio Koyama)....Pages 188-198
Data Fusion Protocols for Cloud Infrastructures (Lidia Ogiela, Makoto Takizawa, Urszula Ogiela)....Pages 199-203
Implementation of Process Migration Method for PC-FPGA Hybrid System (Keisuke Takano, Tetsuya Oda, Ryo Ozaki, Akira Uejima, Masaki Kohata)....Pages 204-210
Speeding-Up of Construction Algorithms for the Graph Coloring Problem (Kazuho Kanahara, Kengo Katayama, Takafumi Miyake, Etsuji Tomita)....Pages 211-222
An On-Board Equipment and Blockchain-Based Automobile Insurance and Maintenance Platform (Wen-Yao Lin, Frank Yeong-Sung Lin, Ting-Huan Wu, Kuang-Yen Tai)....Pages 223-232
An Integrated Fuzzy-Based Simulation System for Driver Risk Management in VANETs Considering Relative Humidity as a New Parameter (Kevin Bylykbashi, Ermioni Qafzezi, Makoto Ikeda, Keita Matsuo, Leonard Barolli, Makoto Takizawa)....Pages 233-243
IoT Device Power Management Based on PSM and EDRX Mechanisms (Kun-Lin Tsai, Fang-Yie Leu, Tz-Yuan Huang, Hao-En Yang)....Pages 244-253
Combining Agile with Traditional Software Development for Improvement Maintenance Efficiency and Quality (Sen-Tarng Lai, Fang-Yie Leu)....Pages 254-264
On Text Tiling for Documents: A Neural-Network Approach (Siang Yun Yoong, Yao-Chung Fan, Fang-Yie Leu)....Pages 265-274
A High Sensing Accuracy Mechanism for Wireless Sensor Networks (Li-Ling Hung, Fang-Yie Leu)....Pages 275-283
A Novel Scheme of Schnorr Multi-signatures for Multiple Messages with Key Aggregation (Rikuhiro Kojima, Dai Yamamoto, Takeshi Shimoyama, Kouichi Yasaki, Kazuaki Nimura)....Pages 284-295
A Fuzzy-Based Approach for Transmission Control of Sensory Data in Resilient Wireless Sensor Networks During Disaster Situation (Daisuke Nishii, Makoto Ikeda, Leonard Barolli)....Pages 296-303
Parasitic Coil Effects on Communication Performance of Table Type 13.56 MHz RFID Reader: A Comparison Study for Different Coil Turns (Yuki Yoshigai, Kiyotaka Fujisaki)....Pages 304-312
Tuning of Output Optical Signal Wavelength Through Resonant Filter for WDM System (Hiroshi Maeda)....Pages 313-320
Design and Implementation of a DQN Based AAV (Nobuki Saito, Tetsuya Oda, Aoto Hirata, Yuto Hirota, Masaharu Hirota, Kengo Katayama)....Pages 321-329
A Dynamic Tree-Based Fog Computing (DTBFC) Model for the Energy-Efficient IoT (Keigo Mukae, Takumi Saito, Shigenari Nakamura, Tomoya Enokido, Makoto Takizawa)....Pages 330-340
An Energy-Efficient Algorithm for Virtual Machines to Migrate Considering Migration Time (Naomichi Noaki, Takumi Saito, Dilawaer Duolikun, Tomoya Enokido, Makoto Takizawa)....Pages 341-354
A Coverage Construction Method Based Hill Climbing Approach for Mesh Router Placement Optimization (Aoto Hirata, Tetsuya Oda, Nobuki Saito, Masaharu Hirota, Kengo Katayama)....Pages 355-364
Review of Intelligent Data Analysis and Data Visualization (Kang Xie, Linshan Han, Maohua Jing, Jingmin Luan, Tao Yang, Rourong Fan)....Pages 365-375
Data Analysis Based on Knowledge Graph (Kang Xie, Qizhen Jia, Maohua Jing, Qilong Yu, Tao Yang, Rourong Fan)....Pages 376-385
Integration of Software-Defined Network and Fuzzy Logic Approaches for Admission Control in 5G Wireless Networks: A Fuzzy-Based Scheme for QoS Evaluation (Phudit Ampririt, Seiji Ohara, Ermioni Qafzezi, Makoto Ikeda, Leonard Barolli, Makoto Takizawa)....Pages 386-396
ICS Testbed Implementation Considering Dataset Collection Environment (Eunseon Jeong, Junyoung Park, Minseong Kim, Chanmin Kim, Soyoung Jung, Kangbin Yim)....Pages 397-406
A Study on Reducing Interest Misleading by Publisher Migration on Mobile Networks (Taichi Iwamoto, Tetsuya Shigeyasu)....Pages 407-415
Cyber Attack Scenarios in Cooperative Automated Driving (Insu Oh, Eunseon Jeong, Junyoung Park, Taeyoung Jeong, Junghoon Park, Kangbin Yim)....Pages 416-425
Implementation of a User Finger Movement Capturing Device for Control of Self-standing Omnidirectional Robot (Kenshiro Mitsugi, Keita Matsuo, Leonard Barolli)....Pages 426-435
Implementation of Control Interfaces for Moving Omnidirectional Access Point Robot (Atushi Toyama, Kenshiro Mitsugi, Keita Matsuo, Leonard Barolli)....Pages 436-443
Proposal and Experimental Results of an Ambient Intelligence for Training on Soldering Iron Holding (Yuto Hirota, Tetsuya Oda, Nobuki Saito, Aoto Hirata, Masaharu Hirota, Kengo Katatama)....Pages 444-453
Design of Education Tool for Reinforcement-Learning Agent Developers (Takahiro Uchiya, Kodai Shimano, Ichi Takumi)....Pages 454-462
Back Matter ....Pages 463-464

Citation preview

Lecture Notes in Networks and Systems 159

Leonard Barolli · Makoto Takizawa · Tomoya Enokido · Hsing-Chung Chen · Keita Matsuo   Editors

Advances on Broad-Band Wireless Computing, Communication and Applications Proceedings of the 15th International Conference on Broad-Band and Wireless Computing, Communication and Applications (BWCCA-2020)

Lecture Notes in Networks and Systems Volume 159

Series Editor Janusz Kacprzyk, Systems Research Institute, Polish Academy of Sciences, Warsaw, Poland Advisory Editors Fernando Gomide, Department of Computer Engineering and Automation—DCA, School of Electrical and Computer Engineering—FEEC, University of Campinas— UNICAMP, São Paulo, Brazil Okyay Kaynak, Department of Electrical and Electronic Engineering, Bogazici University, Istanbul, Turkey Derong Liu, Department of Electrical and Computer Engineering, University of Illinois at Chicago, Chicago, USA; Institute of Automation, Chinese Academy of Sciences, Beijing, China Witold Pedrycz, Department of Electrical and Computer Engineering, University of Alberta, Alberta, Canada; Systems Research Institute, Polish Academy of Sciences, Warsaw, Poland Marios M. Polycarpou, Department of Electrical and Computer Engineering, KIOS Research Center for Intelligent Systems and Networks, University of Cyprus, Nicosia, Cyprus Imre J. Rudas, Óbuda University, Budapest, Hungary Jun Wang, Department of Computer Science, City University of Hong Kong, Kowloon, Hong Kong

The series “Lecture Notes in Networks and Systems” publishes the latest developments in Networks and Systems—quickly, informally and with high quality. Original research reported in proceedings and post-proceedings represents the core of LNNS. Volumes published in LNNS embrace all aspects and subfields of, as well as new challenges in, Networks and Systems. The series contains proceedings and edited volumes in systems and networks, spanning the areas of Cyber-Physical Systems, Autonomous Systems, Sensor Networks, Control Systems, Energy Systems, Automotive Systems, Biological Systems, Vehicular Networking and Connected Vehicles, Aerospace Systems, Automation, Manufacturing, Smart Grids, Nonlinear Systems, Power Systems, Robotics, Social Systems, Economic Systems and other. Of particular value to both the contributors and the readership are the short publication timeframe and the world-wide distribution and exposure which enable both a wide and rapid dissemination of research output. The series covers the theory, applications, and perspectives on the state of the art and future developments relevant to systems and networks, decision making, control, complex processes and related areas, as embedded in the fields of interdisciplinary and applied sciences, engineering, computer science, physics, economics, social, and life sciences, as well as the paradigms and methodologies behind them. ** Indexing: The books of this series are submitted to ISI Proceedings, SCOPUS, Google Scholar and Springerlink **

More information about this series at http://www.springer.com/series/15179

Leonard Barolli Makoto Takizawa Tomoya Enokido Hsing-Chung Chen Keita Matsuo •

•

•

•

Editors

Advances on Broad-Band Wireless Computing, Communication and Applications Proceedings of the 15th International Conference on Broad-Band and Wireless Computing, Communication and Applications (BWCCA-2020)

123

Editors Leonard Barolli Department of Information and Communication Engineering, Faculty of Information Engineering Fukuoka Institute of Technology Fukuoka, Japan Tomoya Enokido Faculty of Business Administration Rissho University Tokyo, Japan

Makoto Takizawa Department of Advanced Sciences, Faculty of Science and Engineering Hosei University Tokyo, Japan Hsing-Chung Chen Department of Computer Science and Information Engineering Asia University Taichung, Taiwan

Keita Matsuo Department of Information and Communication Engineering, Faculty of Information Engineering Fukuoka Institute of Technology Fukuoka, Japan

ISSN 2367-3370 ISSN 2367-3389 (electronic) Lecture Notes in Networks and Systems ISBN 978-3-030-61107-1 ISBN 978-3-030-61108-8 (eBook) https://doi.org/10.1007/978-3-030-61108-8 © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, expressed or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. This Springer imprint is published by the registered company Springer Nature Switzerland AG The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland

Welcome Message of BWCCA-2020 International Conference Organizers

Welcome to the 15th International Conference on Broadband and Wireless Computing, Communication and Applications (BWCCA-2020), which will be held in conjunction with the 15th 3PGCIC-2020 International Conference from October 28 to October 30, 2020 in Yonago City, Tottori Prefecture, Japan. This International Conference is a forum for sharing ideas and research work in the emerging areas of broadband and wireless computing. Information networks of today are going through a rapid evolution. Different kinds of networks with different characteristics are emerging and they are integrating in heterogeneous networks. For these reasons, there are many interconnection problems which may occur at different levels of the hardware and software design of communicating entities and communication networks. These kinds of networks need to manage an increasing usage demand, provide support for a significant number of services, guarantee their QoS, and optimize the network resources. The success of all-IP networking and wireless technology has changed the ways of living the people around the world. The progress of electronic integration and wireless communications is going to pave the way to offer people the access to the wireless networks on the fly, based on which all electronic devices will be able to exchange the information with each other in ubiquitous way whenever necessary. The aim of this conference is to present the innovative research and technologies as well as developments related to broadband networking, and mobile and wireless communications. The organization of an International Conference requires the support and help of many people. A lot of people have helped and worked hard to produce a successful BWCCA-2020 technical program and conference proceedings. First, we would like to thank all authors for submitting their papers, program committee members and reviewers who carried out the most difficult work by carefully evaluating the submitted papers. We thank Web Administrators Co-Chairs and Finance Chair for their excellent work. We would like to express our gratitude to Prof. Makoto Takizawa, Hosei University, Japan, as Honorary Chair of BWCCA-2020 for his support and

v

vi

Welcome Message of BWCCA-2020 International Conference Organizers

help. We give special thanks to Keynote Speakers of BWCCA-2020 and local arrangement team. We hope you will enjoy the conference and have a great time in Yonago City, Japan. Leonard Barolli BWCCA-2020 Steering Committee Chair Tomoya Enokido Farookh Hussain Hsing-Chung Chen BWCCA-2020 General Co-chairs Naohiro Hayashibara Lidia Ogiela Kangbin Yim BWCCA-2020 Program Committee Co-chairs

BWCCA-2020 Organizing Committee

Honorary Chair Makoto Takizawa

Hosei University, Japan

General Co-chairs Tomoya Enokido Farookh Hussain Hsing-Chung Chen

Rissho University, Japan University of Technology Sydney, Australia Asia University, Taiwan

Program Committee Co-chairs Naohiro Hayashibara Lidia Ogiela Kangbin Yim

Kyoto Sangyo University, Japan Pedagogical University of Krakow, Poland SCH University, South Korea

Workshops Co-chairs Keita Matsuo Fang-Yie Leu Tetsuya Shigeyasu

Fukuoka Institute of Technology, Japan Tunghai University, Taiwan Prefectural University of Hiroshima, Japan

Finance Chair Makoto Ikeda

FIT, Japan

vii

viii

BWCCA-2020 Organizing Committee

Web Administrator Co-chairs Kevin Bylykbashi Phudit Ampririt Seiji Ohara Ermioni Qafzezi

Fukuoka Fukuoka Fukuoka Fukuoka

Institute Institute Institute Institute

of of of of

Technology, Technology, Technology, Technology,

Japan Japan Japan Japan

Local Organizing Co-chairs Elis Kulla Akimitsu Kanzaki

Okayama University of Science, Japan Shimane University, Japan

Steering Committee Chair Leonard Barolli

Fukuoka Institute of Technology, Japan

Track Areas Next-Generation Wireless Networks Track Co-chairs Bhed Bista Szu-Yin Lin Sriram Chellappan

Iwate Prefectural University, Japan Chung Yuan Christian University, Taiwan University of South Florida, USA

PC Members Jiahong Wang Shigetomo Kimura Chotipat Pornavalai Danda B. Rawat Gongjun Yan Vamsi Paruchuri Arjan Durresi Shih-Yi James Chien Pei-Ju Lee Chih-Hao Lin Hao-Hsiang Ku Jung-Bin Li Thoshitha Gamage Mukundan Sridharan Brijesh Chejerla Srinivas Chakravarthi Thandu

Iwate Prefectural University, Japan University of Tsukuba, Japan King Mongkut’s Institute of Technology Ladkrabang, Thailand Howard University, USA University of Southern Indiana, USA University of Central Arkansas, USA IUPUI, USA National Sun Yat-sen University, Taiwan National Chung Cheng University, Taiwan Chung Yuan Christian University, Taiwan National Taiwan Ocean University, Taiwan Fu Jen Catholic University, Taiwan Southern Illinois University, USA Samraksh Company, USA Florida Blue, USA Amazon, USA

BWCCA-2020 Organizing Committee

ix

Cloud and Service Computing Track Co-chairs Hwamin Lee Ramesh C. Hansdah Baojiang Cui

Soonchunhyang University, Korea Indian Institute of Science, Bangalore, India Beijing University of Posts and Telecommunications, China

PC Members Gang Wang Jianxin Wang Jie Cheng Shaoyin Cheng Yan Zhang Willy Susilo Kamil Kluczniak Francesco Palmieri Jian Shen Jin Li Fangguo Zhang Xinyi Huang Shengli Liu Zhenjie Huang Joseph K. Liu Yong Yu Ding Wang Tao Jiang Jianfeng Wang S. D. Madhu Kumar Ashutosh Bhatia Amulya Rathna Swain Yogesh Simmhan Soumya K. Ghosh

Nankai University, China Beijing Forestry University, China Shandong University, China University of Science and Technology of China, China Hubei University, China University of Wollongong, Australia Wroclaw University of Technology, Poland University of Salerno, Italy Nanjing University of Information Science and Technology, China Guangzhou University, China Sun Yat-sen University, China Fujian Normal University China Shanghai Jiaotong University China Zhangzhou City University China Institute for Infocomm Research, Australia University of Wollongong, China Peking University, China Xidian University, China Xidian University, China NIT Calicut, India BITS Pilani, Pilani Campus, India KIIT, Bhubaneshwar, India IISc Bangalore, India Indian Institute of Technology, India

Multimedia and Web Applications Track Co-chairs Yoshihiro Okada Chuan-Yu Chang Salem Alkhalaf

Kyushu University, Japan National Yunlin University of Science and Technology, Taiwan Qassim University, Saudi Arabia

x

BWCCA-2020 Organizing Committee

PC Members Kaoru Sugita Tomoyuki Ishida Makoto Nakashima Nobukazu Iguchi Kenzi Watanabe Nobuo Funabiki Shinji Sugawara Li-Wei Kang Chia-Hung Yeh Jun-Wei Hsieh Wu-Chih Hu Chien-Cheng Lee Muhammad Hussain Umair Azfar Khan Shigeru Takano Kosuke Kaneko Akira Haga Wei Shi

Fukuoka Institute of Technology, Japan Fukuoka Institute of Technology, Japan Oita University, Japan Kinki University, Japan Hiroshima University, Japan Okayama University, Japan Chiba Institute of Technology, Japan National Yunlin University of Science and Technology, Taiwan National Taiwan Normal University, Taiwan National Taiwan Ocean University, Taiwan National Penghu University of Science and Technology, Taiwan Yuan-Ze University, Taiwan King Saud University, Saudi Arabia Habib University, Pakistan Kyushu University, Japan Kyushu University, Japan Kyushu University, Japan Kyushu University, Japan

Security and Privacy Track Co-chairs Tianhan Gao Masakatsu Nishigaki Mohamed Abdur Rahman

Northeastern University, China Shizuoka University, Japan Prince Mughrin University, Saudi Arabia

PC Members Nan Guo Zhenhua Tan Jian Xu Hiroaki Kikuchi Takamichi Saito Rashid Tahir Syed Sadiq Md. Mamunur Rashid (Mamun) Akhlaq Ahmad

Northeastern University, China Northeastern University, China Northeastern University, China Meiji University, Japan Meiji University, Japan University of Prince Mugrin Madinah, Saudi Arabia University of Prince Mugrin Madinah, Saudi Arabia King’s Business School, UK Umm Al Qura University Makkah, Saudi Arabia

BWCCA-2020 Organizing Committee

Shyhtsun Felix Wu Zhen-Yu Wu Tsung-Chih Hsiao Kuo-Kun Tseng Akira Otsuka Naonobu Okazaki Masaki Shimaoka

xi

University of California, Davis, USA Penghu University of Science and Technology, Taiwan Southeast University, China Harbin Institute of Technology, China Institute of Information Security, Japan University of Miyazaki, Japan Secom Co., Ltd., Japan

Network Protocols and Performance Analysis Track Co-chairs Tetsuya Shigeyasu Ching-Feng Liang Vamsi Paruchuri

Prefectural University of Hiroshima, Japan Industrial Technology Research Institute, Taiwan University of Central Arkansas, USA

PC Members Xiaoyi Wang Yu Sun Qiang Duan Han-Chieh Wei Masaaki Yamanaka Misako Urakami Tomoya Kawakami Masaaki Noro Nobuyoshi Sato Phone Lin Ray-Guang Cheng Shun-Ren Yang

Nokia Solutions and Networks, USA University of Central Arkansas, USA Pennsylvania State University, USA Dallas Baptist University, USA Japan Coast Guard Academy, Japan Tokuyama College of Technology, Japan Nara Institute of Science and Technology, Japan Fujitsu Corp., Japan Iwate Prefectural University, Japan National Taiwan University, Taiwan National Taiwan University of Science and Technology, Taiwan National Tsing Hua University, TaiwanWhai-En Chen, National ILan University, Taiwan

Intelligent and Cognitive Computing Chairs Lidia Ogiela Takahiro Uchiya Hai Dong

Pedagogical University of Krakow, Poland Nagoya Institute of Technology, Japan RMIT University, Australia

xii

BWCCA-2020 Organizing Committee

PC Members Atsuko Mutoh Shinsuke Kajioka Ryota Nishimura Shohei Kato Francesco Pascale Jan Platoš Pavel Krömer Urszula Ogiela Jana Nowaková Chang Hoon Ko Hae-Duck Joshua Jeong Pengcheng Zhang Sajib Mistry Tooba Aamir Wei Du Wei Zhang Shang-Pin Ma

Nagoya Institute of Technology, Japan Nagoya Institute of Technology, Japan Tokushima University, Japan Nagoya Institute of Technology, Japan University of Salerno, Italy VŠB Technical University of Ostrava, Czech Republic VŠB Technical University of Ostrava, Czech Republic Pedagogical University of Krakow, Poland VŠB Technical University of Ostrava, Czech Republic Choi, Chosun University, Korea Chosun University, Korea Korean Bible University, Korea Hohai University, China University of Sydney, Australia RMIT University, Australia Wuhan University of Technology, China Macquarie University, Australia National Taiwan Ocean University, Taiwan

Distributed and Parallel Computing Track Co-chairs Naohiro Hayashibara Omar Khadeer Hussain

Kyoto Sangyo University, Japan University of New South Wales (UNSW), Australia

PC Members Sazia Parvin Naeem Janjua Alireza Faed Adil Hammadi Lucian Prodan Kanwalinderjit Kaur Gagneja Rohaya Latip Tomoya Enokido Makoto Takizawa Leonard Barolli Akio Koyama Minoru Uehara

Melbourne Polytechnic, Australia Edith Cowan University, Australia Ryerson University, Canada Curtin University, Australia Polytechnic University Timisoara, Romania Florida Polytechnic University, USA Universiti Putra Malaysia, Malaysia Rissho University, Japan Hosei University, Japan Fukuoka Institute of Technology, Japan Yamagata University, Japan Toyo University, Japan

BWCCA-2020 Organizing Committee

xiii

IoT and Smart Environment Track Co-chairs Nadeem Javaid Chun-Wei Tsai

COMSATS University Islamabad, Pakistan National Chung Hsing University, Taiwan

PC Members Zahoor Ali Khan Umar Qasim Farookh Hussain Elis Kulla Keita Matsuo Hsin-Hung Cho Fan-Hsun Tseng Hsin-Te Wu

Higher Colleges of Technology, UAE University of Alberta, Canada University Technology Sydney, Australia Okayama University of Science, Japan Fukuoka Institute of Technology, Japan National Ilan University, Taiwan National Taiwan Normal University, Taiwan National Penghu University of Science and Technology, Taiwan

Database, Data Mining and Big Data Track Co-chairs Antonio Esposito Yao-Chung Fan Morteza Saberi

University of Campania Luigi Vanvitelli, Italy National Chung Hsing University, Taiwan University of New South Wales, Australia

PC Members Mehran Samavati Farshid Hajati Jinnie Hee Yoon Elena Sitnikova Chen-Yi Lin Lun-Chi Chen Huan Chen Luca Tasquier Stefania Nacchia Salvatore Augusto Maisto Salvatore D’Angelo

University of Sydney, Australia Griffith University, Australia Sejong University, Korea UNSW, Australia National Taichung University of Science and Technology, Taiwan National Center for High-performance Computing (NCHC), Taiwan National Chung Hsing University, Taiwan University of Campania Luigi Vanvitelli, Italy University of Campania Luigi Vanvitelli, Italy University of Campania Luigi Vanvitelli, Italy University of Campania Luigi Vanvitelli, Italy

xiv

BWCCA-2020 Organizing Committee

Ubiquitous and Pervasive Computing Track Co-chairs Isaac Woungang Asm Kayes Chyi-Ren Dow

Ryerson University, Canada La Trobe University, Australia Feng Chia University, Taiwan

PC Members Evjola Spaho Makoto Ikeda Elis Kulla Admir Barolli Donald Elmazi Alan Colman Iqbal H. Sarker Eric Pardede Syed Mahbub Patrick Hung Pei-Chun Lin Zhang Kejun Duc-Binh Nguyen

Wei Lu Luca Caviglione Hamed Aly Danda B. Rawat Marcelo Luis Brocardo Glaucio Carvalho

Polytechnic University of Tirana, Albania Fukuoka Institute of Technology, Japan Okayama University of Science, Japan Aleksander Moisiu University of Durres, Albania Fukuoka Institute of Technology, Japan Swinburne University of Technology, Australia Swinburne University of Technology, Australia La Trobe University, Australia La Trobe University, Australia The University of Ontario Institute of Technology, Canada Feng Chia University, Taiwan ZheJiang University, China Thai Nguyen University of Information and Communications Technology (ICTU), Vietnam Keene State College, USA CNIT, Italy Acadia University, Canada Howard University, USA University of Victoria, Canada Ryerson University, Canada

BWCCA-2020 Reviewers Barolli Admir Barolli Leonard Bista Bhed Caballé Santi Chellappan Sriram Chen Hsing-Chung Cui Baojiang Di Martino Beniamino Durresi Arjan Enokido Tomoya

Ficco Massimo Fun Li Kin Funabiki Nobuo Gao Tianhan Gotoh Yusuke Hachaj Tomasz Hussain Farookh Hussain Omar Javaid Nadeem Jeong Joshua

BWCCA-2020 Organizing Committee

Ikeda Makoto Ishida Tomoyuki Izu Tetsuya Kanzaki Akimitsu Kayes Asm Kikuchi Hiroaki Koyama Akio Kulla Elis Lee Kyungroul Leu Fang-Yie Matsuo Keita Moore Philip Koyama Akio Kryvinska Natalia Nishigaki Masakatsu Ogiela Lidia Ogiela Marek Okada Yoshihiro Paruchuri Vamsi Krishna

xv

Rahayu Wenny Rawat Danda Sakamoto Shinji Shibata Yoshitaka Shigeyasu Tetsuya Sato Fumiaki Saito Takamichi Sugawara Shinji Takizawa Makoto Taniar David Uehara Minoru Venticinque Salvatore Vitabile Salvatore Waluyo Agustinus Borgy Wang Xu An Watanabe Koki Woungang Isaac Xhafa Fatos Yim Kangbin

BWCCA-2020 Keynote Talks

Fairness and Efficiency in Network Resource Sharing Masato Tsuru Kyushu Institute of Technology, Japan

Abstract. With the expansion of network users and applications, the network traffic is still growing and a better sharing of limited network resources among multiple users/applications is required. In particular, recent strong demand on Internet of things (IoT) for smart and connected communities along with architectural advancement, such as software-defined networking (SDN) and multiaccess edge computing (MEC), has posed new challenges in fair and efficient resource sharing by multiplexing with complex and heterogeneous settings. In this talk, after briefly reviewing recent trends in communication networks, we discuss the concept of fairness in terms of achieved performance of each user through simple examples in wireless and wired networks. Then, we go into more details in few examples (Multipath-multicast file transfer on OpenFlow network; Wireless shared channel scheduling) and see how a fair and efficient resource sharing can be realized by time-division, space-division and information-coding multiplexing.

xix

Road Status Sensing and V2X Technologies toward Autonomous Driving on Challenged Network Environment Yoshitaka Shibata Iwate Prefectural University, Morioka, Japan

Abstract. Autonomous driving systems are expected as future safe and effective vehicles and have been investigated and developed in industrial countries and actually driving on the exclusive and highway roads with flat surface, clear driving lanes and center lines separated from the opposite direction and on good weather conditions. In the future autonomous driving systems, more general road status and weather status environments such as heavy snow countries in addition to challenged network environment where no public communication network is available must be considered to realize safer and reliable mobility infrastructure. In this talk, in order to resolve the above problems, IoT-based crowd sensing technology using various environmental sensors to precisely identify qualitative and quantitative road status using AI technology is discussed. The next-generation V2X communication technology to exchange and share those road status and GIS information among surrounding vehicles and roadside bases stations is also explained. Finally, a wide road status information sharing platform for challenged weather and network environments based on the 5G and the next-generation high speed LAN is introduced.

xxi

Contents

Performance Evaluation of a Message Relaying Method for Resilient Disaster Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yoshiki Tada, Makoto Ikeda, and Leonard Barolli A Comparison Study of Constriction and Random Inertia Weight Router Replacement Methods for WMNs by WMN-PSOSA-DGA Hybrid Simulation System Considering Chi-square Distribution of Mesh Clients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Admir Barolli, Shinji Sakamoto, Phudit Ampririt, Seiji Ohara, Leonard Barolli, and Makoto Takizawa Multi-source and Multi-target Node Selection in Energy-Efficient Fog Computing Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yinzhe Guo, Takumi Saito, Shigenari Nakamura, Tomoya Enokido, Lei Li, and Makoto Takizawa Epidemic and Topic-Based Data Transmission Protocol in a Mobile Fog Computing Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . Takumi Saito, Shigenari Nakamura, Tomoya Enokido, and Makoto Takizawa

1

11

22

34

The Energy-Efficient Object Replication by Excluding Meaningless Methods in Virtual Machine Environments . . . . . . . . . . . . . . . . . . . . . . Tomoya Enokido and Makoto Takizawa

44

Experiences with a Single-Page Application for Learning Programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Minoru Uehara

55

Approach of a Word2Vec Based Tourist Spot Collection Method Considering COVID-19 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yuki Nagai, Nobuki Saito, Aoto Hirata, Tetsuya Oda, Masaharu Hirota, and Kengo Katayama

67

xxiii

xxiv

Contents

Detecting Distracted Driving from Images by Processing Relative Locations of Objects of Interest Inside Vehicles . . . . . . . . . . . . . . . . . . . Arup Kanti Dey, Bharti Goel, and Sriram Chellappan

76

Cost and Performance Analysis of Cuckoo Search Based File Replication in MANET . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Takeru Kurokawa and Naohiro Hayashibara

87

A New DTN Relay Method Reducing Number of Transmissions Under Existence of Obstacles by Large-Scale Disaster . . . . . . . . . . . . . . Qiang Gao and Tetsuya Shigeyasu

97

Performance Comparison of Multi-class SVM with Oversampling Methods for Imbalanced Data Classification . . . . . . . . . . . . . . . . . . . . . 108 Seunghyun Park and Hyunhee Park Message Transmission Scheduling for Multi-hop Wireless Sensor Network with T-Shaped Topology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120 Linh Vu Nguyen, Masahiro Shibata, and Masato Tsuru Performance Evaluation of Improved V2X Wireless Communication Based on Gigabit WLAN . . . . . . . . . . . . . . . . . . . . . . . 131 Akira Sakuraba, Goshi Sato, Noriki Uchida, and Yoshitaka Shibata Improvement of Dental Treatment Training System Using a Haptic Device . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143 Masaki Nomi and Yoshihiro Okada A Proposal of Air-Conditioning Guidance System Using Discomfort Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154 Samsul Huda, Nobuo Funabiki, Minoru Kuribayashi, Rahardhita Widyatra Sudibyo, Nobuya Ishihara, and Wen-Chun Kao An Efficient Content Sharing Using Dynamic Fog Considering Transition of Number of Mobile Terminals in a City . . . . . . . . . . . . . . . 166 Takuya Itokazu and Shinji Sugawara Oversampling for Detection of Malicious JavaScript in Realistic Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176 Phung Minh Ngoc and Mamoru Mimura DTN Routing Protocol Using Reinforcement Learning . . . . . . . . . . . . . 188 Kenta Henmi and Akio Koyama Data Fusion Protocols for Cloud Infrastructures . . . . . . . . . . . . . . . . . . 199 Lidia Ogiela, Makoto Takizawa, and Urszula Ogiela Implementation of Process Migration Method for PC-FPGA Hybrid System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204 Keisuke Takano, Tetsuya Oda, Ryo Ozaki, Akira Uejima, and Masaki Kohata

Contents

xxv

Speeding-Up of Construction Algorithms for the Graph Coloring Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211 Kazuho Kanahara, Kengo Katayama, Takafumi Miyake, and Etsuji Tomita An On-Board Equipment and Blockchain-Based Automobile Insurance and Maintenance Platform . . . . . . . . . . . . . . . . . . . . . . . . . . . 223 Wen-Yao Lin, Frank Yeong-Sung Lin, Ting-Huan Wu, and Kuang-Yen Tai An Integrated Fuzzy-Based Simulation System for Driver Risk Management in VANETs Considering Relative Humidity as a New Parameter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233 Kevin Bylykbashi, Ermioni Qafzezi, Makoto Ikeda, Keita Matsuo, Leonard Barolli, and Makoto Takizawa IoT Device Power Management Based on PSM and EDRX Mechanisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244 Kun-Lin Tsai, Fang-Yie Leu, Tz-Yuan Huang, and Hao-En Yang Combining Agile with Traditional Software Development for Improvement Maintenance Efficiency and Quality . . . . . . . . . . . . . . 254 Sen-Tarng Lai and Fang-Yie Leu On Text Tiling for Documents: A Neural-Network Approach . . . . . . . . 265 Siang Yun Yoong, Yao-Chung Fan, and Fang-Yie Leu A High Sensing Accuracy Mechanism for Wireless Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275 Li-Ling Hung and Fang-Yie Leu A Novel Scheme of Schnorr Multi-signatures for Multiple Messages with Key Aggregation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 284 Rikuhiro Kojima, Dai Yamamoto, Takeshi Shimoyama, Kouichi Yasaki, and Kazuaki Nimura A Fuzzy-Based Approach for Transmission Control of Sensory Data in Resilient Wireless Sensor Networks During Disaster Situation . . . . . . 296 Daisuke Nishii, Makoto Ikeda, and Leonard Barolli Parasitic Coil Effects on Communication Performance of Table Type 13.56 MHz RFID Reader: A Comparison Study for Different Coil Turns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304 Yuki Yoshigai and Kiyotaka Fujisaki Tuning of Output Optical Signal Wavelength Through Resonant Filter for WDM System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313 Hiroshi Maeda

xxvi

Contents

Design and Implementation of a DQN Based AAV . . . . . . . . . . . . . . . . 321 Nobuki Saito, Tetsuya Oda, Aoto Hirata, Yuto Hirota, Masaharu Hirota, and Kengo Katayama A Dynamic Tree-Based Fog Computing (DTBFC) Model for the Energy-Efficient IoT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 330 Keigo Mukae, Takumi Saito, Shigenari Nakamura, Tomoya Enokido, and Makoto Takizawa An Energy-Efficient Algorithm for Virtual Machines to Migrate Considering Migration Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341 Naomichi Noaki, Takumi Saito, Dilawaer Duolikun, Tomoya Enokido, and Makoto Takizawa A Coverage Construction Method Based Hill Climbing Approach for Mesh Router Placement Optimization . . . . . . . . . . . . . . . . . . . . . . . 355 Aoto Hirata, Tetsuya Oda, Nobuki Saito, Masaharu Hirota, and Kengo Katayama Review of Intelligent Data Analysis and Data Visualization . . . . . . . . . . 365 Kang Xie, Linshan Han, Maohua Jing, Jingmin Luan, Tao Yang, and Rourong Fan Data Analysis Based on Knowledge Graph . . . . . . . . . . . . . . . . . . . . . . 376 Kang Xie, Qizhen Jia, Maohua Jing, Qilong Yu, Tao Yang, and Rourong Fan Integration of Software-Defined Network and Fuzzy Logic Approaches for Admission Control in 5G Wireless Networks: A Fuzzy-Based Scheme for QoS Evaluation . . . . . . . . . . . . . . . . . . . . . . 386 Phudit Ampririt, Seiji Ohara, Ermioni Qafzezi, Makoto Ikeda, Leonard Barolli, and Makoto Takizawa ICS Testbed Implementation Considering Dataset Collection Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 397 Eunseon Jeong, Junyoung Park, Minseong Kim, Chanmin Kim, Soyoung Jung, and Kangbin Yim A Study on Reducing Interest Misleading by Publisher Migration on Mobile Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 407 Taichi Iwamoto and Tetsuya Shigeyasu Cyber Attack Scenarios in Cooperative Automated Driving . . . . . . . . . 416 Insu Oh, Eunseon Jeong, Junyoung Park, Taeyoung Jeong, Junghoon Park, and Kangbin Yim Implementation of a User Finger Movement Capturing Device for Control of Self-standing Omnidirectional Robot . . . . . . . . . . . . . . . . 426 Kenshiro Mitsugi, Keita Matsuo, and Leonard Barolli

Contents

xxvii

Implementation of Control Interfaces for Moving Omnidirectional Access Point Robot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 436 Atushi Toyama, Kenshiro Mitsugi, Keita Matsuo, and Leonard Barolli Proposal and Experimental Results of an Ambient Intelligence for Training on Soldering Iron Holding . . . . . . . . . . . . . . . . . . . . . . . . . 444 Yuto Hirota, Tetsuya Oda, Nobuki Saito, Aoto Hirata, Masaharu Hirota, and Kengo Katatama Design of Education Tool for Reinforcement-Learning Agent Developers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 454 Takahiro Uchiya, Kodai Shimano, and Ichi Takumi Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 463

Performance Evaluation of a Message Relaying Method for Resilient Disaster Networks Yoshiki Tada1 , Makoto Ikeda2(B) , and Leonard Barolli2 1

Graduate School of Engineering, Fukuoka Institute of Technology, 3-30-1 Wajiro-higashi, Higashi-ku, Fukuoka 811-0295, Japan [email protected] 2 Department of Information and Communication Engineering, Fukuoka Institute of Technology, 3-30-1 Wajiro-higashi, Higashi-ku, Fukuoka 811-0295, Japan [email protected], [email protected]

Abstract. In this work, we focus on the transmission of messages by vehicles for multiple users in a disaster area. Previously, we have discussed the effectiveness of message transmission by vehicles based on the Delay Tolerant Networking (DTN). In this paper, we evaluate the network performance considering the impact of closed roads for resilient disaster network. We use Epidemic with recovery function and the proposed Enhanced Dynamic Timer (EDT) as the message delivery protocols. From the simulation results, we found that the delay of proposed EDT is good for both normal and disaster situations.

Keywords: DTN timer · Epidemic

1

· Resilient disaster network · Enhanced dynamic

Introduction

We generally do not have alternative methods of information communication when the network is disrupted by a disaster or equipment trouble. Therefore, it is important to build a network that is resilient to disasters. A resilient disaster network provides a communication lifeline to users, such as quickly switching to an alternative network in the event of a communication disruption [15]. We focus on vehicles as alternative network players in disaster situations. The vehiclebased communications can be Vehicle-to-Vehicle (V2V), Vehicle-to-Pedestrian (V2P) and Vehicle-to-Infrastructure (V2I) [5,8,9,11]. Delay Tolerant Networking (DTN) [7] is effective as a communication method for these vehicle-based communications. By these methods, we can expect in the future new vehiclebased alternative network services. Moreover, the further development of caching method in the relay vehicles used by DTN will have many applications for Information/Content Centric Networking (ICN/CCN). However, DTN has some c The Editor(s) (if applicable) and The Author(s), under exclusive license  to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 1–10, 2021. https://doi.org/10.1007/978-3-030-61108-8_1

2

Y. Tada et al.

problems such as overhead and storage consumption because the vehicles send duplicated messages to neighbors. In [4], the authors proposed a hybrid DTN routing method which selects the Epidemic-based protocol with many replications and the SpW-based protocol with few replications. They consider storage state of the nodes to select the routing protocol. In our previous works [6,10], we have proposed a message relaying method with an Enhanced Dynamic Timer (EDT) considering both grid road and real map scenarios in Vehicular DTN. In these papers, we have not evaluated the impacts of closed roads, which are common in the disaster situations. In this paper, we evaluate the network performance considering the impact of closed roads for resilient disaster network. For evaluation, we use Epidemic with anti-packet and the proposed EDT as the message delivery protocols. The structure of the paper is as follows. In Sect. 2, we give the message relaying methods. In Sect. 3, we provide the description of the evaluation system and the results. Finally, conclusions and future work are given in Sect. 4.

2 2.1

Message Relaying Methods Overview of DTN

DTN can provide a reliable internet-working for space tasks [3,13,17]. The space networks have possibly long delay, frequent link disconnection and frequent disruption. In Vehicular DTN, the intermediate vehicles stored messages in their storage and then sent to other vehicles. The network architecture is specified in RFC 4838 [2]. For DTN, the famous routing protocol is Epidemic routing [12,16]. Epidemic routing uses two control messages to replicate the bundle messages. Nodes periodically broadcast the Summary Vector (SV) message in the network. The SV contains a list of stored messages of each vehicle. When the vehicles receive the SV, they compare received SV to their SV. The vehicle sends a REQUEST message if the received SV contains an unknown message. In Epidemic routing, consumption of network resources and storage usage become a critical problem, because the nodes duplicate messages to neighbors in their communication range. Therefore, received messages remain in the storage and the messages are continuously duplicated even if the destination receives the messages. However, recovery schemes such as timer or anti-packet may delete the duplicate messages in the network. In the case of the timer, messages have a lifetime. The messages are punctually deleted when the lifetime of the messages is expired. However, the setting of a suitable lifetime is difficult. In the case of anti-packet, the destination broadcasts the anti-packet, which contains the list of messages that are received by the destination. Nodes delete the messages according to the anti-packet. Then, the nodes duplicate the anti-packet to other nodes. However, the network resources are consumed by anti-packet.

Performance Evaluation of a Message Relaying Method

3

In this paper, we evaluate a message relaying method with EDT compared to Epidemic equipped with anti-packet under closed road situations to realize the resilient disaster networks. 2.2

Overview of Proposed EDT

In this section, we describe a message relay method considering EDT. The conventional method has a fixed lifetime set for message generation. The EDT controls the bundle drop due considering network conditions. If the lifetime is expired, the vehicle deletes the bundle message in their storage. However, the conventional method does not consider network conditions such as non-signal time, density and so on. The flowchart of the proposed EDT method is shown in Fig. 1. In our approach, each vehicle periodically counts the number of received SVs and measure the non-signal time (NT) from its neighbors. If the current NT is greater than maximum NT (NTmax ), the NTmax will be updated. The EDT is calculated by Eq. (1): EDT = NTmax + Interval,

(1)

where Interval indicates the check interval, which is used for counting the number of received SVs. The EDT method considers the number of neighboring vehicles measured last time (Nprev ) and calculates the change rate of neighboring vehicles by comparing

Fig. 1. Flowchart of proposed message relaying method with EDT.

4

Y. Tada et al.

the value with the Rest Threshold (RT). The timer is reset by the condition of Eq. (2): Nnow ≤ RT. Nprev

(2)

The proposed EDT method can improve the network performance by switching the timer reset condition. In general, if the timer is set in the message, the lifetime is not reset. But, in our proposed method, the EDT reset is allowed even if the lifetime is reset more than one time, which means that there is no reset limit. We consider the Interval to keep the message in the storage for counting the next SV.

(a) Case A under closed roads

(b) Case B under closed roads

(c) Case C under closed roads

Fig. 2. Road models.

3 3.1

Evaluation Results Evaluation Setting

We implemented the proposed EDT method on the Scenargie network simulator [14]. We consider a grid road scenario for both normal and disaster situations (see Fig. 2). For disaster situation, there are some closed roads. On the other hand, there are no closed roads for normal situation. For both situations, the person who needs help in each space area replicates the bundle messages to the relay vehicles and deliver the bundle message to the end area. The end area is assumed to be an evacuation center or an Access Point (AP) for collecting disaster information. For this evaluation, we consider two vehicular densities with 15 and 30 vehicles. Table 1 shows the simulation parameters used for the network simulator. The people who need help are assumed to be the start-points and end area is the end-point. Both start-points and end-point are static. The vehicles move on the road except the closed roads based on the map-based random way-point mobility model.

Performance Evaluation of a Message Relaying Method

5

Table 1. Simulation parameters. Parameter

Value

Simulation time (Tmax )

600 [s]

Number of vehicles

15, 30 [vehicles]

Minimum speed

8.333 [m/s]

Maximum speed

16.666 [m/s]

Number of start-points

5 (Case A), 9 (Case B), 14 (Case C)

Number of end-points

1

Message start and end time

1–400 [s]

Message generation interval

10 [s]

Message size

1, 000 [bytes]

PHY model

IEEE 802.11p

Propagation model

ITU-R P.1411

Antenna model

Omni-directional

EDT: Activated

60 to 600 [s]

EDT: Reset Threshold (RT)

0.5

EDT: Check Interval (Interval) 60 [s]

The start-points send bundle messages to end-point according to ITU-R P.1411 [1] propagation model. When the vehicles receive bundle messages, they store the bundle messages in their storage. Each vehicle periodically broadcasts a SV that stores its own bundle list. Then, the vehicles duplicate the bundles to other vehicles. We considered the interference from obstacles at 5.9 GHz radio channel. We evaluate the performance of delay, Packet Delivery Ratio (PDR), overhead and storage usage for different vehicles. The delay indicates the transmission delay of the bundle to reach the end-area. The PDR indicates the value of the generated bundle messages divided by the delivered bundle messages. The overhead indicates the number of times for sending duplicate bundles. Timer function will be activated after the simulation time is 60 s. The storage usage indicates the average of the storage state of each vehicle. 3.2

Simulation Results

We evaluate the proposed message relaying method with EDT compared to Epidemic with anti-packet for normal and disaster situations. In all cases, all vehicles continue to move on the road during the simulation. We present the simulation results of delay for different cases in Fig. 3. In all cases, the results of the delay improve with increasing numbers of vehicles. For the EDT with 15 vehicles, the delay results are smaller than with Epidemic. On the other hand, for 30 vehicles, we found that EDT has a longer delay compared

6

Y. Tada et al.

to Epidemic. When the road is impassable, the delay is greater than in a normal situation due to the limited mobility of vehicles.

(a) Case A

(b) Case B

(c) Case C

Fig. 3. Delay for different cases.

We show the simulation results of the PDR for different cases in Fig. 4. For Epidemic, the PDR results are significantly higher regardless of the number of vehicles. For normal situation, the proposed EDT method is lower than Epidemic, but the difference is about 10%. On the other hand, for disaster situation, the difference between the PDRs is about 31%. This effect is more significant when the node density is low. The PDR is about 89% with 30 vehicles using proposed EDT. We present the simulation results of overhead for different cases in Fig. 5. The results of overhead increase by increasing the number of vehicles. There is a small difference in overhead in some cases (both Case A and Case B with 15 vehicles), but there is a significant difference for the PDR in the case with 15 vehicles. The Epidemic has the smallest overhead in normal situation due to the anti-packet effect, but the results in a disaster situation are almost the same with the proposed method. In Case C with 15 vehicles, the EDT overhead is 1.5 times higher compared with other cases. As a result, the PDR reaches 95%. We show the simulation results of the storage usage for different cases in Fig. 6. For the Epidemic, the storage usage is larger in disaster situation than in

Performance Evaluation of a Message Relaying Method

(a) Case A

(b) Case B

(c) Case C

Fig. 4. PDR for different cases.

(a) Case A

(b) Case B

(c) Case C

Fig. 5. Overhead for different cases.

7

8

Y. Tada et al.

(a) Case A: 15 vehicles

(b) Case B: 15 vehicles

(c) Case C: 15 vehicles

(d) Case A: 30 vehicles

(e) Case B: 30 vehicles

(f) Case C: 30 vehicles

Fig. 6. Storage usage for different cases.

normal situation. This is due to the delay in anti-packet generation. Then, when the number of vehicles increases in Epidemic, the storage usage is improved. For EDT, we observed that the storage is going up and down at constant intervals in both situations. We think the effect of this constant interval is Eq. (1).

4

Conclusions

In this paper, we evaluated the network performance considering the impact of closed roads for resilient disaster network. For evaluation, we used Epidemic with

Performance Evaluation of a Message Relaying Method

9

anti-packet and the proposed EDT as message delivery protocols. From these results, we found that the delay of EDT is good for both normal and disaster situations. However, there are some problems in PDR and storage reduction timing. In future work, we would like to investigate the impact of the activated time and adapted RT parameters. Acknowledgments. This work has been partially funded by the research project from Comprehensive Research Organization at Fukuoka Institute of Technology (FIT), Japan.

References 1. Rec. ITU-R P.1411-7: Propagation data and prediction methods for the planning of short-range outdoor radiocommunication systems and radio local area networks in the frequency range 300 MHz to 100 GHz. ITU (2013) 2. Cerf, V., Burleigh, S., Hooke, A., Torgerson, L., Durst, R., Scott, K., Fall, K., Weiss, H.: Delay-tolerant networking architecture. IETF RFC 4838 (Informational), April 2007 3. Fall, K.: A delay-tolerant network architecture for challenged Internets. In: Proceedings of the International Conference on Applications, Technologies, Architectures, and Protocols for Computer Communications, SIGCOMM 2003, pp. 27–34 (2003) 4. Henmi, K., Koyama, A.: Hybrid type DTN routing protocol considering storage capacity. In: Proceedings of the 8th International Conference on Emerging Internet, Data and Web Technologies (EIDWT 2020), pp. 491–502, February 2020 5. Hou, X., Li, Y., Chen, M., Wu, D., Jin, D., Chen, S.: Vehicular fog computing: a viewpoint of vehicles as the infrastructures. IEEE Trans. Veh. Technol. 65(6), 3860–3873 (2016) 6. Ikeda, M., Nakasaki, S., Tada, Y., Barolli, L.: Performance evaluation of a message relaying method with enhanced dynamic timer in vehicular DTN. In: Proceedings of the Workshops of the 34th International Conference on Advanced Information Networking and Applications (WAINA-2020), pp. 332–340, April 2020 7. Kawabata, N., Yamasaki, Y., Ohsaki, H.: Hybrid cellular-DTN for vehicle volume data collection in rural areas. In: Proceedings of the IEEE 43rd Annual Computer Software and Applications Conference (COMPSAC-2019), vol. 2, pp. 276–284, July 2019 8. Ku, I., Lu, Y., Gerla, M., Gomes, R.L., Ongaro, F., Cerqueira, E.: Towards software-defined VANET: architecture and services. In: Proceedings of the 13th Annual Mediterranean Ad Hoc Networking Workshop (MED-HOC-NET-2014), pp. 103–110, June 2014 9. Lin, D., Kang, J., Squicciarini, A., Wu, Y., Gurung, S., Tonguz, O.: MoZo: a moving zone based routing protocol using pure V2V communication in VANETs. IEEE Trans. Mob. Comput. 16(5), 1357–1370 (2017) 10. Nakasaki, S., Ikeda, M., Barolli, L.: A message relaying method with enhanced dynamic timer considering decrease rate of neighboring nodes for Vehicular-DTN. In: Proceedings of the 14th International Conference on Broad-Band Wireless Computing, Communication and Applications (BWCCA-2019), pp. 711–720, November 2019

10

Y. Tada et al.

11. Ning, Z., Hu, X., Chen, Z., Zhou, M., Hu, B., Cheng, J., Obaidat, M.S.: A cooperative quality-aware service access system for social internet of vehicles. IEEE Internet of Things J. 5(4), 2506–2517 (2018) 12. Ramanathan, R., Hansen, R., Basu, P., Hain, R.R., Krishnan, R.: Prioritized epidemic routing for opportunistic networks. In: Proceedings of the 1st International MobiSys Workshop on Mobile Opportunistic Networking (MobiOpp 2007), pp. 62–66 (2007) 13. R¨ usch, S., Sch¨ urmann, D., Kapitza, R., Wolf, L.: Forward secure delay-tolerant networking. In: Proceedings of the 12th Workshop on Challenged Networks (CHANTS-2017), pp. 7–12, October 2017 14. Scenargie: Space-time engineering, LLC. http://www.spacetime-eng.com/ 15. Solpico, D., Tan, M.I., Manalansan, E.J., Zagala, F.A., Leceta, J.A., Lanuza, D.F., Bernal, J., Ramos, R.D., Villareal, R.J., Cruz, X.M., dela Cruz, J.A., Lagazo, D.J., Honrado, J.L., Abrajano, G., Libatique, N.J., Tangonan, G.: Application of the V-HUB standard using LoRa beacons, mobile cloud, UAVs, and DTN for disaster-resilient communications. In: Proceedings of the IEEE Global Humanitarian Technology Conference (GHTC-2019), pp. 1–8, October 2019 16. Vahdat, A., Becker, D.: Epidemic routing for partially-connected ad hoc networks. Duke University, Technical report (2000) 17. Wyatt, J., Burleigh, S., Jones, R., Torgerson, L., Wissler, S.: Disruption tolerant networking flight validation experiment on NASA’s EPOXI mission. In: Proceedings of the 1st International Conference on Advances in Satellite and Space Communications (SPACOMM 2009), pp. 187–196, July 2009

A Comparison Study of Constriction and Random Inertia Weight Router Replacement Methods for WMNs by WMN-PSOSA-DGA Hybrid Simulation System Considering Chi-square Distribution of Mesh Clients Admir Barolli1 , Shinji Sakamoto2 , Phudit Ampririt3 , Seiji Ohara3 , Leonard Barolli4(B) , and Makoto Takizawa5 1

Department of Information Technology, Aleksander Moisiu University of Durres, L.1, Rruga e Currilave, Durres, Albania [email protected] 2 Department of Computer and Information Science, Seikei University, 3-3-1 Kichijoji-Kitamachi, Musashino-shi, Tokyo 180-8633, Japan [email protected] 3 Graduate School of Engineering, Fukuoka Institute of Technology, 3-30-1 Wajiro-Higashi, Higashi-Ku, Fukuoka 811-0295, Japan [email protected], [email protected] 4 Department of Information and Communication Engineering, Fukuoka Institute of Technology, 3-30-1 Wajiro-Higashi, Higashi-Ku, Fukuoka 811-0295, Japan [email protected] 5 Department of Advanced Sciences, Faculty of Science and Engineering, Hosei University, Kajino-Machi, Koganei-Shi, Tokyo 184-8584, Japan [email protected] Abstract. Wireless Mesh Networks (WMNs) have many advantages such as: easy maintenance, low upfront cost and high robustness. The connectivity and stability affect directly the performance of WMNs. In our previous work, we implemented a simulation system considering Particle Swarm Optimization (PSO), Simulated Annealing (SA) and Distributed Genetic Algorithm (DGA), called WMN-PSOSA-DGA. In this paper, we evaluate the performance of Constriction Method (CM) and Random Inertia Weight Method (RIWM) for WMNs using WMNPSOSA-DGA hybrid simulation system considering Chi-square distribution of mesh clients. Simulation results show that a good performance is achieved for CM compared with the case of RIWM.

1

Introduction

The wireless networks and devices are becoming increasingly popular and they provide users access to information and communication anytime and c The Editor(s) (if applicable) and The Author(s), under exclusive license  to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 11–21, 2021. https://doi.org/10.1007/978-3-030-61108-8_2

12

A. Barolli et al.

anywhere [1,3,12,13,15,19,22,23]. Wireless Mesh Networks (WMNs) are gaining a lot of attention because of their low cost nature that makes them attractive for providing wireless Internet connectivity. A WMN is dynamically self-organized and self-configured, with the nodes in the network automatically establishing and maintaining mesh connectivity among them-selves (creating, in effect, an ad hoc network). This feature brings many advantages to WMNs such as low up-front cost, easy network maintenance, robustness and reliable service coverage [2]. Mesh node placement in WMN can be seen as a family of problems, which are shown (through graph theoretic approaches or placement problems, e.g. [10,17]) to be computationally hard to solve for most of the formulations [35]. We consider the version of the mesh router nodes placement problem in which we are given a grid area where to deploy a number of mesh router nodes and a number of mesh client nodes of fixed positions (of an arbitrary distribution) in the grid area. The objective is to find a location assignment for the mesh routers to the cells of the grid area that maximizes the network connectivity and client coverage. Network connectivity is measured by Size of Giant Component (SGC) of the resulting WMN graph, while the user coverage is simply the number of mesh client nodes that fall within the radio coverage of at least one mesh router node and is measured by Number of Covered Mesh Clients (NCMC). Node placement problems are known to be computationally hard to solve [14,36]. In previous works, some intelligent algorithms have been investigated for node placement problem [5,11,18,20,21,27,28,37]. In [26], we implemented a Particle Swarm Optimization (PSO) and Simulated Annealing (SA) based simulation system, called WMN-PSOSA. Also, we implemented another simulation system based on Genetic Algorithm (GA), called WMN-GA [5,16], for solving node placement problem in WMNs. Then, we designed a hybrid intelligent system based on PSO, SA and DGA, called WMN-PSOSA-DGA [25]. In this paper, we evaluate the performance of Constriction Method (CM) and Random Inertia Weight Method (RIWM) for WMNs using WMN-PSOSA-DGA simulation system considering Chi-square distribution of mesh clients. The rest of the paper is organized as follows. We present our designed and implemented hybrid simulation system in Sect. 2. The simulation results are given in Sect. 3. Finally, we give conclusions and future work in Sect. 4.

2

Proposed and Implemented Simulation System

Distributed Genetic Algorithm (DGA) has been focused from various fields of science. DGA has shown their usefulness for the resolution of many computationally hard combinatorial optimization problems. Also, Particle Swarm Optimization (PSO) and Simulated Annealing (SA) are suitable for solving NP-hard problems.

A Comparison Study of CM and RIWM by WMN-PSOSA-DGA

2.1

13

Velocities and Positions of Particles

WMN-PSOSA-DGA decides the velocity of particles by a random process considering the area size. For instance, when √ the area size is W × H, the velocity √ is decided randomly from − W 2 + H 2 to W 2 + H 2 . Each particle’s velocities are updated by simple rule [24]. For SA mechanism, next positions of each particle are used for neighbor solution s . The fitness function f gives points to the current solution s. If f (s ) is larger than f (s), the s is better than s so the s is updated to s . However, if f (s) is not larger  than f (s), the s may be updated by using the probability of f (s )−f (s) . Where T is called the “Temperature value” which is decreased exp T with the computation so that the probability to update will be decreased. This mechanism of SA is called a cooling schedule and the next Temperature value of computation is calculated as Tn+1 = α × Tn . In this paper, we set the starting temperature, ending temperature and number of iterations. We calculate α as  α=

SA ending temperature SA starting temperature

1.0/number of iterations .

It should be noted that the positions are not updated but the velocities are updated in the case when the solution s is not updated. 2.2

Routers Replacement Methods

A mesh router has x, y positions and velocity. Mesh routers are moved based on velocities. There are many router replacement methods. In this paper, we use CM and RIWM. Constriction Method (CM) CM is a method which PSO parameters are set to a week stable region (ω = 0.729, C1 = C2 = 1.4955) based on analysis of PSO by M. Clerc et al. [6,9,31]. Random Inertia Weight Method (RIWM) In RIWM, the ω parameter is changing randomly from 0.5 to 1.0. The C1 and C2 are kept 2.0. The ω can be estimated by the week stable region. The average of ω is 0.75 [8,29,33]. Linearly Decreasing Inertia Weight Method (LDIWM) In LDIWM, C1 and C2 are set to 2.0, constantly. On the other hand, the ω parameter is changed linearly from unstable region (ω = 0.9) to stable region (ω = 0.4) with increasing of iterations of computations [7,34]. Linearly Decreasing Vmax Method (LDVM) In LDVM, PSO parameters are set to unstable region (ω = 0.9, C1 = C2 = 2.0). A value of Vmax which is maximum velocity of particles is considered. With increasing of iteration of computations, the Vmax is kept decreasing linearly [8,30,32].

14

A. Barolli et al.

Fig. 1. Model of WMN-PSOSA-DGA migration.

Fig. 2. Relationship among global solution, particle-patterns and mesh routers in PSOSA part.

Rational Decrement of Vmax Method (RDVM) In RDVM, PSO parameters are set to unstable region (ω = 0.9, C1 = C2 = 2.0). The Vmax is kept decreasing with the increasing of iterations as Vmax (x) =



W 2 + H2 ×

T −x . x

Where, W and H are the width and the height of the considered area, respectively. Also, T and x are the total number of iterations and a current number of iteration, respectively [24].

2.3

DGA Operations

Population of individuals: Unlike local search techniques that construct a path in the solution space jumping from one solution to another one through local perturbations, DGA use a population of individuals giving thus the search a larger scope and chances to find better solutions. This feature is also known as “exploration” process in difference to “exploitation” process of local search methods. Selection: The selection of individuals to be crossed is another important aspect in DGA as it impacts on the convergence of the algorithm. Several selection schemes have been proposed in the literature for selection operators trying to cope with premature convergence of DGA. There are many selection methods

A Comparison Study of CM and RIWM by WMN-PSOSA-DGA

15

in GA. In our system, we implement 2 selection methods: Random method and Roulette wheel method. Crossover operators: Use of crossover operators is one of the most important characteristics. Crossover operator is the means of DGA to transmit best genetic features of parents to offsprings during generations of the evolution process. Many methods for crossover operators have been proposed such as Blend Crossover (BLX-α), Unimodal Normal Distribution Crossover (UNDX), Simplex Crossover (SPX). Mutation operators: These operators intend to improve the individuals of a population by small local perturbations. They aim to provide a component of randomness in the neighborhood of the individuals of the population. In our system, we implemented two mutation methods: uniformly random mutation and boundary mutation. Escaping from local optimal: GA itself has the ability to avoid falling prematurely into local optimal and can eventually escape from them during the search process. DGA has one more mechanism to escape from local optimal by considering some islands. Each island computes GA for optimizing and they migrate its gene to provide the ability to avoid from local optimal. Convergence: The convergence of the algorithm is the mechanism of DGA to reach to good solutions. A premature convergence of the algorithm would cause that all individuals of the population be similar in their genetic features and thus the search would result ineffective and the algorithm getting stuck into local optimal. Maintaining the diversity of the population is therefore very important to this family of evolutionary algorithms. In following, we present fitness function, migration function, particle pattern, gene coding and client distributions. 2.4

Fitness and Migration Functions

The determination of an appropriate fitness function, together with the chromosome encoding are crucial to the performance. Therefore, one of most important thing is to decide the determination of an appropriate objective function and its encoding. In our case, each particle-pattern and gene has an own fitness value which is comparable and compares it with other fitness value in order to share information of global solution. The fitness function follows a hierarchical approach in which the main objective is to maximize the SGC in WMN. Thus, the fitness function of this scenario is defined as Fitness = 0.7 × SGC(xij , y ij ) + 0.3 × NCMC(xij , y ij ). Our implemented simulation system uses Migration function as shown in Fig. 1. The Migration function swaps solutions between PSOSA part and DGA part. 2.5

Particle-Pattern and Gene Coding

In order to swap solutions, we design particle-patterns and gene coding carefully. A particle is a mesh router. Each particle has position in the considered area

16

A. Barolli et al.

Fig. 3. Chi-square distribution.

Fig. 4. Simulation results of WMN-PSOSA-DGA for SGC.

and velocities. A fitness value of a particle-pattern is computed by combination of mesh routers and mesh clients positions. In other words, each particle-pattern is a solution as shown is Fig. 2. A gene describes a WMN. Each individual has its own combination of mesh nodes. In other words, each individual has a fitness value. Therefore, the combination of mesh nodes is a solution.

3

Simulation Results

In this section, we show simulation results. In this work, we analyze the performance of CM and RIWM for WMNs by WMN-PSOSA-DGA hybrid intelligent system considering Chi-square client distribution. Our proposed system can generate many client distributions [4]. Here, we consider Chi-square distribution of mesh clients as shown in Fig. 3. The number of mesh routers is considered 16 and the number of mesh clients 48. We conducted simulations 10 times in order to avoid the effect of randomness and create a general view of results. We show the parameter setting for WMN-PSOSA-DGA in Table 1.

A Comparison Study of CM and RIWM by WMN-PSOSA-DGA

17

Table 1. WMN-PSOSA-DGA parameters. Parameters

Values

Clients distribution

Chi-Square

Area size

32.0 × 32.0

Number of mesh routers

16

Number of mesh clients

48

Number of GA islands

16

Number of Particle-patterns 32 Number of migrations

200

Evolution steps

320

Radius of a mesh router

2.0–3.5

Selection method

Roulette wheel method

Crossover method

SPX

Mutation method

Boundary mutation

Crossover rate

0.8

Mutation rate

0.2

SA Starting value

10.0

SA Ending value

0.01

Total number of iterations

64000

Replacement method

CM, RIWM

We show simulation results in Fig. 4 and Fig. 5. We see that for SGC, CM converges faster than RIWM, but the performance is finally almost the same. However, for NCMC, CM performs better than RIWM. The visualized simulation results are shown in Fig. 6. For both replacement methods, all mesh routers are connected, but some clients are not covered. We can see the number of covered mesh clients is larger for CM compared with the case of RIWM.

Fig. 5. Simulation results of WMN-PSOSA-DGA for NCMC.

18

A. Barolli et al.

Fig. 6. Visualized simulation results of WMN-PSOSA-DGA for different replacement methods.

4

Conclusions

In this work, we evaluated the performance of CM and RIWM replacement methods for WMNs using a hybrid simulation system based on PSO, SA and DGA (called WMN-PSOSA-DGA) considering Chi-square distribution of mesh clients. Simulation results show that a good performance was achieved for CM compared with the case of RIWM. In our future work, we would like to evaluate the performance of the proposed system for different parameters and patterns.

References 1. Ahmed, S., Khan, M.A., Ishtiaq, A., Khan, Z.A., Ali, M.T.: Energy harvesting techniques for routing issues in wireless sensor networks. Int. J. Grid Util. Comput. 10(1), 10–21 (2019) 2. Akyildiz, I.F., Wang, X., Wang, W.: Wireless mesh networks: a survey. Comput. Netw. 47(4), 445–487 (2005) 3. Barolli, A., Sakamoto, S., Barolli, L., Takizawa, M.: A hybrid simulation system based on particle swarm optimization and distributed genetic algorithm for WMNs: performance evaluation considering normal and uniform distribution of mesh clients. In: International Conference on Network-Based Information Systems, pp. 42–55. Springer (2018) 4. Barolli, A., Sakamoto, S., Barolli, L., Takizawa, M.: Performance analysis of simulation system based on particle swarm optimization and distributed genetic algorithm for WMNs considering different distributions of mesh clients. In: International Conference on Innovative Mobile and Internet Services in Ubiquitous Computing, pp. 32–45. Springer (2018) 5. Barolli, A., Sakamoto, S., Ozera, K., Barolli, L., Kulla, E., Takizawa, M.: Design and implementation of a hybrid intelligent system based on particle swarm optimization and distributed genetic algorithm. In: International Conference on Emerging Internetworking, Data & Web Technologies, pp. 79–93. Springer (2018)

A Comparison Study of CM and RIWM by WMN-PSOSA-DGA

19

6. Barolli, A., Sakamoto, S., Durresi, H., Ohara, S., Barolli, L., Takizawa, M.: A comparison study of constriction and linearly decreasing Vmax replacement methods for wireless mesh networks by WMN-PSOHC-DGA simulation system. In: International Conference on P2P, Parallel, Grid, Cloud and Internet Computing, pp. 26–34. Springer (2019) 7. Barolli, A., Sakamoto, S., Ohara, S., Barolli, L., Takizawa, M.: Performance analysis of WMNs by WMN-PSOHC-DGA simulation system considering linearly decreasing inertia weight and linearly decreasing Vmax replacement methods. In: International Conference on Intelligent Networking and Collaborative Systems, pp. 14–23. Springer (2019) 8. Barolli, A., Sakamoto, S., Ohara, S., Barolli, L., Takizawa, M.: Performance analysis of WMNs by WMN-PSOHC-DGA simulation system considering random inertia weight and linearly decreasing Vmax router replacement methods. In: Conference on Complex, Intelligent, and Software Intensive Systems, pp. 13–21. Springer (2019) 9. Clerc, M., Kennedy, J.: The particle swarm-explosion, stability, and convergence in a multidimensional complex space. IEEE Trans. Evol. Comput. 6(1), 58–73 (2002) 10. Franklin, A.A., Murthy, C.S.R.: Node placement algorithm for deployment of twotier wireless mesh networks. In: Proceedings of Global Telecommunications Conference, pp. 4823–4827 (2007) 11. Girgis, M.R., Mahmoud, T.M., Abdullatif, B.A., Rabie, A.M.: Solving the wireless mesh network design problem using genetic algorithm and simulated annealing optimization methods. Int. J. Comput. Appl. 96(11), 1–10 (2014) 12. Gorrepotu, R., Korivi, N.S., Chandu, K., Deb, S.: Sub-1GHz miniature wireless sensor node for IoT applications. Internet of Things 1, 27–39 (2018) 13. Islam, M.M., Funabiki, N., Sudibyo, R.W., Munene, K.I., Kao, W.C.: A dynamic access-point transmission power minimization method using PI feedback control in elastic WLAN system for IoT applications. Internet of Things 8(100), 089 (2019) 14. Maolin, T., et al.: Gateways placement in backbone wireless mesh networks. Int. J. Commun. Netw. Syst. Sci. 2(1), 44 (2009) 15. Marques, B., Coelho, I.M., Sena, A.D.C., Castro, M.C.: A network coding protocol for wireless sensor fog computing. Int. J. Grid Util. Comput. 10(3), 224–234 (2019) 16. Matsuo, K., Sakamoto, S., Oda, T., Barolli, A., Ikeda, M., Barolli, L.: Performance analysis of WMNs by WMN-GA simulation system for two WMN architectures and different TCP congestion-avoidance algorithms and client distributions. Int. J. Commun. Netw. Distrib. Syst. 20(3), 335–351 (2018) 17. Muthaiah, S.N., Rosenberg, C.P.: Single gateway placement in wireless mesh networks. In: Proceedings of 8th International IEEE Symposium on Computer Networks, pp. 4754–4759 (2008) 18. Naka, S., Genji, T., Yura, T., Fukuyama, Y.: A hybrid particle swarm optimization for distribution state estimation. IEEE Trans. Power Syst. 18(1), 60–68 (2003) 19. Ohara, S., Barolli, A., Sakamoto, S., Barolli, L.: Performance analysis of WMNs by WMN-PSODGA simulation system considering load balancing and client uniform distribution. In: International Conference on Innovative Mobile and Internet Services in Ubiquitous Computing, pp. 25–38. Springer (2019) 20. Ozera, K., Sakamoto, S., Elmazi, D., Bylykbashi, K., Ikeda, M., Barolli, L.: A fuzzy approach for clustering in MANETs: performance evaluation for different parameters. Int. J. Space-Based Situated Comput. 7(3), 166–176 (2017)

20

A. Barolli et al.

21. Ozera, K., Bylykbashi, K., Liu, Y., Barolli, L.: A fuzzy-based approach for cluster management in VANETs: performance evaluation for two fuzzy-based systems. Internet of Things 3, 120–133 (2018) 22. Ozera, K., Inaba, T., Bylykbashi, K., Sakamoto, S., Ikeda, M., Barolli, L.: A WLAN triage testbed based on fuzzy logic and its performance evaluation for different number of clients and throughput parameter. Int. J. Grid Util. Comput. 10(2), 168–178 (2019) 23. Petrakis, E.G., Sotiriadis, S., Soultanopoulos, T., Renta, P.T., Buyya, R., Bessis, N.: Internet of Things as a Service (iTaaS): challenges and solutions for management of sensor data on the cloud and the fog. Internet of Things 3, 156–174 (2018) 24. Sakamoto, S., Oda, T., Ikeda, M., Barolli, L., Xhafa, F.: Implementation of a new replacement method in WMN-PSO simulation system and its performance evaluation. In: The 30th IEEE International Conference on Advanced Information Networking and Applications (AINA 2016), pp. 206–211 (2016). https://doi.org/ 10.1109/AINA.2016.42 25. Sakamoto, S., Barolli, A., Barolli, L., Takizawa, M.: Design and implementation of a hybrid intelligent system based on particle swarm optimization, hill climbing and distributed genetic algorithm for node placement problem in WMNs: a comparison study. In: The 32nd IEEE International Conference on Advanced Information Networking and Applications (AINA 2018), pp. 678–685. IEEE (2018) 26. Sakamoto, S., Ozera, K., Ikeda, M., Barolli, L.: Implementation of intelligent hybrid systems for node placement problem in WMNs considering particle swarm optimization, hill climbing and simulated annealing. Mob. Netw. Appl. 23(1), 27–33 (2018) 27. Sakamoto, S., Barolli, A., Barolli, L., Okamoto, S.: Implementation of a web interface for hybrid intelligent systems. Int. J. Web Inf. Sys. 15(4), 420–431 (2019) 28. Sakamoto, S., Barolli, L., Okamoto, S.: WMN-PSOSA: an intelligent hybrid simulation system for WMNs and its performance evaluations. Int. J. Web Grid Serv. 15(4), 353–366 (2019) 29. Sakamoto, S., Ohara, S., Barolli, L., Okamoto, S.: Performance evaluation of WMNs by WMN-PSOHC system considering random inertia weight and linearly decreasing inertia weight replacement methods. In: International Conference on Innovative Mobile and Internet Services in Ubiquitous Computing, pp. 39–48. Springer (2019) 30. Sakamoto, S., Ohara, S., Barolli, L., Okamoto, S.: Performance evaluation of WMNs by WMN-PSOHC system considering random inertia weight and linearly decreasing Vmax replacement methods. In: International Conference on NetworkBased Information Systems, pp. 27–36. Springer (2019) 31. Sakamoto, S., Ohara, S., Barolli, L., Okamoto, S.: Performance evaluation of WMNs WMN-PSOHC system considering constriction and linearly decreasing inertia weight replacement methods. In: International Conference on Broadband and Wireless Computing, Communication and Applications, pp. 22–31. Springer (2019) 32. Schutte, J.F., Groenwold, A.A.: A study of global optimization using particle swarms. J. Glob. Optim. 31(1), 93–108 (2005) 33. Shi, Y.: Particle swarm optimization. IEEE Connections 2(1), 8–13 (2004) 34. Shi, Y., Eberhart, R.C.: Parameter selection in particle swarm optimization. In: Evolutionary Programming VII, pp. 591–600 (1998) 35. Vanhatupa, T., Hannikainen, M., Hamalainen, T.: Genetic algorithm to optimize node placement and configuration for WLAN planning. In: The 4th IEEE International Symposium on Wireless Communication Systems, pp. 612–616 (2007)

A Comparison Study of CM and RIWM by WMN-PSOSA-DGA

21

36. Wang, J., Xie, B., Cai, K., Agrawal, D.P.: Efficient mesh router placement in wireless mesh networks. In: Proceedings of IEEE International Conference on Mobile Adhoc and Sensor Systems (MASS 2007), pp. 1–9 (2007) 37. Yaghoobirafi, K., Nazemi, E.: An autonomic mechanism based on ant colony pattern for detecting the source of incidents in complex enterprise systems. Int. J. Grid Util. Comput. 10(5), 497–511 (2019)

Multi-source and Multi-target Node Selection in Energy-Efficient Fog Computing Model Yinzhe Guo1(B) , Takumi Saito1 , Shigenari Nakamura2 , Tomoya Enokido3 , Lei Li1 , and Makoto Takizawa1

2

1 Hosei University, Tokyo, Japan [email protected], [email protected], [email protected], [email protected] Tokyo Metropolitan Industrial Technology Research Institute, Tokyo, Japan [email protected] 3 Rissho University, Tokyo, Japan [email protected]

Abstract. In the fog computing model to realize the IoT, each fog node supports application processes to calculate output data on input data received from a fog node and sends the output data to another fog node. In our previous studies, types of the TBFC (Tree-Based Fog Computing) models are proposed to reduce the electric energy consumption and execution time of fog nodes and servers and to be tolerant of node faults. In the TBFC models, the tree structure of fog nodes is not changed even if some fog node is overloaded and underloaded. In this paper, we consider the DNFC (Dynamic Network-based Fog Computing) model. Here, there is one or more than one possible target fog node for each fog node and also one or more than one possible source node for each target node. A pair of a source node and target node which exchange data have to be selected. In this paper, we propose an MSMT (Multi-Source and MultiTarget node selection) protocol among multiple source and target nodes. Here, a pair of a source node and a target node are selected so that the total energy consumption of the nodes can be reduced. In the evaluation, we show the total energy consumption and total execution time by target nodes can be more reduced in the MSMT protocol. Keywords: IoT (Internet of Things) · Energy-efficient fog computing model · Dynamic Network-based Fog Computing (DNFC) model · MSMT (Multi-source and Multi-target node selection) protocol

1

Introduction

The IoT (Internet of Things) [2,4] is composed of not only computers but also devices like sensors and actuators installed in various things [9]. Compared with c The Editor(s) (if applicable) and The Author(s), under exclusive license  to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 22–33, 2021. https://doi.org/10.1007/978-3-030-61108-8_3

Multi-source and Multi-target Node Selection

23

traditional network systems like the cloud computing model [1], the IoT is so scalable that from huge amount of sensor data are transmitted billions of devices to servers in networks. The fog computing (FC) model is proposed to efficiently realize the IoT [12]. In order to not only increase the performance and reliability but also reduce the electric energy consumption of the IoT, types of the TBFC (Tree-Based Fog Computing) models are proposed in our previous studies [6,9–11]. Here, fog nodes are hierarchically structured in a tree. A root node shows a cloud of servers and leaf nodes indicate edge nodes which receive data from sensors and send actions to actuators. Each fog node receives input data from child nodes and calculates output data on the input data. Then, the fog node sends the output data to a parent node. A child-parent relation of fog nodes in a tree structure shows an output-input relation of data on the fog nodes. Even if some fog node is heavily congested while other fog nodes are lightly loaded, input data to the congested fog node cannot be distributed to other fog nodes due to the static tree structure of the TBFC model. The DNFC (Dynamic Network-based Fog Computing) model [7] is proposed to make the FC model more flexible. Here, there are one or more than one target node to which each source node can send output data and there are also one or more than one source node from which each target node can receive input data. In a set of the source nodes and target nodes, a pair of source fog node fi and a target fog node fj which send output data and receives the data, respectively, are decided. So that the total energy to be consumed by the source and target nodes can be reduced. In this paper, we propose an MSMT (Multi-Source and MultiTarget node selection) protocol to make pairs of source and target fog nodes through the negotiation among source and target fog nodes. In the evaluation, we show the total energy consumption and total execution time of fog nodes in the MSMT protocol is smaller than the DNFCN [8] and random (RD) protocols. In Sect. 2, we present the FC model. In Sect. 3, we discuss the computation and power consumption models of a fog node. In Sect. 4, we propose the MSMT protocol in the DNFC model. In Sect. 5, we evaluate the MSMT protocol.

2

Fog Computing Models

The fog computing (FC) model [12] to efficiently realize the IoT [9] is composed of fog nodes in addition to sensor and actuator devices and clouds of servers. Clouds are composed of servers like the cloud computing (CC) model [1]. Various devices equipped with sensors and actuators are connected to edge fog nodes. A sensor node sends sensor data to an edge fog node. Each fog node fi supports an application process p(fi ) to calculate output data on input data from sensor nodes and other fog nodes. Then, the fog node fi sends the output data to another fog node fj which can calculate on the data. Fog nodes are interconnected with other fog nodes in types of networks including wireless networks. Fog nodes also move in networks as discussed in the MFC (Mobile FC) model [5] (Fig. 1).

24

Y. Guo et al.

Fig. 1. Fog node model.

Let F be a set of fog nodes f1 , ..., fn (n ≥ 1) in a system. A fog node fi receives input data idik from each source fog node sfik (∈ F )(k = 1, ..., li )(li ≥ 1). IDi shows a collection of input data idi1 ..., idi,li (li ≥ 1) of the fog node fi . An application process p(fi ) supported by the fog node fi calculates a collection ODi of output data {odi1 , ..., odi,mi } (mi ≥ 1) on the input data IDi . Then, the fog node fi sends each output data odik of the output data ODi to a target fog node tfik (∈ F ). Here, fi ⇒ tfik shows that a fog node tfik is a target fog node of fi . Each target fog node tfik supports a process p(tfik ) which can calculate on the output data odik from the fog node fi . Let T N (fi ) and SN (fi ) be sets of target nodes tfi1 , ..., tfi,mi and source nodes sfi1 , ..., sfi,li of a fog node fi , respectively. A notation |d| shows the size [Byte] of data d. The ratio ori = |ODi |/|IDi | is the output ratio of a fog node fi . In the TBFC (Tree-Based FC) model, fog nodes are structured in a tree as shown in Fig. 2 [11]. A root node f shows a cloud of servers and a leaf node indicates an edge node which communicates with sensors and actuators. A child fog node fi sends output data to a parent fog node fj , i.e. fi ⇒ fj . That is, fj is a target fog node of fi and fi is a source fog node of fj .

Multi-source and Multi-target Node Selection

25

Fig. 2. TBFC model.

3 3.1

Computation and Power Consumption Models Computation Model

Let F be a set of fog nodes f1 , ..., fn (n ≥ 1) in a system. Each fog node fi receives input data IDi = {idi1 , ..., idi,li } of size sii (= |IDi |) from source fog nodes sfi1 , ..., sfi,li (li ≥ 0), respectively, where each node sfij ∈ F . Each fog node fi calculates the output data ODi = {od1 , ..., odi,mi } of size soi (= |ODi |) on the input data IDi by the application process p(fi ). Where soi = ori · sii for the output ratio ori . Then, the fog node fi sends each output data odik in the output data set ODi to each target node tfik (∈ F )(k = 1, ..., mi ). Let CRi show the computation rate of a fog node fi to a root node f , i.e. server. Let CR be the computation rate of the root node f . We assume the computation rate CR to be one. CRi ≤ CR for every fog node fi . This means, the server node fi is CRi times slower than the server node f . T Pi (x) shows the execution time [sec] of a fog node fi to calculate output data ODi on input data IDi of size x. That is, it takes T Pi (x) [sec] to perform a process p(fi ). The execution time T Pi (x) is cti · Ci (x) where cti is 1/CRi . In this paper, Ci (x) is assumed to be x or x2 . Ci (x) stands for the computation complexity of a process p(fi ) of a node fi . The execution time T Iik (x) and T Oik (x) [sec] of a fog node fi (∈ F ) to receive and send data of size x from a source node sfik and to a target node tfik (∈ F ), respectively, are proportional to the data size x, i.e. T Iik (x) = rci + rti · x and T Oik (x) = sci + sti · x. Here, rci , rti , sci , and sti are constants for fi . A fog node fi receives input data idi1 , ..., idi,li from source nodes sfi1 , ..., sfi,li (li ≥ 1), respectively, where xk = |idik |(k = 1, ..., li ) and x = x1 + · · · + xli . It

26

Y. Guo et al.

takes time T Ii (x) = T Ii1 (x1 ) + · · · + T Ii,li (xli ) [sec] to receive the input data idi1 , ..., idi,li from the source nodes sfi1 , ..., sfi,li , respectively. After finishing the calculation, a fog node fi sends output data odi , ..., odi,mi to target nodes tfi1 , ..., tfi,mi , respectively. It takes time T Oi (x) = T Oi1 (x1 )+· · ·+T Oi,mi (xmi ) for a fog node fi to send the output data odi , ..., odi,mi to the target nodes tfi1 , ..., tfi,mi , respectively, where xh = |odih | for h = 1, ..., mi . The total execution time T Ti (x) [sec] of a fog node fi for input data IDi of size x is given as follows: T Ti (x) = T Ii (x) + T Ci (x) + δi · T Oi (ori · x).

(1)

Here, δi = 0 if the fog node fi is a root node, otherwise δi = 1. 3.2

Energy Consumption Model

Next, we discuss the energy consumption model of a fog node. Let EIi (x), ECi (x), and EOi (x) show electric energy [J] consumed by a fog node fi to receive, calculate on, and send data of size x, respectively. First,we discuss the electric energy ECi (x) to be consumed by a fog node fi to calculate output data odi on input data IDi of size x (= |IDi |). In this paper, we assume each fog node fi follows the SPC (Simple Power Consumption) model [2–4]. A fog node fi consumes the maximum power maxEi [W] to do the calculation on input data IDi . On the other hand, the fog node fi consumes the minimum electric power minEi (< maxEi ) [W] if there is no input data. Since a fog node fi consumes the electric power maxEi [W] for time T Ci (x) [sec] to calculate output data on input data IDi of size x, the fog node fi consumes the energy EC(x)[J] = maxEi [w] · T Ci (x)[sec]. A fog node fi consumes electric power P Ii and P Oi [W] to receive and send data where P Ii = rei ·maxEi and P Oi = sei ·maxEi , respectively, where rei ≤ 1 and sei ≤ 1. It takes time T Ii (x) and T Oi (x) to receive and send data of size x from li (≥ 1) source nodes sfi1 , ..., sfi,li and to mi (≥ 1) target nodes tfi1 , ..., tfi,mi . A fog node fi consumes electric energy EIi (x) = P Ii ·T Ii (x) = rei ·maxEi · (li · rci + rti · x) [J] and EOi (x) = P Oi · T Oi (x) = sei · maxEi · (mi · sci + sti · x) [J] of a fog node fi to receive and send data of size x (> 0), respectively. A fog node fi consumes electric energy EEi (x) [J] to receive and calculate output data odi1 , ..., odi,mi on the input data idi1 , ..., idi,li of size x1 , ..., xli , respectively, where x = x1 + · · · + xli and to send the output data odi1 , ..., odi,mi of size ori · x to target fog nodes: EEi (x) = EIi (x) + ECR (x) + δi · EOi (ori · x) = (rei · T Ii (x) + T Ci (x) + δi · sei · T Oi (ori · x)) · maxEi = ((rei · (li · rci + rti · x)) + cti · Ci (x) + δi · sei · (mi · sci + sti · ori · x)) · maxEi .

(2)

Multi-source and Multi-target Node Selection

4 4.1

27

MSMT (Multi-source and Multi-target Selection) Protocol DNFC Model

Let F be a set of fog nodes in a system. Each fog node fi supports an application process p(fi ) which calculates output data odi on a collection IDi of input data received from other source nodes. Then, the output data odi is delivered to another target fog node fj whose process p(fj ) can do the calculation on the data odi . Here, the fog node fj is a target node of the source fog node fi (fi → fj ), i.e. the output data of fi is the input data of fj . Let T N (fi ) be a set {fj |fi → fj }(⊆ F ) of target nodes of a fog node fi . Let SN (fi ) be a set {fj |fj → fi }(⊆ F ) of source nodes of a fog node fi . A fog node may not communicate with every fog node due to the scalability. A fog node fj with which a fog node fi can communicate is an acquaintance fog node. An acquaintance relation “fi ↔ fj ” means that a pair of fog nodes fi and fj can communicate with each other. For example, if a pair of mobile fog nodes fi and fj are in the wireless communication range of each other, fi ↔ fj . In this paper, we assume the acquaintance relation ↔ is symmetric. Let AN (fi ) be a set {fi |fi ↔ fj } of acquaintance fog nodes with which a fog node fi can communicate. Suppose a pair of fog nodes and are interconnected in wireless networks. A pair fog nodes fi and fj which are in the communication range of each other can communicate with each other (fi ↔ fj ). For each fog node fi , let T AN (fi ) be a set {fj |fi → f j and fi ↔ fj } (⊆ F ) of acquaintance target nodes, i.e. fi and fj ↔ T AN (fi ) = T N (fi ) AN (fi ). In turn, SAN (fi ) is a set {fj |fj →  fi } (⊆ F ) of acquaintance source nodes, i.e. SAN (fi ) = SN (fi ) AN (fi ) as shown in Fig. 3. Thus, each source fog node fi can send output data odi to one or more than one target fog node while each target fog node fj may receive input data from one or more than one source fog node. A ready source fog node fi is a node which has output data odi to be sent to a target fog node. We have to make every pair fi , fj of a source fog node fi SAN (fj ) and a target fog node fj , where the source node fi in T AN (fi ) sends the output data odi to the target node fj which satisfies the following conditions: 1. Every ready source node fi can send output data odi to one target node. 2. Each target node fj can receive input data from multiple source nodes.

28

Y. Guo et al.

Fig. 3. Source and target nodes.

4.2

Negotiation Protocol

A fog node fi has to send the output data odij (= idji ) of size xi (= |odij |) to a target node tfij in the set T AN (fi ). Suppose each target node tfij in the set T AN (fi ) holds its own input data IDj to be calculated on before receiving the output data odij from the source node fi . If the source fog node fi sends the output data odij to the target fog node tfij , the target fog node tfij has to calculate output data odj on the input data idij (= odij ) from the fog node fi in addition to its own input data IDij of size Xi . Hence, the target fog node tfij totally consumes energy EEij (Xj + xi ) [J] for time T Tij (Xj + xi ) [sec] to calculate on both the input data IDj and odij where Xi = |IDj | and xi = |odij |. On the other hand, a target fog node fj may receive input data idjk from a source fog node sfjk in the set SAN (fj ) in addition to input data idji from another source fog node sfji . Thus, a target fog node fj which holds its own input data IDj of size Xj may receive input data from multiple source nodes. Suppose each source node sfji in the set SAN (fj ) selects a target fog node fj in the set T AN (fi ) since the energy consumption EEj (Xj + xi ) is minimum. Here, if every source fog node sfji in the set SAN (fj ) sends output data odi of size xi to the target fog node fj , the target fog node fj consumes energy EEj (Xj + sfji ∈SAN (fj ) xi ) which might be too large. We have to distribute output data of source nodes to different target nodes so that each of the target nodes does not get heavily loaded. A pair of a source fog node fi in SAN (fj ) and a target fog node fj in T AN (fi ) are selected to communicate with each other in an MSMT (Multisource and Multi-target node selection) protocol as follows: [MSMT protocol]

Multi-source and Multi-target Node Selection

29

Fig. 4. Source node fi .

[Source node fi ] [Fig. 4] 1. Let xi be the size |odi | of the output data odi of a source fog node fi . T AN (fi ) is a set of target nodes which can communicate with the source fog node fi . 2. The source fog node fi sends a processing request Qi (xi ) with the size xi of the output data odi to every target fog node tfij in the set T AN (fi ). 3. The source fog node fi waits for confirmations from target fog nodes. Suppose the node fi receives confirmations Ci1 (Ei1 ), ..., Ci,li (Ei,li ) from target fog nodes tfi1 , ..., tfi,li (li ≥ 1) in the set T AN (fi ), respectively. Here, Eij shows the expected energy consumption EEij (Xj + xi ) of each target fog node tfij to calculate on the input data odi of size xi and its own input data of IDij of size Xj . Let T N (⊆ T AN (fi )) be the set of target nodes tfi1 , ..., tfi,li , which send the confirmations to the source fog node fi . 4. The source fog node fi selects a target fog node tfij in the set T N where the expected energy consumption Eij is minimum. The source fog node fi sends a DOij (xi ) message to the target fog node tfij and a N O message to the other target fog nodes in the set T N . 5. The source fog node fi waits for an OK message from the selected target fog node tfij . On receipt of the OK message from the target fog node tfij , the source fog node sends the output data odi to the target fog node tfij . Otherwise, T AN (fi ) = T AN (fi ) − {tfij } and go to 2. [Target node fj ] [Fig. 5] 1. A target fog node fj receives a processing request Qji (xi ) from a source fog node sfji in the set SAN (fj ). 2. Let SN be a set of source fog nodes sfj1 , ..., sfj,lj which send the processing requests to the target fog node fj . The target fog node fj obtains the expected energy consumption Eij = EEj (Xj + xi ) to calculate on both its own input data IDj of size Xj and the input data idji of size xi from the source fog node sfji .

30

Y. Guo et al.

Fig. 5. Target node fj .

3. The target fog node fj sends a confirmation Cj (Eij ) to every source fog node sfji in the set SN . 4. The target fog node fj waits for a DOij (xi ) message from each source fog node sfji in the set SN . The target fog node fj receives a DOij (xi ) message from a source fog node sfji in the set SN . 5. Let SD be a set of the source nodes sfj1 , ..., sfj,mj from which the target node fj receives DO messages. a. The target node fj receive DO messages from multiple source nodes, i.e. |SD| ≥ 2. The target fog node fj selects a pair of source fog nodes sfji and sfjk where the expected energy consumption Eji = EEj (Xj + xi ) and Ejk = EEj (Xj + xk ) of the two source fog nodes sfji and sfjk are minimum and maximum, respectively. The target fog node fj sends an OK message to the source fog nodes sfji and sfjk . SD = SD − {sfji , sfjk }; T D = {sfji , sfjk }; b. The target fog node fj receives a DO message from one source fog node sfji , i.e. |SD| = 1. The target fog node fj sends an OK message to the source fog node sfji . SD = SD − {sfji }; T D = {sfji }; c. Unless the target fog node fj receives an OK message from any source fog node, i.e. |SD| = 0, go to 1. 6. The target fog node fj sends a N O message to every source fog node sfjk in the set SD, which is not selected. 7. On receipt of output data odji from every source node sfji in the set T D to which the target node fj sends OK, the target node fj calculates input data odj on the output data IDj and {idji |sfji ∈ T D}.

Multi-source and Multi-target Node Selection

5

31

Evaluation

We evaluate the MSMT protocol in terms of energy consumption E and execution time T of fog nodes. Suppose that there are a set SAN of source fog node sf1 , ..., sfl (l ≥ 1) and a set T AN of target fog nodes tf1 , ..., tfm (m ≥ 1) where each pair of fog nodes sfi and tfj are source and target nodes of tfj and sfi , respectively, and all the source and target fog nodes are in the communication range of one another. Each source fog node sfi holds output data odi of size xi . Each target fog node tfj holds its own input data IDj of size Xj . In the evaluation, we assume every fog node is implemented on a Raspberry Pi 4 model B. Every fog node has the same computation rate CR = 0.43[MB/sec] and the maximum power consumption maxE. The computation complexity Cij (x) of the process p(tfij ) of each target fog node tfij is x2 for size x of input data. Each source fog node sfi in the set SAN sends a processing request Qi (xi ) to every target fog node tfj in the set T AN . Each target fog node tfj selects a source fog node sfi whose xi is the smallest, i.e. the energy consumption Eij = EEj (Xj + xi ) is largest. The target fog node tfj sends the confirmation Cj (Eij ) to the source fog node sfi . Then, the source fog node sfi selects a target fog node tfj whose energy consumption Eij is smallest and sends DOij (xi ) to the target fog node tfj . The target fog node tfj sends OK to a pair of source nodes sfi and sfk whose xi and xk are the largest and smallest, respectively, in the source nodes which send DO messages to the target node tfj . On receipt of an OK message from a target node tfj , the source fog node sfi sends the output data odi to the target fog node tfj . The target fog node tfj consumes energy Ej = EEj (Xj + xi + xk ) to calculate on its own input data of size Xj and idi of size xi and idk of size xk . It takes time Tj = T Tj (Xj + xi + xk ) [sec] for the target node tfj . The total energy consumption E = E1 + ... + El [w · sec] and total execution time T = T1 + ... + Tl [sec] of the target fog nodes tf1 , ..., tfl are obtained in the evaluation. In the evaluation, we consider l source fog nodes and m target fog nodes and l = m. The size xi of output data odi of each source fog node sfi and the size Xj of input data idj of each target fog node tfj are randomly taken out of 1, 2, ..., 10[MB]. We consider a random (RD) protocol and the DNFCN protocol [8] in addition to the MSMT protocol. In the RD protocol, each source fog node sfi randomly selects a target fog node tfj in the set T AN . In the DNFCN protocol, each target fog node tfj receives data from only one source fog node. As shown in Fig. 6, the total energy consumption of l target fog nodes is smaller in the MSMT protocol than the DNFCN and RD protocols. The total energy consumption of target fog nodes in the MSMT protocol is smaller than the DNFCN and RD protocols. For example, the total energy consumption of the MSMT protocol is 30% and 20% smaller than the DNFCN protocol and the RD protocol, respectively, for l = 80. Fig. 7 shows the total execution time T of l target nodes. As shown in Fig. 7, the total execution time T of the target nodes is shorter than the DNFCN and RD protocols.

32

Y. Guo et al.

Fig. 6. Energy consumption of target nodes.

Fig. 7. Execution time of target nodes.

6

Concluding Remarks

In the DNFC (Dynamic Network-based Fog Computing) model [8], each source fog node dynamically selects a target fog node, each time the source fog nodes send output data. In order to more reduce the total energy consumption of fog nodes, we proposed the MSMT protocol to do the negotiation among source and target fog nodes to select pairs of a target fog node and a source fog node which exchange data with each other in this paper. A target fog node can receive input data from multiple source fog nodes the MSMT protocol while each target fog node can receive data from one source fog node in the DNFCN protocol [8]. In the evaluation, we showed the total energy consumption and the total execution

Multi-source and Multi-target Node Selection

33

time of target fog nodes can be reduced in the MSMT protocol compared with the random (RD) and DNFCN protocols.

References 1. Creeger, M.: Cloud computing: an overview. Queue 7(5), 3–4 (2009) 2. Enokido, T., Ailixier, A., Takizawa, M.: A model for reducing power consumption in peer-to-peer systems. IEEE Syst. J. 4, 221–229 (2010) 3. Enokido, T., Ailixier, A., Takizawa, M.: Process allocation algorithms for saving power consumption in peer-to-peer systems. IEEE Trans. Ind. Electron. 58(6), 2097–2105 (2011) 4. Enokido, T., Ailixier, A., Takizawa, M.: An extended simple power consumption model for selecting a server to perform computation type processes in digital ecosystems. IEEE Trans. Ind. Inf. 10, 1627–1636 (2014) 5. Gima, K., Oma, R., Nakamura, S., Enokido, T., Takizawa, M.: A model for mobile fog computing in the IoT. In: Proceedings of the 22nd International Conference on Network-Based Information Systems (NBiS 2019), pp. 447–458 (2019) 6. Guo, Y., Oma, R., Nakamura, S., Duolikun, D., Enokido, T., Takizawa, M.: A twoway flow model for fog computing. In: Proceedings of the Workshops of the 33rd International Conference on Advanced Information Networking and Applications (WAINA 2019), pp. 612–620 (2019) 7. Guo, Y., Oma, R., Nakamura, S., Enokido, T., Takizawa, M.: Distributed approach to fog computing with auction method. In: Proceedings of IEEE the 34nd International Conference on Advanced Information Networking and Applications (AINA 2020), pp. 268–275 (2020) 8. Guo, Y., Saito, T., Nakamura, S., Enokido, T., Takizawa, M.: A dynamic networkbased fog computing model for energy-efficient IoT. In: Proceedings of the 23rd International Conference on Network-Based Information System (NBiS 2020) (2020) 9. Oma, R., Nakamura, S., Duolikun, D., Enokido, T., Takizawa, M.: An energyefficient model for fog computing in the internet of things (IoT). Internet of Things 1–2, 14–26 (2018) 10. Oma, R., Nakamura, S., Duolikun, D., Enokido, T., Takizawa, M.: Evaluation of an energy-efficient tree-based model of fog computing. In: Proceedings of the 21st International Conference on Network-Based Information Systems (NBiS 2018), pp. 99–109 (2018) 11. Oma, R., Nakamura, S., Enokido, T., Takizawa, M.: A tree-based model of energyefficient fog computing systems in IoT. In: Proceedings of the 12th International Conference on Complex, Intelligent, and Software Intensive Systems (CISIS 2018), pp. 991–1001 (2018) 12. Rahmani, A., Liljeberg, P., Preden, J.-S., Jantsch, A.: Fog Computing in the Internet of Things. Springer (2018)

Epidemic and Topic-Based Data Transmission Protocol in a Mobile Fog Computing Model Takumi Saito1(B) , Shigenari Nakamura2 , Tomoya Enokido3 , and Makoto Takizawa1

2

1 Hosei University, Tokyo, Japan [email protected], [email protected] Tokyo Metropolitan Industrial Technology Research Institute, Tokyo, Japan [email protected] 3 Rissho University, Tokyo, Japan [email protected]

Abstract. In the fog computing (FC) models, a fog node supports application processes to calculate output data on input data from sensors and other fog nodes and sends the output data to target fog nodes which can calculate on the output data. In this paper, we consider the MPSFC (Mobile topic-based PS (publish/subscribe) FC) model where mobile fog nodes communicate with one another by publishing and subscribing messages in wireless networks. Subscription topics of a fog node denote input data on which the fog node can calculate and publication topics of a message show data carried by the message. A fog node only receives a message whose publication topics shares a common topic with the subscription topics. In the TBDT protocol proposed in our previous studies, a fog node only publishes a message of the output data to a target fog node in the communication range. Here, while a fewer number of messages are transmitted, the delivery ratio of messages in the TBDT protocol is smaller than the epidemic routing protocol. In this paper, we propose an ETBDT (Epidemic and Topic-Based Data Transmission) protocol in order to increase the delivery ratio, where mobile fog nodes forward messages to not only target nodes but also non-target nodes in the communication range. In the evaluation, we show the delivery ratio in the ETBDT protocol is larger than the TBTD protocol. Keywords: IoT · Mobile fog computing (MFC) model · ETBDT (Epidemic Topic-Based Data Transmission) protocol · Mobile topic-based publish/subscribe fog computing (MPSFC) model

1

Introduction

Fog computing (FC) models [7,12] to efficiently realize the IoT are composed of fog nodes in addition to server clouds and sensor and actuator devices. Output c The Editor(s) (if applicable) and The Author(s), under exclusive license  to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 34–43, 2021. https://doi.org/10.1007/978-3-030-61108-8_4

Topic-Based Processing Protocol in a Mobile Fog Computing Model

35

data is calculated on sensor data by an application process supported by a fog node and then is sent to other fog nodes to do further calculation on the output data. In the TBFC (Tree-based Fog Computing) model [5,6,8,11], fog nodes are hierarchically structured in a tree to reduce the energy consumption of the fog nodes. In order to make the TBFC model tolerant of faults, the FTBFC (Fault-tolerant TBFC) model is also proposed [9–11]. The MFC (Mobile FC) model [3,4] is composed of mobile fog nodes which communicate with other fog nodes in wireless communication links. Mobile fog nodes communicate with other fog nodes only in the communication range of wireless ad-hoc networks. Thus, mobile fog nodes communicate with other fog nodes by taking advantage of opportunistic routing protocol [2,16]. A PS (Publish/Subscribe) model is a contents-aware, event-driven model of a distributed system [17,18]. In topic-based PS models [15,19], data carried by messages are denoted by topics. We consider the P2PPS (P2P (peer-to-peer) type of topic-based publish/subscribe (PS)) model [17,18] to realize the FC model. Each fog node fi is a peer which can publish a message m with publication topics m.P and subscribe messages by specifying subscription topics fi .S. The subscription topics fi .S denote input data on which the fog node fi can calculate. The publication topics m.P denote output data odi of the source fog node fi . A fog node fi only receives a message m published by a fog node fj whose publication topics m.P and the subscription topics fi .S include some common topic, i.e. m.P ∩ fi .S = φ. Here, the fog node fi is a target fog node of the source node fj . Thus, topics denote data on which a fog node calculates and which a message carries to a target fog node. In the TBDT (Topic-Based Data Transmission) protocol [13,14], each fog node fi sends a message of the output data to only a target node fj in the communication range. Here, while a fewer number of messages are transmitted, the delivery ratio of each message is smaller than the epidemic routing protocol [1]. Fog nodes which receive messages forward the messages to servers. In this paper, we propose an ETBDT protocol where a fog node sends a message of the output data to a fog node in the communication range even if the fog node is not a target node. In the evaluation, we show the number of messages transmitted in the ETBDT protocol is fewer and the delivery ratio of messages is larger than the TBDT protocol [1]. In Sect. 2, we present the MPSFC model. In Sect. 3, we propose the ETBDT protocol in the MPSFC model. In Sect. 4, we evaluate the ETBDT protocol compared with the epidemic routing protocol.

2

Moblile Publish/Subscribe Fog Computing (MPSFC) Model

In this paper, we consider the MPSFC (Mobile topic-based Publish/Subscribe Fog Computing) model [14] to realize the mobile fog computing (MFC) model [3,4] of the IoT by taking advantage of the PS (publish/subscribe) model [17,18]. The MPSFC model is composed of mobile fog nodes which communicate with one another by publishing and subscribing messages in wireless communication

36

T. Saito et al.

Fig. 1. MFC model.

networks. Mobile fog nodes communicate with one another only in the communication range of wireless ad-hoc networks [1]. A mobile fog node is also equipped with sensors and actuators. A fog node collects data from sensors and activates actuators. In addition, a fog node fi supports an application process p(fi ) to calculate output data on the sensor data and sends the output data to other fog nodes. A fog node fi receives input data idj from each source fog node fj . Then, the fog node fi calculates the output data odi on the input data and then forwards the output data odi to other fog nodes [Fig. 1]. A fog node fi specifies subscription topics fi .S which denote input data on which the fog node fi can calculate. A message m is characterized by publication topics m.P which denote output data odj of the source fog node fj . A fog node fi only receives a message m whose m.P ∩ fi .S = φ. This means, the fog node fi supports a process p(fi ) to calculate on data odj carried by the message m. Here, the fog node fi is a target node of the fog node fj and fj is a source fog node of fi (fj → fi ). A fog node fi receives a collection IDi of input data, i.e. IDi = {idij | fi can receive input data idij from fj to fi and fj → fi }. Then, on receipt of messages, the fog node fi calculates output data odi on input data IDi carried by the messages. The output data odi is characterized by publication topics odi .P . The publication topics fi .P of a fog node fi show types of output data of a process p(fi ). The publication topics odi .P indicate types of output data odi calculated by a process p(fi ). Let D be a collection of data in a system. A process p(fi ) supported by a fog node fi is considered to be as a function which uses input data idi1 , ..., idi,li (li ≥ 1) and output data odi , i.e. odi = p(fi ) (idi1 , ..., idi,li ), where idij ∈ D(j = 1, ..., li ) and odi ∈ D. Let Tij (∈ 2T ) be a subset of topics denoting input data idij and Ti (∈ T ) be a subset of topics of output data odi . Here, the process p(fi ) is specified as p(fi ): Ti1 × · · · × Ti,li → Ti . A set Ti1 ∪ · · · ∪ Ti,li of topics gives the subscription topics fi .S and Ti gives the publication topics m.P of a message m carrying output data of a fog node fi , i.e. m.P = fi .P .

Topic-Based Processing Protocol in a Mobile Fog Computing Model

37

A fog node fi precedes a fog node fj (fi → fj ) if the publication topic fi .P of the fog node fi is one of the subscription topics fj .S, i.e. fi .P ∩ fj .S = φ. Then, the process p(fi ) of a fog node fi precedes the process p(fj ) of a fog node fj (p(fi ) ⇒ p(fj )) if and only if (iff) the publication topics fi .P is one of the input sorts of the subscription topics fj .S (Fig. 2).

Fig. 2. Fog node.

A fog node fi can communicate with another fog node fj (fi ↔ fj ) only if the fog node fj is in the communication range of the fog node fi . Let F N (fi ) be a set of fog nodes which are in the communication range of a node fi , i.e. {fj | fi ↔ fj }. Each fog node fi can only communicate with another fog node fj in the set F N (fi ). A message m published by a source fog node fi is only received by a target fog node fj in the communication range, where fi → fj , i.e. m.P ∩ fj .S = φ. Let T N (fi ) (⊆ F ) be a set {fj | fi → fj } of target fog nodes of a fog node fi . Let T F N (fi ) (⊆ F ) be a set of target fog nodes of a fog node fi with which the fog node fi can communicate, i.e. {fj | fi → fj and fj ↔ fi }, i.e. T F N (fi ) = T N (fi ) ∩ F N (fi ). A fog node fi can only deliver a message m to a target fog node fj in the set T F N (fi ).

3

Epidemic and Topic-Based Data Transmission (ETBDT) Protocol

In the TBDT (Topic-based Data Transmission) protocol proposed in our previous papers [13,14], each fog node fi only delivers a message m of output data odi to a target fog node fj only if the target fog node fj is in the communication range, i.e. fi → fj and fi ↔ fj . Here, a fewer number of messages are

38

T. Saito et al.

transmitted but the delivery ratio of messages to the root node is smaller than the epidemic routing protocol [1]. In this paper, in order to increase the delivery ratio of each message, some message is delivered to fog nodes even if the fog nodes are not target nodes. As presented in the preceding section, each fog node fi calculates output data odi and publishes a message m carrying the data odi to a target fog node fj . Finally, a fog node fi publishes a message of output data odi to the root node f , i.e. server. Output data odi to be just delivered to the root node f is referred to as f inal data. In this paper, we propose an ETBDT (Epidemic and Topic-Based Data Transmission) protocol where final data can be forwarded to not only the root node but also non-target fog nodes. Suppose a fog node fj comes in the communication range of a fog node fi which holds final data odi and the fog node fj is not a target node of the node fi (fi  fj ). In the TBDT protocol, the fog node fj does not receive a message m published by the fog node fi since fj  fi . In the ETBDT protocol, the fog node fj receives the message m from the fog node fi if the fog node fj supports enough memory to store the message m. Then, if the fog node fj eventually finds another fog node fk in the communication range, the fog node fj delivers the message m to the fog node fk in a similar way as the epidemic routing protocol [1]. If the fog node fk comes in the communication range of the root node f , the fog node fk delivers the message m of the final data odi to the root node f . First, suppose a fog node fi obtains input data from sensors and calculates output data odi on the sensor data. Then, a fog node fi sends the output data odi to a target fog node fj if the target fog node fj is in the communication range of the source fog node fi (fi ↔ fj ). Suppose a target fog node fj is in the communication range of a fog node fi , i.e. fi ↔ fj . Here, the fog node fi includes the output data odi to a message mi . Thus, a message mi carries a collection of data di1 , . . . , di,li . The fog node fi publishes the message mi to the target fog node fj . Topics denoting data dij are denoted by dij .P . The publication mi .P of the message mi is a set of the publication topics di1 .P, . . . , di,li .P of the data di1 , . . . , di,li carried by the message m. Then, the message mi arrives at the fog node fj . The fog node fj checks if the publication mi .P of the message mi and the subscription fj .S of the fog node fj include a common topic, i.e. mi .P ∩ fj .S = φ. Here, if mi .P ∩ fj .S = φ, the message mi is received by a fog node fj . Then, each data dij in the message mi is stored in the memory fj .M only if dij .P ∩ fj .S = φ. A fog node fi obtains final data odi by calculating on the input data. The final data odi has to be delivered to the root node. Here, the publication odi .P of the final odi is empty, i.e. odi .P = φ. Each fog node fi publishes and receives a message m and calculates output data on input data carried by the message m as follows: [Fog node fi publishes a message mi to a target fog node fj ] 1. A fog node fi finds that a target fog node fj is in the communication range of the fog node fi , i.e. fi ↔ fj ; 2. If the fog node fi has output data odi in the memory fi .M , the fog node fi adds the output data odi in the memory fi .M to a message mi . {odi } ∈ fi .M ;

Topic-Based Processing Protocol in a Mobile Fog Computing Model

39

3. If odi is not final, the publication mi .P is a set of topics odi .P (⊆ T ), i.e. mi .P = odi .P else if odi is final, i.e. mi .P = φ; 4. The fog node fi publishes the message mi to the target fog node fj ; 5. If the fog node fi has final data odk (k = i), where odk .P = φ, the fog node fi publishes the message mk with the final data odk , where mk .P = φ, to the fog node fj ; [Fog node fi receives a message mj from a source fog node fj ] 1. A message mj from a source node fj arrives at a fog node fi ; 2. If mj .P ∩ fi .S = φ or mj .P = φ, i.e. mj carries the final data, the fog node fi receives the message mj , otherwise, the fog nodes fi neglects the message mj ; 3. If a memory fi .M of the fog node fi is full, the fog node fi deletes the data dk which is the oldest and fi .P ∩ dk .T = φ in the memory fi .M ; 4. The fog node fi stores the data odj as th input data idi in the memory fi .M ; In the second step, the condition “mj .P = φ” means that the message m carries final data. [Fog node fi calculates input data idi ] 1. The fog node fi calculates output data odi on the input data idi in the message mj using the process p(fi ); 2. If the process p(fi ) is in the final stage, the topics odi .P of the final output data odi is empty, i.e. odi .P = φ; 3. The output data odi is stored in the memory fi .M ;

4

Evaluation

We evaluate the ETBDT protocol of the MPSFC model in terms of the number of messages exchanged among fog nodes and delivery ratio of messages compared with the TBDT [13,14] protocol and the epidemic routing protocol [1]. There are mobile fog nodes f1 , ..., fn (n ≥ 1) on an m×m mesh M . We assume the distance between a pair of neighboring points is one in the mesh M . Each fog node fi moves on the mesh M in a random walk way. Let cri be the communication range of a fog node fi . Each fog node fi moves with speed si in a random walk. In order to evaluate the ETBDT protocol, we develop a time-based simulator in C language. Let T be a set of all topics t1 , . . . , tl (l ≥ 1) in a system. Let D be also a collection of data in a system. First, each fog node fi has one data di and a topic T (di ) of the data di is randomly taken from the set T . Then, each fog node fi has a process p(fi ). Each fog node fi supports a process p(fi ) : Ti1 × · · · × Ti,li → Ti . The subscription fi .S and the publication fi .P of a fog node fi are a collection {Ti1 , . . . Ti,li } of input sorts and the output sort Ti , respectively. Next, every fog node fi randomly moves in the mesh M with the moving speed si for each simulation step. Let dij show the distance between a pair of fog nodes fi and fj . If each fog node fi finds a fog node fj in the communication

40

T. Saito et al.

range cri , i.e. dij ≤ cri , the fog node fi sends a message mi to the fog node fj . Here, the message m carries the output data odi of the fog node fi . The publication mi .P of the message mi is the publication topics odi .P of the output data odi . Then, the message mi arrives at the fog node fj . The fog node fj receives a message mi and stores output data odi as the input data idi in the memory fj .M . The fog node fj calculates the output data odj on the input data idi by using the process p(fj ). The output data odj is stored in the memory fj .M . Finally, the delivery ratios of messages in the ETBDT protocol, the TBDT protocol, and the epidemic routing protocol [1] are calculated. In the ETBDT and TBDT protocols, on arrival of a message m, the fog node fi checks if the publication topics m.P of the message m and the subscription topics fi .S of the fog node fi include a common topic. Here, if m.P ∩ fi .S = φ, the fog node fi receives a message m and stores the message m in the memory fi .M , i.e. the message m is delivered to the fog node fi . In the ETBDT protocol, even if m.P ∩fi .S = φ, if m.P = φ, the fog node fi also receives a message m and stores final data carried by the message m in the memory fi .M . If m.P ∩fi .S = φ, the fog node fi checks if m.P = φ. If m.P = φ, the message m carries final data. The fog node fi receives the message m. If m.P = φ, the fog node fi neglects the message m. In the epidemic routing protocol [1], if a message m arrives at the fog node fi , the fog node fi receives the message m and stores the message m in the memory fi .M . In the evaluation, we consider fifteen processes p1 , p2 , . . . , p15 . Here, a process is composed of three stages. In the first stage, the process calculates the average value on input data. In the second stage, the process merges input data into output data. In the third stage, the process joins input data into output data. The processes p1 , . . . , p5 do the first stage. The processes p6 , . . . , p10 do the second stage. The processes p11 , . . . , p15 do the third stage. The processes p1 , . . . , p5 have five sorts of input data id1 , . . . , id5 and five sorts of output data od1,1 , . . . , od1,5 . The processes p6 , . . . , p10 also have five sorts of input data id1,1 , . . . , id1,5 and five sorts of output data od1,6 , . . . , od1,10 . The processes p11 , . . . , p15 also have five sorts of input data id1,6 , . . . , id1,10 and five sorts of output data od1,11 , . . . , od1,15 . Then, each fog node fi has a memory fi .M . In the simulator, size of each memory fi .M is 10 and size of each data di is 1. There are thirty mobile fog nodes (n = 30) on a 300 × 300 mesh (m = 300), and twenty topics (l = 20). Each fog node has one process and each process is supported by two fog nodes. Then, the speed si of each fog node fi is 1 for one simulation step and the communication range cri of each fog node fi is 3. Figure 3 shows the delivery ratios of messages in the ETBDT protocol, the TBDT protocol, and the epidemic routing protocol. The delivery ratio of the ETBDT protocol is about 20% larger than the TBDT protocol since final data is forwarded to non-target fog nodes with an epidemic way in the ETBDT protocol. However, the delivery ratio of the ETBDT protocol is still smaller than the epidemic routing protocol. Figure 4 shows the number of messages exchanged among fog nodes in the ETBDT protocol, the TBDT protocol, and the epidemic routing protocol. The

Topic-Based Processing Protocol in a Mobile Fog Computing Model

41

Fig. 3. Delivery ratio.

Fig. 4. Number of messages exchanged.

numbers of messages in the ETBDT protocol, the TBDT protocol and the epidemic routing protocol linearly increase for each simulation step. The larger number of messages are exchanged among fog nodes in the epidemic routing protocol than the ETBDT protocol while more number of messages are transmitted in the ETBDT protocol than the TBDT protocol.

5

Concluding Remarks

In this paper, we considered the MFC (Mobile Fog Computing) model [3,4] to efficiently realize the IoT, where mobile fog nodes like vehicles communicate with other nodes in wireless networks. Here, each fog node calculates the output data on the input data received from other fog nodes and forwards the output data to target fog nodes in the epidemic routing way. In the TBDT protocol [13,14],

42

T. Saito et al.

each fog node delivers a message with output data to only a target fog node. The number of messages can be reduced in the TBDT protocol but the delivery ratio of messages is smaller in the TBDT protocol than the epidemic routing protocol. In this paper, we newly proposed the ETBDT (Epidemic and TopicBased Data Transmission) protocol in the MPSFC (Mobile Publish/Subscribe Fog Computing) model of fog nodes where each fog node fi is delivered not only messages of data on which the fog node fi can calculate by taking advantage of the topic-based PS model but also data on which the fog node fi cannot calculate. In the ETBDT protocol, fog nodes forward final data to even nontarget fog nodes in the communication range in order to increase the delivery ratio of messages to the root node. In the evaluation, we showed the number of messages can be reduced in the ETBDT protocol compared with the TBDT and epidemic routing protocols.

References 1. Amin, V., David, B.: Epidemic routing for partially-connected adhoc networks. Technical Report (2000) 2. Dhurandher, S.K., Sharma, D.K., Woungang, I., Saini, A.: An energy-efficient history-based routing scheme for opportunistic networks. Int. J. Commun. Syst. 30(7) (2015) 3. Gima, K., Oma, R., Nakamura, S., Enokido, T., Takizawa, M.: A model for mobile fog computing in the IoT (accepted). In: Proceedings of the 22nd International Conference on Network-Based Information Systems (NBiS 2019) (2019) 4. Gima, K., Oma, R., Nakamura, S., Enokido, T., Takizawa, M.: Parallel data transmission protocols in the mobile fog computing model. In: Proceedings of the 14th International Conference on Broad-Band Wireless Computing, Communication and Applications (BWCCA 2019), pp. 494–503 (2019) 5. Guo, Y., Oma, R., Nakamura, S., Duolikun, D., Enokido, T., Takizawa, M.: Data and subprocess transmission on the edge node of TWTBFC model. In: Proceedings of the 11th International Conference on Intelligent Networking and Collaborative Systems (INCoS 2019), pp. 80–90 (2019) 6. Guo, Y., Oma, R., Nakamura, S., Duolikun, D., Enokido, T., Takizawa, M.: Evaluation of a two-way tree-based fog computing (TWTBFC) model. In: Proceedings of the 13th International Conference on Innovative Mobile and Internet Services in Ubiquitous Computing (IMIS 2019), pp. 72–81 (2019) 7. Oma, R., Nakamura, S., Duolikun, D., Enokido, T., Takizawa, M.: An energyefficient model for fog computing in the internet of things (IoT). Internet of Things 1–2, 14–26 (2018) 8. Oma, R., Nakamura, S., Duolikun, D., Enokido, T., Takizawa, M.: Evaluation of an energy-efficient tree-based model of fog computing. In: Proceedings of the 21st International Conference on Network-Based Information Systems (NBiS 2018), pp. 99–109 (2018) 9. Oma, R., Nakamura, S., Duolikun, D., Enokido, T., Takizawa, M.: Energy-efficient recovery algorithm in the fault-tolerant tree-based fog computing (FTBFC) model. In: Proceedings of the 33rd International Conference on Advanced Information Networking and Applications (AINA 2019), pp. 132–143 (2019)

Topic-Based Processing Protocol in a Mobile Fog Computing Model

43

10. Oma, R., Nakamura, S., Duolikun, D., Enokido, T., Takizawa, M.: A fault-tolerant tree-based fog computing model (accepted). Int. J. Web Grid Serv. (IJWGS) (2019) 11. Oma, R., Nakamura, S., Enokido, T., Takizawa, M.: A tree-based model of energyefficient fog computing systems in IoT. In: Proceedings of the 12th International Conference on Complex, Intelligent, and Software Intensive Systems (CISIS 2018), pp. 991–1001 (2018) 12. Rahmani, A., Liljeberg, P., Preden, J.-S., Jantsch, A.: Fog Computing in the Internet of Things. Springer (2018) 13. Saito, T., Nakamura, S., Enokido, T., Takizawa, M.: Topic-based processing protocol in a mobile fog computing model. In: Proceedings of the 23nd International Conference on Network-Based Information Systems (NBiS 2020) (2020, accepted) 14. Saito, T., Nakamura, S., Enokido, T., Takizawa, M.: A topic-based publish/subscribe system in a fog computing model for the IoT. In: Proceedings of the 14th International Conference on Complex, Intelligent and Software Intensive Systems (CISIS 2020), pp. 12–21 (2020) 15. Setty, V., van Steen, M., Vintenberg, R., Voulgais, S.: PolderCast: fast, robust, and scalable architecture for P2P topic-based pub/sub. In: Proceedings of ACM/IFIP/USENIX 13th International Conference on Middleware (Middleware 2012), pp. 271–291 (2012) 16. Spaho, E., Barolli, L., Kolici, V., Lala, A.: Evaluation of single-copy and multiplecopy routing protocols in a realistic VDTN scenario. In: Proceedings of the 10th International Conference on Complex, Intelligent, and Software Intensive Systems (CISIS 2016), pp. 285–289 (2016) 17. Tarkoma, S.: Publish/Subscribe System: Design and Principles, 1st edn. Wiley, Hoboken (2012) 18. Tarkoma, S., Rin, M., Visala, K.: The publish/subscribe internet routing paradigm (PSIRP): designing the future internet architecture. In: Future Internet Assembly, pp. 102–111 (2009) 19. Yamamoto, Y., Hayashibara, N.: Merging topic groups of a publish/subscribe system in causal order. In: Proceedings of the 31st International Conference on Advanced Information Networking and Applications Workshops (WAINA 2017), pp. 172–177 (2017)

The Energy-Efficient Object Replication by Excluding Meaningless Methods in Virtual Machine Environments Tomoya Enokido1(B) and Makoto Takizawa2 1

Faculty of Business Administration, Rissho University, 4-2-16, Osaki, Shinagawa-ku, Tokyo 141-8602, Japan [email protected] 2 Department of Advanced Sciences, Faculty of Science and Engineering, Hosei University, 3-7-2, Kajino-cho, Koganei-shi, Tokyo 184-8584, Japan [email protected] Abstract. In order to realize various types of distributed application services, various kinds of data are gathered from various types of devices. Each gathered data unit is stored in an object like database systems. Each object is replicated on multiple virtual machines in a system to provide reliable and available application services. However, a large mount of electric energy is consumed since replicas of each object are manipulated in multiple virtual machines in a system. In this paper, the EIEEQS-VM (Extended IEEQS-VM) algorithm is newly proposed to reduce the total electric energy consumption of servers by omitting meaningless read and write methods. Keywords: Meaningless methods · Server cluster Quorum · Virtual machines · Green computing

1

· Replication ·

Introduction

There are various kinds of distributed applications like vehicle network services [1]. In these distributed applications, various kinds of data like humidity and temperature are gathered from various types of devices like sensors and smartphones. Each gathered data unit is stored in an object [2] like database systems. An object is an encapsulation of data and methods to manipulate the data in the object. A distributed application is composed of large number of objects and each object is shared by a huge number of application processes. Hence, scalable and fault-tolerant computing systems like cloud computing systems [3–6] are required to realize current distributed application services. In order to realize a cloud computing system, a server cluster system supporting virtual machines [3–5] are widely used. Here, each object is replicated to multiple virtual machines in a server cluster system to provide reliable and available application services. A transaction is an atomic sequence of methods to manipulate objects. Conflicting transactions have to be serialized to keep replicas of each object mutually c The Editor(s) (if applicable) and The Author(s), under exclusive license  to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 44–54, 2021. https://doi.org/10.1007/978-3-030-61108-8_5

The EIEEQS-VM Algorithm

45

consistent [7–9]. The quorum-based locking (QBL) protocol [10] is proposed to keep replicas of each object mutually consistent. However, the total amount of electric energy consumption is larger than the non-replication systems since each method issued by a transaction to manipulate an object has to be performed on multiple virtual machines which hold replicas of the object. In our previous studies, the IEEQS-VM (Improved Energy-Efficient Quorum Selection with Virtual Machines) algorithm [5] is proposed to reduce the total electric energy consumption of a system by omitting meaningless write methods which are not required to be performed on each replica of an object. In this paper, we additionally define meaningless read methods based on the precedent relation and semantics of methods. Then, the EIEEQS-VM (Extended IEEQS-VM) algorithm is proposed to furthermore reduce the total electric energy consumption of servers by omitting meaningless read and write methods. We evaluate the EIEEQS-VM algorithm compared with the IEEQS-VM algorithm. The evaluation results show the total electric energy consumption of a server cluster, the average execution time of each transaction, and the average number of aborted transaction instances in the EIEEQS-VM algorithm can be more reduced than the IEEQS-VM algorithm. In Sect. 2, we discuss the system model of this paper. In Sect. 3, we propose the EIEEQS-VM algorithm. In Sect. 4, we evaluate the EIEEQS-VM algorithm compared with the IEEQS-VM algorithm.

2 2.1

System Model Object Replication

A server cluster S is composed of multiple servers s1 , ..., sn (n ≥ 1). Each server st is equipped with a multi-core CPU. Each server st holds a set Ct of cores c1t , ..., cnct t where nct (≥1) is the total number of cores in the server st . Let ctt (ctt ≥ 1) be the total number of threads on each core clt in a server st . Each server st holds a set T Ht of threads th1t , ..., thntt t where ntt (ntt ≥ 1) is the total number of threads in a server st , i.e. ntt = nct ·ctt . Threads th(h−1)·ctt +1 , ..., thh·ctt (1 ≤ h ≤ nct ) are performed on a core clt . A set Vt of virtual machines V M1t , ..., V Mntt t is installed in a server st . In this paper, we assume each virtual machine V Mkt is exclusively performed on one thread thkt in a server st . Let O denote a set of objects o1 , ..., om (m ≥ 1) [2]. Each object oh is an encapsulation of data and methods to manipulate the data in the object oh . In this paper, methods are classified into read (r) and write (w) methods. Write methods are furthermore classified into full write (wf ) and partial write (wp ) methods, i.e. w ∈ {wf , wp }. In a full write method, a whole data in a object is fully written while a part of data is written in a partial write method. A notation op(oh ) indicates a state obtained by performing a method op on an object oh . A composition op1 ◦ op2 means that a method op1 is performed before another method op2 . A pair of methods op1 and op2 on an object oh are compatible if and only if (iff) op1 ◦ op2 (oh ) = op2 ◦ op1 (oh ) for every state of the object oh . Otherwise, a method

46

T. Enokido and M. Takizawa

op1 conflicts with another method op2 . In this paper, a pair of read methods are compatible. Otherwise, a pair of methods are conflict. The conflicting relation is symmetric. Each object oh is replicated on multiple virtual machines in a server cluster S to provide a reliable and available application service. Let R(oh ) be a set of replicas o1h , ..., odh (1 ≤ d ≤ n) [9] of an object oh . nR(oh ) = |R(oh )|. Let SVh be a subset of virtual machines which hold a replica of an object oh (SVh ⊆ Vt ). A transaction is an atomic sequence of methods [7]. Each execution of a transaction is referred to as transaction instance. A transaction T i issues read (r) and write (w) methods to manipulate replicas of objects. Multiple conflicting transactions have to be serializable [7,8] to keep replicas of each object mutually consistent. Let T be a set of transactions T 1 , ..., T v (v ≥ 1). A transaction T i precedes another transaction T j (T i →sch T j ) in a schedule sch iff a method opi issued by the transaction T i is performed before a method opj issued by the transaction T j and opi conflicts with opj . A schedule sch is a partially ordered set T, →sch of transactions. A schedule sch of T is serializable iff the precedent relation →sch (⊆ T2 ) is acyclic [7]. The quorum-based locking (QBL) protocol [10] is proposed to serialize multiple conflicting transactions. A quorum Qop h (⊆ R(oh )) is a subset of replicas of an object oh to be locked by a method op (op ∈ {r, w}). op Let nQop h (= |Qh |) be the quorum number of a method op on a object oh . The quorums have to satisfy the following constraints: (1) Qrh ⊆ R(oh ), Qw h ⊆ R(oh ), r w r w = R(o ). (2) nQ + nQ > nR(o ), i.e. Q ∩ Q and Qrh ∪ Qw h h h h h h h = φ. (3) > nR(o )/2. nQw h h The QBL Protocol. A transaction T i locks replicas of an object oh by the following procedure [10]: op 1. A quorum Qop h for a method op is constructed by selecting nQh replicas in a set R(oh ) of replicas. 2. If every replica in the quorum Qop h can be locked by the method op, the replicas in the quorum Qop h are manipulated by the method op. 3. When the transaction T i commits or aborts, the locks on the replicas in the quorum Qop h are released.

At least one newest replica is surely included in every quorum in the QBL protocol. Each replica oqh holds a version number vhq . Suppose a transaction T i reads an object oh . The transaction T i selects nQrh replicas in the set R(oh ), i.e. read(r) quorum Qrh . If every replica in the r-quorum Qrh can be locked, the transaction T i reads data in a replica oqh whose version number vhq is the maximum in the r-quorum Qrh . Next, suppose a transaction T i writes data in an object oh . The transaction T i selects nQw h replicas in the set R(oh ), i.e. write(w) . If every replica in the w-quorum Qw quorum Qw h h can be locked, the transaction T i writes data in a replica oqh whose version number vhq is maximum in the wq q quorum Qw h and the version number vh of the replica oh is incremented by one. q The updated data and version number vh of the replica oqh are sent to every other replica in the w-quorum Qw h . Then, data and version number of each replica in are replaced with the newest values. the w-quorum Qw h

The EIEEQS-VM Algorithm

2.2

47

Data Access Model

The DAVM (Data Access model for Virtual Machine environments) model [3] is proposed to perform read and write methods on multiple virtual machines in a i i (oqh ) and wkt (oqh ) be read (r) and write server st in our previous studies. Let rkt i (w) methods issued by a transaction T to read and write data of a replica oqh stored in a virtual machine V Mkt in a server st , respectively. Methods which are being performed on a virtual machine V Mkt are current at time τ . Let RMkt (τ ) and W Mkt (τ ) be sets of current r and w methods on a virtual machine V Mkt at time τ , respectively. Let RMt (τ ) and W Mt (τ ) be sets of current ntt r and w RMkt (τ ) methods on a server st at time τ , respectively, i.e. RMt (τ ) = k=1 ntt and W Mt (τ ) = k=1 W Mkt (τ ). A notation Mt (τ ) shows a set of current r and w methods on a server st at time τ , i.e. Mt (τ ) = RMt (τ ) ∪ W Mt (τ ). In i (oqh ), data in a replica oqh stored in a virtual machine V Mkt each method rkt i i is read at rate RRkt (τ ) [B/sec] at time τ . In each method wkt (oqh ), data is q i (τ ) [B/sec] written in a replica oh stored in a virtual machine V Mkt at rate W Rkt at time τ . Let maxRRt and maxW Rt be the maximum read and write rates i (τ ) (≤ maxRRt ) and [B/sec] of a server st , respectively. The read rate RRkt i write rate W Rkt (τ ) (≤ maxW Rt ) are f rt (τ ) · maxRRt and f wt (τ ) · maxW Rt , respectively. Here, f rt (τ ) and f wt (τ ) are degradation ratios. 0 ≤ f rt (τ ) ≤ 1 and 0 ≤ f wt (τ ) ≤ 1. The degradation ratios f rt (τ ) and f wt (τ ) are 1 / (|RMt (τ )| + rwt · |W Mt (τ )|) and 1 / (wrt · |RMt (τ )| + |W Mt (τ )|), respectively. 0 ≤ rwt ≤ 1 and 0 ≤ wrt ≤ 1. 2.3

Electric Power Consumption Model

The PCDAVM (Power Consumption model for Data Access in Virtual Machine environment) model [3] is proposed to perform read and write methods on multiple virtual machines in a server st in our previous studies. Notations maxEt and minEt show the maximum and minimum electric power [W] of a server st , respectively. Let act (τ ) and att (τ ) be the number of active cores and the number of active threads in a server st at time τ , respectively. Let minCt be the electric power [W] where at least one core cht is active on a server st . A notation cEt shows the electric power [W] consumed by a server st to make one core active. A notation tEt indicates the electric power [W] consumed by a server st to make one thread active. The base electric power BEt (τ ) of a server st at time τ is given as formula (1): BEt (τ ) = minEt + γt · (minCt + act (τ ) · cEt + att (τ ) · tEt )[W ].

(1)

If at least one virtual machine is active, γt = 1. Otherwise, γt = 0. The base electric power BEt (τ ) [W] of a server st depends on the numbers of active threads and cores in the server st at time τ . Let Et (τ ) be the electric power [W] consumed by a server st at time τ . A notation REt [W] shows the electric power of a server st to perform only read methods on multiple virtual machines in the server st . A notation W Et [W] shows the electric power of a server st to perform only write methods on multiple

48

T. Enokido and M. Takizawa

virtual machines in the server st . The electric power Et (τ ) [W] consumed by a server st at time τ to perform read and write methods on multiple virtual machines is given as formula (2): ⎧ BEt (τ ) + W Et if |W Pt (τ )| ≥ 1 and |RPt (τ )| = 0. ⎪ ⎪ ⎪ ⎨BE (τ ) + W RE (α(τ )) if |W P (τ )| ≥ 1 and |RP (τ )| ≥ 1. t t t t Et (τ ) = (2) ⎪ (τ ) + RE if |W P (τ )| = 0 and |RP (τ )| ≥ 1. BE t t t t ⎪ ⎪ ⎩ if |W Pt (τ )| = |RPt (τ )| = 0. BEt (τ ) A server st consumes the minimum electric power minEt [W] if no method is performed on the server st , i.e. the electric power consumed in the idle state. The server st consumes the electric power BEt (τ ) + REt [W] if only r methods are performed on multiple virtual machines in the server st . The server st consumes the electric power BEt (τ ) + W Et [W] if only w methods are performed on multiple virtual machines in the server st . The server st consumes the electric power BEt (τ ) + W REt (α(τ )) = (1 − α(τ )) · REt + α(τ ) · W Et [W] where α(τ ) = |W Pt (τ )|/|(W Pt (τ ) + RPt (τ ))| if both at least one r method and at least one w method are concurrently performed. Here, REt ≤ W REt (α(τ )) ≤ W Et . The processing power P Et (τ ) [W] of a server st at time τ is Et (τ ) − minEt . The total processing electric energy T P Et (τ1 , τ2 ) [J] of a server st from time τ1 to τ2 is given as Σττ2=τ 1 P Et (τ ). The total processing electric energy laxity tpeelt (τ ) shows how much electric energy a server st has to consume to perform all current r and w methods performed on every virtual machine in the server st at time τ . The T P ECLt algorithm [6] is proposed to obtain the total processing electric energy laxity tpeelt (τ ) of a server st at time τ where r and w methods are performed on the server st in our previous studies. In this paper, the T P ECLt algorithm is used to obtain the total processing electric energy laxity tpeelt (τ ) of a server st at time τ .

3

Energy-Efficient Object Replication Algorithm

In our previous studies, the EEQS-VM (Energy-Efficient Quorum Selection with Virtual Machines) algorithm [4] is proposed to select members of a quorum for each method so that the total electric energy consumption of a server cluster to perform read (r) and write (w) methods on replicas of each object can be reduced. In addition, the meaningless write methods [5] are defined, which are not required to be performed on each replica of an object based on the precedent relation and semantics of methods. Then, the IEEQS-VM (improved EEQS-VM) algorithm [5] is proposed to reduce the total electric energy consumption of servers by omitting meaningless write methods. In this paper, we additionally define meaningless read methods. Next, the EIEEQS-VM (Extended IEEQS-VM) algorithm is proposed to furthermore reduce the total electric energy consumption of servers by omitting meaningless read and write methods. In the EIEEQS-VM algorithm, members of a quorum for each method are selected by the EEQS-VM algorithm [4].

The EIEEQS-VM Algorithm

49

A method op1 precedes another method op2 in a schedule H(op1 →H op2 ) iff 1) the method op1 is issued before op2 by the same transaction T i , 2) the method op1 issued by a transaction T i conflicts with the method op2 issued by a transaction T j and T i →H T j , or 3) op1 →H op3 →H op2 for some method op3 . Let Hh be a local schedule of methods which are performed on an object oh in a schedule H. Definition. A method op1 locally precedes another method op2 in a local schedule Hh of an object oh (op1 →Hh op2 ) iff op1 →H op2 . Definition – A read method op1 absorbs another read method op2 in a local subschedule Hh on an object oh if op1 →Hh op2 and there is no write method op such that op1 →Hh op →Hh op2 , or op1 absorbs op and op absorbs op2 for some method op . – A full write method op2 absorbs another partial or full write method op1 in a local subschedule Hh on an object oh if op1 →Hh op2 and there is no read method op such that op1 →Hh op →Hh op2 , or op2 absorbs op and op absorbs op1 for some method op . Definition. A method op is meaningless iff the method op is absorbed by another method op in the local subschedule Hh of an object oh . r i (o ) Suppose a transaction T i issues a read method ri (oh ) to a quorum Qh h for manipulating an object oh . The read method ri (oh ) is performed on a replica oqh r i (o )

whose version number is the maximum in the quorum Qh h . Suppose another read method rj (oh ) issued by another transaction T j is concurrently performed on the replica oqh and ri (oqh ) →Hh rj (oqh ). Here, the read method rj (oqh ) issued by the transaction T j is meaningless since the read method ri (oqh ) issued by the transaction T i is being performed on the replica oqh and ri (oqh ) absorbs the read method rj (oqh ). Hence, the read method rj (oqh ) is not performed on the replica oqh and a result obtained by performing the read method ri (oqh ) is sent to a pair of transaction T i and T j . This means the meaningless read method rj (oqh ) is omitted on the replica oqh . Let oqh .CR be a read method rti (oqh ) issued by a transaction T i , which is being performed on a replica oqh in a server st . Let oqh .DW be a write method wti (oqh ) issued by a transaction T i to replace data of a replica oqh in a server st with the updated data dh , which is waiting for the next method op to be performed on the replica oqh . Suppose a transaction T i issues a method op to manipulate an object oh . In the EIEEQS-VM algorithm, the method op is performed on the replica oqh whose version number is the maximum in a quorum Qop h by the EIEEQS-VM Perform procedure as shown in Algorithm 1.  Each time a replica oqh in a write quorum Qw h receives the newest version  q number vh and updated data dh from another replica oqh , the replica oqh manipulate the version number vhq and updated data dh by the IEEQS-VM Replace procedure [5] as shown in Algorithm 2:

50

T. Enokido and M. Takizawa

Algorithm 1 : The EIEEQS-VM Perform procedure Input: op(oqh ) procedure EIEEQS-VM Perform(op(oqh )) if op(oqh ) = r then if oqh .DW = φ then if oqh .CR = φ then oqh .CR = op(oqh ); perform(op(oqh )); oqh .CR = φ; else  /* a read method is omitted */ a result of oqh .CR is sent to a transaction T i end if else  /* oqh .DW = φ */ q q perform(oh .DW ); oh .DW = φ; oqh .CR = op(oqh ); perform(op(oqh )); oqh .CR = φ; end if else  /* op(oqh ) = w */ q q vh = vh + 1; if oqh .DW = φ and oqh .DW is not meaningless then perform(oqh .DW ); perform(op(oqh )); oqh .DW = φ; else  /* oqh .DW = φ or oqh .DW method is omitted */ perform(op(oqh ));  q vh and updated data dh are sent to every replica oqh in a quorum Qop h ; end if end if end procedure

Algorithm 2 : The EIEEQS-VM Perform procedure Input: vhq , dh , wti (oqh ) procedure IEEQS-VM Replace(vhq , dh , wti (oqh ))  vhq = vhq ;  if oqh .DW = φ then  oqh .DW = wti (oqh ); else  if wti (oqh ) absorbs oqh .DW then  oqh .DW = wti (oqh ); else   perform(oqh .DW ); oqh .DW = wti (oqh ); end if end if end procedure

4

Evaluation

We evaluate the EIEEQS-VM algorithm in terms of the total processing electric energy consumption of a server cluster S, the average execution time of each

The EIEEQS-VM Algorithm

51

transaction, and the average number of aborted transaction instances compared with the IEEQS-VM [5] algorithm. In this evaluation, we consider a homogeneous server cluster S which is composed of ten servers s1 , ..., s10 (n = 10). Every server st (t = 1, ..., 8) in the server cluster S follows the same DAVM model and PCDAVM model as shown in Tables 1 and 2, respectively. The parameters of each server st are given based on the experimentations [3]. There are fifty objects o1 , ..., o50 (m = 50). The size of data in each object oh is randomly selected between 50 and 100 [MB]. There are four replicas of each object. Replicas of each object are randomly distributed on five virtual machines which are performed on different servers in the server r cluster S. The quorum numbers nQw h and nQh on every object oh are three, respectively. Table 1. Parameters of a server st for the DAVM model. maxRRt

maxW Rt

wrt

rwt

98.5 [MB/sec] 85.3 [MB/sec] 0.077 0.667

Table 2. Parameters of a server st for the PCDAVM model. minEt minCt

cEt

tEt

W Et

REt

maxEt

17 [W] 1.1 [W] 0.6 [W] 0.5 [W] 4 [W] 1 [W] 24.3 [W]

The number numT of transactions (0 ≤ numT ≤ 500) are issued to manipulate objects. Each transaction issues three methods randomly selected from one-hundred fifty methods on the fifty objects. The total amount of data of an object oh is fully read and written by each read (r) and full write (wf ) methods, respectively. On the other hand, a half size of data of an object oh is written by each partial write (wp ) method. The starting time of each transaction T i is randomly selected in a unit of one second between 1 and 360 [sec]. Figure 1 shows the average total processing electric energy consumption of the server cluster S to perform the total number numT of transactions in the EIEEQS-VM and IEEQS-VM algorithms. The average total processing electric energy consumption of the server cluster S can be more reduced in the EIEEQSVM algorithm than the EEQS-VM algorithm for 0 ≤ numT ≤ 500. In the EIEEQS-VM algorithm, not only the meaningless write methods but also the meaningless read methods are omitted. As a result, the average total processing electric energy consumption of the server cluster S can be more reduced in the EIEEQS-VM algorithm than the IEEQS-VM algorithm. Figure 2 shows the average execution time [sec] of each transaction to perform the total number numT of transactions in the EIEEQS-VM and IEEQS-VM algorithms. For 0 < numT ≤ 500, the average execution time of each transaction can be more shorter in the EIEEQS-VM algorithm than the IEEQS-VM

52

T. Enokido and M. Takizawa

Total processing electric energy [KJ]

35000 EIEEQS-VM 30000

IEEQS-VM

25000 20000 15000 10000 5000 0 0

100

200

300

400

500

Number numT of transactions

Fig. 1. Average total processing electric energy consumption [KJ].

algorithm. In the EIEEQS-VM algorithm, each transaction can commit without waiting for performing not only the meaningless write methods but also meaningless read methods. Hence, the average execution time of each transaction can be more reduced in the EIEEQS-VM algorithm than the IEEQS-VM algorithm. In this evaluation, a transaction T i aborts if the transaction T i could not lock every replica in a quorum for each method. Then, the transaction T i is restarted after δ time units. The time units δ [sec] is randomly selected between twenty and thirty seconds in this evaluation. Every transaction T i is restarted until the transaction T i commits. Each execution of a transaction is referred to as transaction instance. Figure 3 shows the average number of aborted transaction instances for each transaction to perform the total number numT of transactions in the EIEEQS-VM and IEEQS-VM algorithms. For 0 < numT ≤ 500, the average number of aborted instances for each transaction can be more reduced in the EIEEQS-VM algorithm than the IEEQS-VM algorithm. The average execution time of each transaction can be shorter in the EIEEQS-VM algorithm than the IEEQS-VM algorithm. As a result, the number of aborted transactions can be more reduced in the EIEEQS-VM algorithm than the IEEQS-VM algorithm since the number of transaction to be concurrently performed can be more reduced in the EIEEQS-VM algorithm than the IEEQS-VM algorithm. Following the evaluation, the total processing electric energy consumption of a server cluster, the average execution time of each transaction, and the number of aborted transaction instances in the EIEEQS-VM algorithm can be more reduced than the IEEQS-VM algorithm, respectively. Hence, the EIEEQS-VM algorithm is more useful than the IEEQS-VM algorithm.

20 EIEEQS-VM IEEQS-VM

15

10

5

0 0

100

200

300

400

500

Number numT of transactions

Fig. 2. Average execution time [sec] of each transaction.

5

Average number of aborted transaction instances

Average response time of each transaction [sec]

The EIEEQS-VM Algorithm

53

25 EIEEQS-VM 20

IEEQS-VM

15

10

5

0 0

100

200

300

400

500

Number numT of transactions

Fig. 3. Average number of aborts for each transaction.

Concluding Remarks

In this paper, the EIEEQS-VM algorithm is newly proposed to furthermore reduce the total electric energy consumption of a server cluster to manipulate replicas of each object by omitting meaningless read and write methods. The evaluation results show the total electric energy consumption of a server cluster, the average execution time of each transaction and the average number of aborted transactions can be more reduced in the EIEEQS-VM algorithm than the IEEQS-VM algorithm [5] previously proposed. Following the evaluation results, the EIEEQS-VM algorithm is more useful than the IEEQS-VM algorithm.

References 1. Bylykbashi, K., Spaho, E., Barolli, L., Xhafa, F.: Routing in a many-to-one communication scenario in a realistic VDTN. J. High Speed Netw. 24(2), 107–118 (2018) 2. Object Management Group Inc.: Common object request broker architecture (CORBA) specification, version 3.3, part 1 – interfaces. (2012). http://www.omg. org/spec/CORBA/3.3/Interfaces/PDF 3. Enokido, T., Takizawa, M.: The power consumption model of a server to perform data access application processes in virtual machine environments. In: Proceedings of the 34th International Conference on Advanced Information Networking and Applications (AINA-2020), pp. 184–192 (2020) 4. Enokido, T., Takizawa, M.: Energy-efficient quorum-based locking protocol in virtual machine environments. In: Proceedings of the 14th International Conference on Complex, Intelligent,and Software Intensive Systems (CISIS-2020), pp. 22–30 (2020) 5. Enokido, T., Takizawa, M.: The energy-efficient object replication scheme by omitting meaningless write methods in virtual machine environments. In: Accepted for publication in Proceedings of The 23rd International Conference on Network-Based Information Systems (NBiS-2020) (2020)

54

T. Enokido and M. Takizawa

6. Enokido, T., Duolikun, D., Takizawa, M.: Energy consumption laxity-based quorum selection for distributed object-based systems. Evol. Intell. 13, 71–82 (2020) 7. Bernstein, P.A., Hadzilacos, V., Goodman, N.: Concurrency Control and Recovery in Database Systems. Addison-Wesley, Boston (1987) 8. Gray, J.N.: Notes on database operating systems. Lect. Notes Comput. Sci. 60, 393–481 (1978) 9. Schneider, F.B.: Replication management using the state-machine approach. In: Distributed Systems, 2nd edn. ACM Press (1993) 10. Tanaka, K., Hasegawa, K., Takizawa, M.: Quorum-based replication in objectbased systems. J. Inf. Sci. Eng. 16(3), 317–331 (2000)

Experiences with a Single-Page Application for Learning Programming Minoru Uehara(B) Toyo University, Kawagoe, Saitama, Japan [email protected]

Abstract. In recent years, the need for programming has increased. Self-study in programing is important for mastering it. However, many sites for learning programming only support personal computers. JavaScript Development Environment (JDE) was developed in our previous research. JDE is a single-page application that allows a user to learn programming on a smartphone. It imposes a small load on the server, thus many people can use it at the same time. Because it does not use a computer, it can be used in a classroom. Furthermore, JDE is suitable for distance learning (or e-learning). The number of remote classes has increased in Japan in 2020, as a result of the coronavirus pandemic. In this paper, we introduce an example of distance learning using JDE and verify the effectiveness of JDE.

1 Introduction Recently, the necessity of programming education has been increasing [1]. Programming education is a part of science, technology, engineering, and mathematics (STEM) education. Among STEM subjects, programming education has a substantial impact [2]. Programs can visualize teaching materials and can explain the causal relationship between phenomena. In primary education, many teaching materials have been developed and are actually used, using visual languages such as Scratch and Google Blockly. However, there are few examples of its use in higher education institutions, such as universities. There are several reasons for this. (R1) Most programmers develop applications using text-based languages, rather than visual languages. Therefore, most of the teaching materials mentioned above are not suitable. (R2) Many sites for learning programming have limited application. Programmers learn new languages on such sites, and many of them provide remote execution environments. However, in many cases, those execution environments only support simple input/output (I/O). They are fully usable to teach grammar and basic concepts, but they cannot teach advanced contents. More advanced contents are more difficult to learn using these environments. © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 55–66, 2021. https://doi.org/10.1007/978-3-030-61108-8_6

56

M. Uehara

(R3) The cost of providing a learning environment is high if the application is not restricted. Because it is necessary to construct a dedicated learning environment on its own, its development costs are added to the tuition fees. It is difficult to provide services when the number of learners cannot be predicted. For example, when using graphics other than HTML or using a special device, then remote execution is difficult. (R4) Otherwise, learners will need to prepare their own learning environment. If the learners’ skills are inadequate, then they will fail to set up the environment in the early stages of learning. Many universities avoid these methods for reasons R1 or R3. However, reasons R2 and R4 can be partially relaxed. Many universities, despite R2, use learning sites partially as teaching materials for home study. Additionally, despite R4, a classroom equipped with personal computers (PCs) provides an exercise environment that is difficult for students to construct. In our previous research [15], we developed the JavaScript Development Environment (JDE). JDE is a single-page application (SPA) that allows a user to learn programming on a smartphone. It imposes a small load on the server, so many people can use it at the same time. Because it does not use a PC, it can be used in regular classrooms (those that are not equipped with PCs). That is, problems R2 and R4 (above) can be partially solved at the same time. The number of remote classes has increased in Japan in 2020, as a result of the coronavirus (COVID-19) pandemic. JDE is suitable for distance learning. In this paper, we introduce an example of distance learning using JDE and verify the effectiveness of JDE. The paper is organized as follows. Section 2 describes related research. Section 3 describes JDE in detail. Section 4 introduces actual classes using JDE. Section 5 describes the experience gained through the lessons. We conclude in Sect. 6.

2 Related Work In this section, we describe related work. 2.1 Web-Based Integrated Development Environments Web-based integrated development environments (Web IDEs) are increasing in popularity because of the spread of cloud computing. Arvue [3] is a Java-based Web IDE. Eclipse Che is based on Eclipse. Adinda [4] is collaborative. Cloud9 can be used as a web IDE. However, Cloud9 is basically an editor and does not include a languagespecific development kit. Yanagisawa et al. are developing a web-based programming environment [5–9]. Web IDEs are also available on Chromebooks. However, they are not suitable for smartphones because of their small screens. Furthermore, some web-based editors do not display a virtual keyboard, thus they are not available for smartphones.

Experiences with a Single-Page Application for Learning Programming

57

2.2 Code-Sharing Sites Code-sharing sites provide the functions of code sharing and code execution. However, not all code-sharing sites provide these features; they can be classified into three types depending on the combination. (1) Sites that just share code (e.g., GitHub). (2) Sites that can execute shared code (e.g., CodePen and JSFiddle.net). CodePen specializes in front-end XX. (3) Sites on which code can be executed but cannot be shared (e.g., Paiza.io). Of the above systems, (2) and (3) can be used as Web IDEs. 2.3 Chrome Extensions A Chrome extension is an application that is used by incorporating it into the Chrome web browser. Chrome extensions include many development tools. Web Maker is a Chrome extension that can be used as a development tool. It is considered to be a standalone version of CodePen and works offline. However, it is not available on Chrome for smartphones. 2.4 Single-Page Applications An SPA is an application composed of a single web page (i.e., an HTML file). Strictly speaking, the back end exists on the server side, so the application on the client side consists of a single page. The views are dynamically generated on the client side, allowing the application to exhibit excellent response time. An SPA communicates asynchronously with the back end. Asynchronous communication can hide latency, and thereby shorten the response time. Increased responsiveness improves the user experience and makes an SPA easy to use. An SPA is often developed using SPA frameworks, such as React and Vue.js. Currently, JDE (described below) does not use these SPA frameworks. Rewriting JDE using an SPA framework is a topic for the future. 2.5 Flip Teaching Flip teaching [10, 11] is a type of blended learning that mixes distance learning and face-to-face learning. Basically, self-study is performed using prepared video materials, and later reviewed by face-to-face classes. Khan Academy offers a variety of materials that can be used in flipped classrooms. There are many cases of reversing programming [12].

3 JavaScript Development Environment We developed JDE in a previous study [15]. This section describes its design and implementation.

58

M. Uehara

3.1 Features JDE has the following features. (F1) JDE can execute code in a web browser. (F2) The JDE screen is suitable for smartphones. (F3) The operation of JDE is suitable for smartphones. These characteristics endow JDE with several advantages. JDE is usable offline. JDE does not overload the server, and can therefore be used by many users at the same time. JDE is usable on smartphones, so students can study while commuting between school and home. According to the Japan Student Services Organization, the average commuting time for Japanese college students is 1 h [13]. According to another survey, half of the students use the railway. Therefore, students can use their smartphones to study while commuting to school. 3.2 Screen Figure 1 shows a part of the screen running JDE on the iPhone X. The JDE screen consists of four parts: (1) menu, (2) code field, (3) output field, and (4) console field. Various buttons are lined up in the menu. The user inputs the code in the code field and edits it. Virtual keyboards on smartphones are not suitable for entering code, so JDE provides some buttons that help users to enter code. A virtual keyboard does not allow the user to enter a tab for indentation, so a Tab button is used for this purpose. The Enter button will enter an indentation after a line break. The +return button adds a return on the last line. The For button creates a for statement. For example, nested loops can be created from a list of variable names. The If button creates an if statement. The Syntax button assists with entering reserved words. The Op button assists with entering operators. The Brackets button creates brackets of various types, and positions the cursor between them. The Snippets button inserts frequently used code. Values generated by code execution can be added to the console and displayed. The final result of the code is displayed in the output field as the return value of the function. If no return value is specified, then the value displayed is undefined. 3.3 Library One of the problems with learning programming on smartphones is code input. An example library was developed to eliminate the need for code input (No Code Input). Figure 2 shows the screen of the example library. The example library is a regular web application, not an SPA. Students can use sections of code registered by the teacher. Currently, more than 200 sample sections of code are registered. The copy is performed as follows. First, the sample code is saved in the local storage of the browser with its tag name, which is an identifier that corresponds to the file name. JDE loads the sample code from local storage by using the tag name.

Experiences with a Single-Page Application for Learning Programming

59

Fig. 1. JDE screen.

3.4 Summary JDE is an SPA JavaScript development environment. JDE can be run only in a web browser and can be used offline. Students need to access the server to browse the example library, but once saved, they can then browse the library offline. JDE is designed to operate on a smartphone. Therefore, using JDE, teachers can usually teach in regular classrooms. The number of Wi-Fi connections is usually limited in these classrooms. However, because JDE runs on a smartphone, students can connect to the server with IP telephony (4G). Once they download the SPA, it can be used offline.

4 A Classroom Using JDE JDE has been used in the actual course “Introduction to Programming” since 2017. The course is offered in the first (Spring) semester. (In Japan, the school year starts in April.)

60

M. Uehara

Fig. 2. JDE library.

It has a total of 15 lessons with two credits. In these lessons, both lectures and exercises are offered. The purpose of the course is for students with no programming experience to learn the basics of programming. Although it uses JavaScript as its programming language, it emphasizes the learning of language-independent principles. The contents include basic concepts, such as data, variables, I/O, conditionals, loops, arrays, functions, and combinations thereof. The contents of each lesson are as follows. 1. 2. 3. 4. 5. 6. 7. 8. 9. 10.

Introduction Basic Concepts Getting started with JavaScript/Output Expressions/Variables/Input Conditional Statements and Logic Loops Nested Loops Computational Complexity Arrays Objects

Experiences with a Single-Page Application for Learning Programming

11. 12. 13. 14. 15.

61

Functions Advanced I/O by HTML Form Quizzes Checking the Answers Summary

The number of students taking the course in 2017, 2018, 2019, and 2020 was 78, 82, 72, and 78, respectively. Although it is not a compulsory subject, many students choose it. The same course is offered in three courses in parallel. Students are automatically assigned to a course by student ID number. Therefore, the number of students enrolled in each course is approximately one-third of the entire capacity of 260. The features of this course are as follows. (P1) JavaScript is used as the programming language. (P2) JDE is used as the development environment. (P3) Students can also practice on smartphones. (P4) Classes can be held in regular classrooms, instead of PC rooms. (P5) New concepts are combined with known concepts. (P6) Many examples are used. JDE satisfies the features of P1–P4. In the course, we use many examples for P5 and P6. For example, if students learn loops after learning conditional statements, they also learn applications in which a loop is combined with a conditional statement. After more classes, the number of combinations increases. Many sample sections of code are used to learn the combinations. The sample code can be found in the text distributed by the learning management system (LMS) called manaba [14]. This LMS prohibits the use of script tags and style tags, for security reasons. Therefore, it is necessary to manually copy and paste code from that text into JDE to run it. At that time, extra blank lines may be copied. Therefore, in 2019, we added a function to delete blank lines. Copying and pasting with a smartphone is not always easy. It is difficult to specify a sample code section accurately when it extends over 100 lines. Therefore, in 2019, we developed the example library, which avoids the need for students to enter the code. Students can also modify an example and try it out. Features added in 2019 are detailed in Ref. [15]. For example, the Open button has been added to create a window and display HTML. It is difficult to use the Run button, which executes JavaScript, to execute the I/O code of a Form.

5 Experiences of a Classroom Using JDE In 2020, many universities offered classes online because of the coronavirus pandemic. We also held classes online. In this section, we describe our experiences of an online class using JDE. Online classes can be classified into two types: live-streaming classes and ondemand-type classes. The former streams lessons using a video conferencing system,

62

M. Uehara

such as Zoom, Google Meet, Microsoft Teams, or Cisco Webex. This type of class is suitable for conducting traditional face-to-face lessons online. The latter distributes lesson videos using video on demand (VoD), such as YouTube or Microsoft Stream. This type is suitable for flipped classes: in the flipped classroom, video materials are originally delivered on demand. The subject “Introduction to Programming”, described in Sect. 4, uses on-demand classes. With this type of class, the content distribution method is important. The materials are shared with students on Google Drive, to limit the audience. The file size of a 90-min MP4 video (Full HD) is around 1 GB. Assuming that approximately 80 people download it at the same time, 80 GB of traffic will be generated at the beginning of the class, which is difficult to serve with an on-premises file server. Ideally, a VoD platform should be used. However, any platform that can handle a large amount of traffic can be used instead. For example, cloud-based online storage can be used as an alternative platform. Therefore, Google Drive can be used for content distribution. Our university and many others use G Suite for Education. JDE traffic will be described next. According to our previous research [15], JDE can be downloaded by each user in around 300 ms. The load on the server is proportional to the number of users, which we measured from the server log. Table 1 shows the results of classifying URLs from the logs for a period of 2 weeks (July 8–July 22, 2020). Here, the URL indicates an address pattern. The URL “/” matches addresses of all pages on this site. The JDE URL is “/jde”, which includes the addresses of the JavaScript library and the example library. The corresponding URLs are “/jde/js” and “/jde/lib”. Therefore, the number of accesses to “/jde/”, excluding the library, is only 392, at most. Students accessed it only five times on average. Because the classes were held twice every 2 weeks, the students only accessed it 2.5 times per class. By implementing JDE as an SPA, we were able to reduce the number of downloads significantly. However, the problem is that JavaScript libraries were downloaded more often than SPAs. This JavaScript library displays line numbers in the code field, but it is not absolutely necessary. Traffic can be reduced by removing unnecessary libraries. Table 1. URLs accessed and the number of accesses to each. URL

Accesses

/

26843

/jde

3731

/jde/js

1820

/jde/lib

1519

Figure 3 shows the dates of accesses. Classes were held every Friday, including July 10 and July 17. Although access was distributed, peaks occurred on class days. The number of accesses was almost equal to the number of students taking the course. Therefore, it follows that almost all the students had access. Because the system time

Experiences with a Single-Page Application for Learning Programming

63

zone is Coordinated Universal Time (UTC), accesses on class days were divided between July 16 and 17. In general, the number of accesses on weekends was small.

22-Jul-20

21-Jul-20

20-Jul-20

19-Jul-20

18-Jul-20

17-Jul-20

16-Jul-20

15-Jul-20

14-Jul-20

13-Jul-20

12-Jul-20

11-Jul-20

10-Jul-20

09-Jul-20

08-Jul-20

90 80 70 60 50 40 30 20 10 0

Fig. 3. Access dates.

Figure 4 shows the distribution of access times. Because the time zone is UTC, the time difference from Japanese Standard Time (JST) is +9. Class hours 16:30–18:00 (JST) are 7:30–9:00 (UTC). The first peak occurred during regular school hours. The next peak was at 14:00 (UTC) or 23:00 (JST). Students have a habit of learning before bed. The on-demand method is ideal for the lifestyle of such students.

45 40 35 30 25 20 15 10 5 0 00:0002:0004:0006:0008:0010:0012:0014:0016:0018:0020:0022:00 Fig. 4. Access times.

64

M. Uehara

Table 2 shows the types of client operating systems. JDE can be used on smartphones, but many students use PCs. However, the total number of iPhone and iPad accesses exceeded the number of Macintosh accesses. The Linux accesses were from IP addresses starting with “64” octets. According to Table 3, they were from Google and were not student accesses. Table 2. Client operating systems and the number of accesses. Client

Accesses

Windows

211

Macintosh

47

iPad

37

iPhone

27

Android

5

Other(Linux)

65

Table 3. Client IP addresses. Address pattern Carrier

Accesses

126.77.137.x

Softbank 37

64.233.172.x

Google

17+12+12

118.156.139.x

KDDI

15

110.132.0.x

Jupiter

12

Table 3 shows the IP addresses of the clients with the most accesses. The last octet is hidden. These were all class A addresses. The “Carrier” column indicates the administrator, as displayed by WHOIS. All except Google are mobile carriers. Google uses multiple addresses to collect files, such as favicons, and should be excluded. The results in Table 2 suggest that students were using JDE from their PCs. However, the results in Table 3 suggest that a significant number of students used mobile carriers, but not all of them did so. JDE can be used on smartphones, but the communication charges for smartphones are high, and students always act to reduce them. When students watch a 1-GB lesson video on their smartphone, their communications charges are very high, and therefore many students watch the video on their PC instead. The PC is connected to a flat-rate home Internet connection and includes a large screen, with video and JDE displayed side by side at the same time. The above suggests that there are few advantages to using a smartphone for home learning. Rather, it is more advantageous to use it in a regular classroom.

Experiences with a Single-Page Application for Learning Programming

65

6 Conclusions This paper described experiences of distance learning using JDE. Students downloaded JDE only 2.5 times per class on average. Therefore, the load on the server was very small. Furthermore, student accesses were distributed outside of class days. It was often accessed late at night and during class hours. With JDE, students can study both in the classroom and at home. However, with home learning, smartphones have no advantage over PCs. Therefore, the advantage of smartphones is that they can be used in a regular classroom. The topics for future work are as follows. JDE will be combined with other services. For example, JDE will make mutual copies of code with LMSs such as manaba. This will allow students to test programming problems. We will also improve the user interface, using React or Vue.js. JDE can be used as a smartphone application by rebuilding it with Flutter or React native XX. Acknowledgments. We thank Edanz Group (https://en-author-services.edanzgroup.com/ac) for editing a draft of this manuscript.

References 1. Cabinet Office, Programming Education. http://www.mext.go.jp/a_menu/shotou/zyouhou/ detail/1375607.htm. Accessed 4 July 2019 2. Vee, A.: Coding Literacy: How Computer Programming Is Changing Writing. The MIT Press, Cambridge (2017) 3. Aho, T., et al.: Designing IDE as a service. Commun. Cloud Softw. 1(1) (2011) 4. van Deursen, A., Mesbah, A., Cornelissen, B., Zaidman, A., Pinzger, M., Guzzi, A.: Adinda: a knowledgeable, browser-based IDE. In: Proceedings of the 32nd ACM/IEEE International Conference on Software Engineering, (ICSE 2010), vol. 2, pp. 203–206. ACM, New York (2010). http://dx.doi.org/10.1145/1810295.1810330 5. Yanagisawa, H., Uehara, M., Mori, H.: Web-based collaborative development environment for designing processors. GESTS Int. Trans. Comput. Sci. Eng. 30(1), ISSN 17386438 (2006) 6. Yanagisawa, H., Uehara, M., Mori, H.: Interface implementation using ajax for web-based instruction set simulator. In: Proceedings of 2nd International Workshop on Telecommunication Networking, Applications and Systems (TeNAS2008), pp. 1511–1516 (2008) 7. Yanagisawa, H., Uehara, M., Mori, H.: Web-based collaborative development environment for an ISA simulator. In: Cooperative Internet Computing, pp. 125–138 (2008). https://doi. org/10.1142/9789812811103_0009 8. Yanagisawa, H.: Evaluation of a web-based programming environment. In: 2012 15th International Conference on Network-Based Information Systems, Melbourne, VIC 2012, pp. 633–638 (2012). https://doi.org/10.1109/nbis.2012.67 9. Yanagisawa, H., Kondo, K.: Implementation of web-based interactive interface as software execution environment. In: 2013 16th International Conference on Network-Based Information Systems, Gwangju, pp. 383–388 (2013). https://doi.org/10.1109/nbis.2013.61 10. Lage, M., Platt, G., Treglia, M.: Inverting the classroom: a gateway to creating an inclusive learning environment. J. Econ. Educ. 31, 30–43 (2000)

66

M. Uehara

11. Baker, J.W.: The “Classroom Flip”: using web course management tools to become the guide by the side, p. 15. Communication Faculty Publications (2000) 12. Çakıro˘glu, Ü., Öztürk, M.: Flipped classroom with problem based activities: exploring selfregulated learning in a programming language course. J. Educ. Technol. Soc. JSTOR, 20(1), 337–349 (2017). www.jstor.org/stable/jeductechsoci.20.1.337. Accessed 21 July 2020 13. Japan Student Services Organization, “Student Life Survey Report 2019. (in Japanese). https://www.jasso.go.jp/about/statistics/gakusei_chosa/__icsFiles/afieldfile/2020/03/16/dat a18_all.pdf. Accessed 07 July 2020 14. AsahiNet, manaba. http://manaba.jp/. Accessed 07 July 2020 15. Uehara, M.: JavaScript development environment for programming education using smartphones. In: 2019 Seventh International Symposium on Computing and Networking Workshops (CANDARW), Nagasaki, Japan, pp. 292–297 (2019). https://doi.org/10.1109/candarw. 2019.00058

Approach of a Word2Vec Based Tourist Spot Collection Method Considering COVID-19 Yuki Nagai1(B) , Nobuki Saito1 , Aoto Hirata2 , Tetsuya Oda1 , Masaharu Hirota3 , and Kengo Katayama1 1

3

Department of Information and Computer Engineering, Okayama University of Science (OUS), 1-1 Ridaicho, Kita-ku, Okayama 700-0005, Japan {t18j057ny,t17j033sn}@ous.jp, {oda,katayama}@ice.ous.ac.jp 2 Engineering Project Course, Okayama University of Science (OUS), 1-1 Ridaicho, Kita-ku, Okayama 700-0005, Japan [email protected] Department of Information Science, Okayama University of Science (OUS), 1-1 Ridaicho, Kita-ku, Okayama 700-0005, Japan [email protected] Abstract. The Novel Coronavirus disease 2019 (COVID-19) is raging around the world and is seriously affecting daily lives and economic activities for people. For example, the work and class are now taking place online, and major lifestyle changes are also taking place. It is possible that COVID-19 will continue to change lifestyle habits in the future. To cope with this change, COVID-19 needs to be taken into account in daily life and in economic activities, and measures need to be taken based on a variety of information. A variety of approaches are currently being tried and tested around the world to assist in combating COVID19. This paper focuses on measures taken in the tourism industry and aims to propose a tourist spot collection method that takes into account COVID-19.

1

Introduction

In this year, the Novel Coronavirus disease (COVID-19) is spreading around the world [1–3]. This has had a serious impact on daily life and economic activities, requiring people to refrain from going out of the house, and has led to major lifestyle changes. The refrain from going out has also had an economic impact on restaurants, hotels and tourist attractions. On the other hand, since the economy cannot be stopped for long periods of time, daily life and economic activities are being resumed in areas where the number of people infected is declining, taking into account COVID-19. The academic associations also support COVID-19 infection prevention. For example, the Association for the Advancement of Artificial Intelligence (AAAI) provides a dataset on COVID-19 that summarizes the background and symptoms c The Editor(s) (if applicable) and The Author(s), under exclusive license  to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 67–75, 2021. https://doi.org/10.1007/978-3-030-61108-8_7

68

Y. Nagai et al.

Fig. 1. Proposed system.

of infection in infected individuals to help in the COVID-19 infection prevention. In Japan, the “Go To Campaign” has been running since July 22, 2020 in all regions except Tokyo, where the number of people infected with the disease is severe. This is a project to attract people to restaurants and tourist sites that had been economically affected by the refraining of outdoor activities, with some limited government subsidies for food and travel expenses. However, the risk of infection COVID-19 has not been completely eliminated. Therefore, tourism must be conducted while taking into account the impact of COVID-19. In this paper, we propose a Word2Vec [4] based tourist spot collection method considering COVID-19 infection prevention.

2

Proposed System

In this section, the processing of the proposed method is described. The process in this paper is divided into four major parts: collection of text data and training of Word2Vec, extraction of similarity from the distributed representation of training words, and extraction of tourist attractions from the word vectors of similarity (Table 1).

Word2Vec Based Tourist Spot Collection Method Considering COVID-19

69

Table 1. Experimental parameters. Parameteras

Value

Depth of search 3 10 Waiting time [s] 1 Minimum length of string Maximum length of strings 10000

2.1

Collection of Textual Data

In this section, we describes the data collection system and the data to be collected. The structure of the COVID-19 related textual data collection system is shown in Fig. 1. The crawler in proposed system load the HyperText Markup Language (HTML) in the website by crawling the website and scraping it to collect the Uniform Resource Locator (URL), strings and the date and time the strings were written. The extracted strings are converted into words using MeCab of a morphological analysis system. Then, the collected data are then classified by the indexer system into URL, strings and words. Also, the string and word data are accompanied by the date and time they were posted. The indexer in proposed system assign identification numbers to collected URL, words, and strings, and recognizes them to determine where the data is stored and to prevent data duplication. The structure for storing data is based on the Resource Description Framework (RDF) in Linked Open Data (LOD) [6–10], and each URL, word, and string in the same site has the same data structure. This allows for efficient retrieval of interlinked data. In proposed system, the electronic bulletin board “Open2ch” [12] is the target of the textual data collection. The period of time covered was 6 month period from February to July, when COVID-19 begins to spread, and all COVID-19 related posts posted during this period were considered. The following is an example of RDF in LOD. @prefix bp: . @prefix rdfs: . @prefix xsd: . @prefix geo: . rdfs:label “XXXXXX”@ja; bp:title “COVID-19 Information and Resources”; bp:url “http://XXXXX.ac.jp”; bp:words “Coronavirus Disease (COVID-19)”; bp:time “1 February, 2020 at 00:00”;

70

Y. Nagai et al.

Fig. 2. CBOW model of Word2Vec.

The indexer in proposed system sends the URL to be searched to the crawler system in the order in which they are found. The proposed system collects data by repeating these processes. The data collected here will be sent to Word2Vec as the training data. Table 2. Training parameters. Parameteras

Value

Number of words Number of training Window size Batch size Node of middle class Initial weight

1257401 139270 5 100 100 Random

Word2Vec Based Tourist Spot Collection Method Considering COVID-19

71

Fig. 3. Recommendation method.

2.2

Learning Textual Data Based on Word2Vec

In this section, we describe learning textual data based on Word2Vec. The proposed system is shown in Fig. 2. We use Word2Vec to obtain a distributed representation from the collected text data, and the training model is the CBOW model [11]. In addition to data collected from the Open2ch data is used for training. The textual data used in open2ch, writings posted between February and August 9, which is used for the training. The data are combined and trained in Word2Vec as one training data. In order to improve the accuracy of word predictions, if the submitted text contains symbols, the symbols and emojis are removed in advance. The training parameters for Word2Vec are shown in Table 2. 2.3

Recommendation Method for Tourist Spot Considering COVID-19 Using Word2Vec

In this section, we will discuss how to extract tourist spot. The process of recommending a tourist destination is shown in Fig. 3. Determine the cosine similarity on the variance representation from the input words. We prepared a database of tourist attractions in advance, and by referring to a database of tourist attractions from a list of words with high cosine similarity, we finally extracted only the tourist attractions. We also obtain the cosine similarity of COVID-19 related words from the variance representation by entering the COVID-19 related words. The extracted word vectors of tourist spots minus the word vectors of related words in COVID-19 are recommended as tourist attractions with low relevance to COVID-19.

72

Y. Nagai et al. 5 Error per Number of learning

Cross entropy error

4

3

2

1

0 0

20000

40000

60000 80000 Number of learning

100000

120000

Fig. 4. Result of cross entropy error.

3

Case Study

In this section, we present a case study of the proposed system. 3.1

Word2Vec Learning Results

We describes the results of the study of Word2Vec. Firstly, crawling of COVID19 related pages of Open2ch was performed by the proposed system. As a result, we scraped 615 pages and collected 206051 lines of text. Based on this string, we generated training data for 1257401 words and used this data to train Word2Vec. Next, the training process of Word2Vec is shown in Fig. 4 and the post-training distributed representation is shown in Fig. 5. Figure 4 shows the cross-entropy error for each training session, and it can be seen that the error decreases as the training progresses and Word2Vec is learning. The study was repeated 139270 times in a sequence of forward to reverse propagation and took 348 minutes to complete. The variance representation in Fig. 5 is represented using the weights of the input layer of Word2Vec after training, with each dot representing a word. From the coordinates of the words in this distributed representation, we calculate the cosine similarity between the words.

Word2Vec Based Tourist Spot Collection Method Considering COVID-19

73

Fig. 5. Result of distributed representation.

Fig. 6. Tourist spots and degree of similarity.

3.2

Experimental Results

We describe the experimental results of the proposed system. In this paper, “Travel” was used to extract tourist spots and “Corona Virus” was used as

74

Y. Nagai et al. Table 3. Tourist spots and COVID-19 related words. Travel

Corona

Gokayama Yakushima Tushima Sado Higashihonngannzi Kansai Kinpu-san Minamiizu Yunoyamaonnsen Tochigi Yokohama Shirakawago Osaka-Jo Tohukuzi Kyushu San’in Kusiro Makuhari Messe Shiretoko Kawai Kisarazu Wakasa Ichihino Hiroshima Takehu Niseko Kabuki-za Ichihuzi

Virus New model Pollen Pneumonia Measures Bacteria Each country China Prefectures Infection Confusion Mutation Gene Stock price Company Trend GoToTravel Treatment Prevalence Kanto Hospital Proportion Vaccine Splash Inspection Japan High fever Hokkaido

a related term of COVID-19. A list of tourist spots extracted from “Travel” and examples of words extracted from “Corona Virus” are shown in Table 3. From “Travel” it was possible to extract 28 tourist spots, which included the names of tourist spots and the place names of tourist spots. Examples of words extracted from “Corona Virus” are randomly selected from the extracted words. The result of subtracting the words extracted from “Corona” from the list of tourist spots is shown in Fig. 6. The graph shows the degree of similarity between the tourist sites and “Corona Virus”. In this experiment, we were able to extract 16 places. The 16 tourist spots extracted, 3 were tourist spots and 16 were place

Word2Vec Based Tourist Spot Collection Method Considering COVID-19

75

names of tourist spots. The maximum similarity to “Corona Virus” was 73.15 for “infection” and the average similarity was 22.98, so these 16 tourist sites were extracted as having a low association with “Corona Virus”.

4

Conclusion

In this paper, we proposed a of Word2Vec based tourist spot collection method considering COVID-19. The proposed system allowed us to extract 16 tourist sites in this paper. Future goals are to increase the amount of data we use for training so that we can increase the number of places we can extract, and to learn the coordinates of places so that we can make recommendations that take into account the location of tourist spot. Acknowledgement. This work was supported by Grant for Promotion of Okayama University of Science (OUS) Research Project (OUS-RP-20-3).

References 1. Mehta, P., et al.: COVID-19: consider cytokine storm syndromes and immunosuppression. Lancet 395(10229), 1033–1034 (2020) 2. Bai, Y., et al.: Presumed asymptomatic carrier transmission of COVID- 19. J. Am. Med. Assoc. (JAMA) 323(14), 1406–1407 (2020) 3. Fauci, A., et al.: COVID-19 - navigating the uncharted. New Engl. J. Med. 382(13), 1268–1269 (2020) 4. Mikolov, T., et. al.: Distributed representations of words and phrases and their compositionality. In: Proceedings of the 26th International Conference on Neural Information Processing Systems (NIPS-2013), vol. 2, pp. 3111–3119 (2013) 5. Kudo, T.: MeCab: yet another part-of-speech and morphological analyzer. https:// taku910.github.io/mecab/. Accessed 20 Aug 2020 6. Fernandez, J., et. al.: Compact representation of large RDF data sets for publishing and exchange. In: Proceedings of The 9th International Semantic Web Conference (ISWC-2010), pp. 193–208 (2010) 7. Ermilov, I., et. al.: Linked open data statistics: collection and exploitation. In: Proceedings of The 4th International Conference on Knowledge Engineering and the Semantic Web (KESW-2013), pp. 242–249 (2013) 8. Ristoski, P., et al.: Mining the web of linked data with rapidminer. J. Web Semant. 35(3), 142–151 (2015) 9. Kamdar, M., et al.: Enabling web-scale data integration in biomedicine through linked open data. npj Digit. Med. 2(90), 1–14 (2019) 10. Maki, T., et al.: Resource propagation algorithm considering predicates to complement knowledge bases in linked data. Int. J. Space-Based Situat. Comput. 8(2), 115–121 (2018) 11. Goldberg, Y., Levy, O.: Word2Vec explained: deriving Mikolov et. al.’s negativesampling word-embedding method. arXiv:1402.3722, pp. 1–5 (2014) 12. Open2ch. https://open2ch.net/. Accessed 20 Aug 2020

Detecting Distracted Driving from Images by Processing Relative Locations of Objects of Interest Inside Vehicles Arup Kanti Dey(B) , Bharti Goel, and Sriram Chellappan University of South Florida, Tampa, FL, USA {arupkantidey,bharti,sriramc}@usf.edu

Abstract. Distracted driving on roads is a problem that is common across the world now. With increasing use of smarter and connected devices, coupled with their miniature form factors, humans are now increasingly using these devices under mobility. When operating a vehicle, using smart devices can pose serious threats to road safety. Another contributing factor to distracted driving today stems from the urge to eat, drink and fall asleep while driving. In this paper, we use a popular and publicly available image dataset captured from embedded cameras inside cars that indicate instances of distracted driving or not. Different from existing works that look at the entire image to classify distracted driving, we first localize objects within the image that impact distracted driving. There are three broad categories we localize, namely, external entities (smartphones and bottles); entities within the car (steering wheel) and human-centered entities (left and right hand). Our approach to localize objects is based on Regional-Convolutional Neural Networks (R-CNNs). Once we localize these objects, we then design simpler machine learning techniques to process the relative locations of these objects within the image to detect instances of distracted driving. Our resulting performance evaluations demonstrate the validity of our approach. To the best of our knowledge, our work in this paper is unique, and we believe, provides more contextual relevance towards detecting instances of distracting driving, and could possibly yield newer approaches to educate drivers on safe driving.

1

Introduction

Distracted driving is any activity that deviates a driver’s attention from the road while driving. Many studies attest to the seriousness of the problem today [1]. There is increasing consensus that technologies are needed to notify distracted drivers including where, when and how they have been distracted, so that drivers can understand and correct dangerous behaviors. In this paper, we look to leverage assistance of computer vision techniques to detect instances of distracted driving using images captured from embedded cameras inside of cars. c The Editor(s) (if applicable) and The Author(s), under exclusive license  to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 76–86, 2021. https://doi.org/10.1007/978-3-030-61108-8_8

Detecting Distracted Driving from Images by Processing Relative Locations

77

The dataset we used for our problem is American University of Cairo (AUC) distracted driving dataset [2]. It has a total of 10, 555 training images separated into nine classes that indicate distracted driving, and one safe driving class. The images were captured using an embedded camera inside the cars. Real drivers simulated various aspects of distracted and safe driving while the car was static (for safety reasons). This dataset is popular and used extensively in peer research [2–5]. For our study we choose images from only five classes that contribute most to distracted driving. These classes include talking on phone with left hand, talking on phone with right hand, texting using left hand, texting using right hand, and drinking while driving. The sixth class is the safe driving class without distractions. Using this dataset, we design a multi-tiered approach for detecting instances of distracted driving. Specifically, we first use the notion of Region-based Convolutional Neural Network (Faster R-CNN) algorithm, to process images and localize objects that impact distracted driving. These objects include smartphones and bottles (external devices), steering wheel (internal to the car) and left & right hand (human-centric). Once we detect and localize these objects, we design simpler machine learning techniques to learn from their locations in the image relative to each other to classify an image as indicative of distracted driving or not. To the best of our knowledge, this is the first paper we are aware of that localizes objects of interest in an image and using those positions relative to each other to detect distracted driving. We believe that our approach will be able to provide superior contextual feedback to drivers, and such educational programs are part of our future work. The paper is organized as follows: Sect. 2 discusses related work of distracted driving detection and object detection, Sect. 3 discusses experimental design and, concurrently describe the dataset used and related architecture, Sect. 4 discusses evaluation results. Finally, in Sect. 5 we conclude our paper with closing remark and potential scope for future work.

2

Related Work

We now present important related work in the space of detecting distracted driving using technological assistance. Renato et al. [6] have worked on automatically processing images from internal video surveillance systems released by State Farm [7]. They mainly focused on identifying driver’s distraction due to mobile phone activity like talking and texting. They proposed a Convolutional Neural Network (CNN) to classify when a driver is using mobile phone. With the use of 9841 images, they attained 99% detection accuracy for distracted driving. Sarfaraz et al. [8] proposed a machine learning model using CNNs to identify driver’s distraction on the same dataset as mentioned above. However, they did not limit distractions only to mobile phone activities, but considered a total of 9 distraction activities and one safe driving activity. They have used VGG-16[9] and VGG-19 [9] architecture model to classify distraction and have achieved around 99% accuracy for most of the distraction class.

78

A. K. Dey et al.

Abouelnaga et al. [2] developed a novel genetically weighted ensemble based on CNNs to detect distracted driving from a dataset of images captured from a smartphone attached to top of passenger side seat. They trained Alexnet [10] and Inception V3 model [11] and calculated weighted sum of all the networks to the data and achieved classification accuracy of 94.25% for detecting distracted driving. This dataset consists of 10, 555 images, and is open sourced, and is used in this paper. However, they classify the entire image, instead of contextualizing the classification like we do in this paper. There are also other works like [12–16] where inertial sensors have been used to detect distracted driving, but since these are not based on images, we do not elaborate them in this paper. To the best of our knowledge, our work in this paper is the first to localize important objects of interest within an image, and then use the relative locations of those images to detect instances of distracted driving. This approach we believe enables more contextualization of distracted driving, and may provide newer resources for driver education.

3 3.1

Algorithm Design for Detecting Distracted Driving Data Set Description

In this experiment, we have used AUC Distracted Driving Dataset presented above that is open sourced [2]. The image data was collected using the camera of a smartphone, which in this case was an ASUS ZenPhone (Model Z00UD) emplaced on top of passenger side seat. Initially, data was collected in a video format then cut into individual image of size 1920 × 1080 each. Images were labeled in 10 different classes, among them one was safe driving class and rest were classified as distracted driving labeled as texting right hand, texting left hand, talking right hand, talking left hand, adjusting radio, talk passenger, doing makeup, drinking, reaching behind. For this paper, we chose five classes among the above for distracted driving. These classes included calling left-hand, calling right-hand, texting left-hand, texting right-hand and drinking while driving. We chose these five classes for distracted driving because, phone usage and eating/ drinking are known to be the most significant sources for accidents [17], and so we focused on these five classes only. The sixth class in our model was obviously the safe driving class. 3.2

Faster R-CNN Architecture

Our goal in this paper is to enable contextual detection of distracted driving. To do so, rather than classify an image as merely distracted or not, we wanted to design an approach, where objects in an image that most likely impact distracted driving are localized first, and the relative locations of those objects are utilized for classification. For instance, consider the Fig. 1 below that is indicative of distracted driving from the AUC dataset. Instead of a traditional approach that

Detecting Distracted Driving from Images by Processing Relative Locations

79

indicates this image is indicative of distracted driving (on the left), we want to detect objects of interest, namely hand and phone, and measure the relative locations of these objects, and since they are close to each other, we want to tag this image as indicative of distracted driving (on the right). We believe that this approach provides superior contextual feedback to drivers for subsequent training programs. This is the core novelty of our work in this paper.

Fig. 1. Traditional classification of distracted driving vs. classification based on object detection and localization

To enable our approach, the first step is object detection and localization. Faster R-CNN (Faster Regional-Convolutional Neural Networks) technique [18] is a popular and widely used object detection architecture, that can simultaneously localize objects of interest within the image. The Faster R-CNN architecture is composed of three parts and they are 1) Convolution layer 2) Region Proposal Network (RPN) and 3) Class and Bounding box prediction. The basic idea is the following, and incurs a series of steps. The first step is to derive feature maps from the entire training image dataset using a state of the art CNN. Then, we emplace numerous bounding boxes of arbitrary sizes in each training image and manually label the foreground and background class. The foreground class comprises of our objects of interest (hands, steering wheel, phone and bottle), and the background class contains other objects. Then we use the notion of Regional Proposal Networks to train a simpler CNN to learn feature maps that separate the foreground from the background class. The, the next step is to employ the notion of Region Of Interest (ROI) pooling to convert variable sized feature maps to fixed size feature maps. The last step in training is a simple Regressor CNN to tighten the anchor boxes, which is an important step, because the tighter the anchor boxes are, the better will be accuracy of processing relative locations of these anchor boxes. All of these steps are explained below in the process of localizing the objects of interest (see Fig. 2).

80

A. K. Dey et al.

Fig. 2. Illustration of our overall methodology

3.2.1 Detailed Explanation of Steps to Localize Objects in the Image In order to train and test the neural network architecture for localization, we selected 302 images equally distributed among all the six classes of interest to this study. These included 40 images from each class to make up a total of 240 images for training, and 62 more images among the classes were used for validation purposes. First, for feature extraction from these training images, we used the popular Inception V2 network [19]. It has a total of 164 layers, and is well studied in the literature. In order to learn feature maps corresponding to the foreground and background class, we used the labelImg tool [20] to first label the corresponding components in the training images. Regarding other critical parameters, we employed Stochastic gradient descent with learning rate of 0.0019 for first 90000 steps in the training with a batch size of 1. We set 0.9 for momentum optimizer with random horizontal flip as only data augmentation. The entire model was trained in a graphical processing unit (GPU) cluster which, has 4 nodes of GeRorce GTX TITAN X each with 12GB of memory. Cross entropy loss was used for classification and smooth L1 loss was used for regression. After training the model for 86531 steps we test the accuracy with 36 additional unseen images, results of which are presented later when we discuss results. 3.3

Classification of Images as Indicative of Distracted Driving or Not, Using Traditional Machine Learning Models

Post localization of objects of interest, the last and final task to classify whether or not the relative positions of objects in the image indicate distracted driving. To do so, we utilize a traditional machine learning algorithm - Random Forests [21] which gave us best accuracy. For this algorithm, we identified six features presented in Table 1. Note that for distance calculation, we determine the center of each bounding box in the localized object of interest and calculate distance between the centers.

Detecting Distracted Driving from Images by Processing Relative Locations

81

Table 1. List of features for machine learning model

4 4.1

Features

Description

Left Hand-Mobile Phone Left hand- Bottle Left Hand-Steering-Wheel Right Hand- Mobile Phone Right Hand-Bottle Right Hand-Steering-Wheel

Distance Distance Distance Distance Distance Distance

between between between between between between

Left hand and Mobile Phone Left hand and Bottle Left hand and Steering wheel Right hand and Mobile Phone Right hand and Bottle Right hand and Steering wheel

Discussion of Results Results of Object Localization

We separated out 36 unseen images from the dataset from all six classes for testing. Results are presented now. First, we present the metrics used to evaluate our localization component. For the localization component, the metric we use to evaluate is the mean Average Precision (mAP). Before, we define mAP, we define a few other metrics first. These are Precision, Recall and Intersection over Union (IOU). These three are defined below, where TP, FP and FNs are True Positives, False Positives and False Negatives respectively: P recision = Recall =

TP TP + FP

TP TP + FN

Intersection Over U nion (IOU ) =

area of intersection area of union

(1) (2) (3)

Here, IOU measures the overlap between two anchor boxes. Recall that the ground truth has manually emplaced anchor boxes for the testing dataset, and the AI algorithm emplaces its own anchor boxes after classification. The IOU takes the ratio of the area of intersection between these two boxes over the area of the union. Higher IOU implies superior localization performance. An IOU of 0.6 to 0.7 is considered state of the art today. To define our final metric, the Mean Average Precision (mAP), we define another metric, Average precision (AP), which is the average of all the Precision values for a range of Recall (0 to 1 for our problem) at a certain preset IOU threshold for a particular class among the ones for our problem (i.e., smartphone, bottle, hands and steering wheel). This metric essentially balances both Precision and Recall for a particular value of IOU for one class. Finally, the Mean Average Precision (mAP) is the average of AP values among all our classes. Table 2 shows the mAP score for different IOU thresholds for our testing image dataset. For IOU of 0.5, the performance is indeed good with an mAP of 0.84. As the IOU is increased (meaning more

82

A. K. Dey et al.

stringent evaluation), the mAP decreases. To summarize here, we are confident that our technique can indeed detect and localize sources that impact distraction in car to a high degree of accuracy. Table 2. Mean average precision comparison Model

mAP IOU = 0.50 mAP IOU = 0.75

Faster RCNN with InceptionV2 0.8401

0.4089

Figure 3 show instances of the output of our above technique for object detection and localization. Results shown here are representative of other images in our testing dataset. We believe that these methods of detecting and localizing objects will pique interest of drivers to learn from mistakes and improve their safer driving skills, and such educational programs are part of our future work in this space.

Fig. 3. Results of object detection and localization for distracted driving

4.2

Results of Image Classification as Indicative of Distracted Driving or Not

After detecting and localizing the objects of interest, we design traditional machine learning algorithms using six features presented in Table 1. We tested four different ML models and their overall accuracies are shown in Table 3.

Detecting Distracted Driving from Images by Processing Relative Locations

83

Table 3. Accuracy of different models for multi-class classification Model

Accuracy

K-nearest neighbors Decision Tree SVM Random forest

0.55 0.44 0.36 0.75

As we can see, the Random Forest classifier is providing best accuracy. From Fig. 4 we can see the confusion matrix of our model. From Fig. 5, we see the Precision, Recall and F1 score (that integrates both Precision and Recall) for the Random Forest model. Apart from multi-class classification, we also tried with binary class classification where we only used Distracted and Not Distracted class and tried classification with three ML models. The overall accuracy of those models is shown in Table 4. The confusion matrix and Precision, Recall and F1 score are presented in Fig. 6 and Fig. 7 respectably.

Fig. 4. Confusion matrix for multi-class classifier

84

A. K. Dey et al.

Fig. 5. Precision, Recall and F1 score for multi-class classifier

Table 4. Accuracy of different model for binary class classification Model

Accuracy

Decision tree 0.72 0.80 Random forest K-nearest neighbor 0.75

Fig. 6. Confusion matrix for binary-class classifier

Detecting Distracted Driving from Images by Processing Relative Locations

85

Fig. 7. Precision, Recall and F1 score for binary classifier

5

Conclusions

In this paper, we design a novel technique to detect instances of distracted driving from images captured inside of cars. Different from existing works in this space, we first localize objects of interest in the image that could impact distracted driving (smartphones, hands, bottles and steering wheel), and use their relative locations to design algorithms to detect distracted driving. We present our results using several metrics. Using ideas presented in this paper to improve education for drivers is part of our future work where we will create a much larger corpus for training dataset by incorporating significantly greater number of images.

References 1. Vegega, M., Jones, B., Monk, C.: Understanding the effects of distracted driving and developing strategies to reduce resulting deaths and injuries: a report to congress. Art. no. DOT HS 812 053, December 2013. Accessed 16 July 2020 2. Abouelnaga, Y., Eraqi, H.M., Moustafa, M.N.: Real-time distracted driver posture classification. arXiv:1706.09498, November 2018, Accessed 14 July 2020 3. Moslemi, N., Azmi, R., Soryani, M.: Driver distraction recognition using 3D convolutional neural networks. In: 2019 4th International Conference on Pattern Recognition and Image Analysis (IPRIA), Tehran, Iran, pp. 145–151 (2019). https://doi. org/10.1109/PRIA.2019.8786012. 4. Leekha, M., Goswami, M., Shah, R.R., Yin, Y., Zimmermann, R.: Are you paying attention? Detecting distracted driving in real-time. In: 2019 IEEE Fifth International Conference on Multimedia Big Data (BigMM), Singapore, Singapore, pp. 171–180 (2019). https://doi.org/10.1109/BigMM.2019.00-28. 5. Mase, J.M., Grazziela, P.F., Chapman, P., Torres, M.: A hybrid deep learning approach for driver distraction detection. ResearchGate. https://www. researchgate.net/publication/340917598 A Hybrid Deep Learning Approach for Driver Distraction Detection. Accessed 15 July 2020

86

A. K. Dey et al.

6. Torres, R., Ohashi, O., Carvalho, E., Pessin, G.: A deep learning approach to detect distracted drivers using a mobile phone. In: Lintas, A., Rovetta, S., Verschure, P.F.M.J., Villa, A.E.P. (eds.) Artificial Neural Networks and Machine Learning ICANN 2017, vol. 10614, pp. 72–79. Springer, Cham (2017) 7. State farm distracted driver detection. https://kaggle.com/c/state-farmdistracted-driver-detection. Accessed 14 July 2020 8. Masood, S., Rai, A., Aggarwal, A., Doja, M.N., Ahmad, M.: Detecting distraction of drivers using convolutional neural network. Pattern Recognit. Lett. S0167865517304695 (2018). https://doi.org/10.1016/j.patrec.2017.12.023. 9. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition (2014). arXiv preprint arXiv:1409.1556 10. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. In: Advances in Neural Information Processing Systems, pp. 1097–1105 (2012) 11. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the inception architecture for computer vision. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2818–2826 (2016) 12. Goel, B., Dey, A.K., Bharti, P., Ahmed, K.B., Chellappan, S.: Detecting distracted driving using a wrist-worn wearable. In: 2018 IEEE International Conference on Pervasive Computing and Communications Workshops (PerCom Workshops), pp. 233–238. IEEE (2018) 13. Ahmed, K.B., Goel, B., Bharti, P., Chellappan, S., Bouhorma, M.: Leveraging smartphone sensors to detect distracted driving activities. IEEE Trans. Intell. Transp. Syst. 20(9), 3303–3312 (2018) 14. Goel, B., Dey, A.K., Chellappan, S.: Detecting routes taken by users on public vehicles from their wearables. In: 2017 8th IEEE Annual Information Technology, Electronics and Mobile Communication Conference (IEMCON), pp. 451–457. IEEE (2017) 15. Fazeen, M., Gozick, B., Dantu, R., Bhukhiya, M., Gonz´ alez, M.C.: Safe driving using mobile phones. IEEE Trans. Intell. Transp. Syst. 13(3), 1462–1468 (2012) 16. Singh, P., Juneja, N., Kapoor, S.: Using mobile phone sensors to detect driving behavior. In: Proceedings of the 3rd ACM Symposium on Computing for Development, pp. 1–2 (2013) 17. 3 dangerous activities you might do while driving—Lytx. https://www.lytx.com/ en-us/news-events/press-release/2014/lytx-data-finds-three-dangerous-activitiesyou-may. Accessed 16 July 2020 18. Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: towards real-time object detection with region proposal networks. arXiv:1506.01497, January 2016. Accessed 14 July 2020 19. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the inception architecture for computer vision. arXiv:1512.00567, December 2015, Accessed 14 July 2015 20. Tzutalin. LabelImg. Git code (2015). https://github.com/tzutalin/labelImg 21. Breiman, L.: Random forests. Mach. Learn. 45(1), 5–32 (2001)

Cost and Performance Analysis of Cuckoo Search Based File Replication in MANET Takeru Kurokawa and Naohiro Hayashibara(B) Kyoto Sangyo University, Kyoto, Japan {i1986061,naohaya}@cc.kyoto-su.ac.jp

Abstract. Mobile ad-hoc network consists of a collection of mobile devices that are interconnected each other. It has attracted attention in various areas, such as information sharing and intelligent transportation. In particular, information sharing on MANET in an emergency is crucial for victims. In this situation, information on relief supplies and safety confirmation should be replicated to keep it on the network as long as possible. On the other hand, the cost of replicated information should be taken into account since the storage of each mobile device is limited. In this paper, we evaluate the performance of the replication protocol based on the cuckoo search that is a meta-heuristic algorithm inspired by the egg-laying habits of cuckoos to improve the availability of low demand data in MANET. We also discuss the tradeoff between file availability and the costs for storage and communication associated with the replication protocol.

1

Introduction

Mobile ad-hoc network (MANET) is widely used by many applications such as the information sharing system in an emergency [12], indoor location tracking [1] and intelligent transportation in a smart city. Now, we suppose the information sharing in the case of emergency (e.g., disaster). In this case, there is no network infrastructure available. So, MANET that consists of mobile phones is an alternative to a network infrastructure (e.g., wifi access points) for sharing information among victims. Victims in a shelter upload information on disaster relief supplies, medical support, livelihood support, and confirmation of individual safety. The number of replication of data (files) is often in proportion to the number of requests in common replication schemes, for example, Path replication [4] and Owner replication [10]. Thus, files that are frequently required by users create many replicas. For instance, information on disaster relief supplies and medical, livelihood support is required by many people and replicated actively. Contrary to this, files with low demand create a small number of replicas. This type of file could be disappeared easily in MANETs by node’s leaving or churn. On the other hand, those files could be important for specific individuals, for example, the safety information on family and relatives. c The Editor(s) (if applicable) and The Author(s), under exclusive license  to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 87–96, 2021. https://doi.org/10.1007/978-3-030-61108-8_9

88

T. Kurokawa and N. Hayashibara

In general, the number of replicas is correlated with the availability of the files. The availability of information increases as the number of its replicas increases. For improving the availability of low demand files, the total number of replicas must be increased. As a result, the cost of the storage of each node and the network becomes high. This is a serious problem in MANETs. This work aims to improve the availability of low demand files without imposing a heavy load on the devices and the network. Now, we assume a mobile ad-hoc network that consists of mobile devices and information is shared on it in the case of emergency (e.g., disaster). In this case, we need to take into account the storage cost in each device because it is strictly limited. We also make low demand (but still important) information available on the network as long as possible at the same time. For this purpose, we proposed a novel approach for data replication on MANETs based on Cuckoo search (CSPR) [8] in the previous work. Cuckoo search [13] is a meta-heuristic algorithm inspired by the egg-laying habits of cuckoos. It is known as an efficient solution for nonlinear global optimization problems, and it is proven that it guarantees the global convergence by using L´evy flight [5]. Roughly speaking, we use Cuckoo search for allocating the replicated files to nodes in MANET. Each node evaluates other nodes that are close to it based on the residual of their storage and battery, then sends copies of files, instead of eggs, to the best nodes to improve the file availability, especially for low demand files. On the other hand, we need to consider the tradeoff between file availability and the costs for replication since the storage and the bandwidth are limited in MANET. In this paper, we measure the availability of files, the storage cost, and the communication cost of CSRP to evaluate that it is suitable for MANETs. Moreover, we also compare several existing replication protocols with CSRP regarding the performance metrics stated above.

2

Related Work

First, we introduce some of the replication protocols in MANETs and P2P systems. Then we explain the Cuckoo search for solving optimization problems. 2.1

Replication in MANETs and P2P Systems

Many data replication protocols in MANETs and P2P systems have been developed so far. Owner replication [10] is a simple replication protocol which replicates data to the node that issues a search request for it. It means that one replicated data corresponds to the search query. Thus, the cost of replicating data is low. However, this protocol may require many messages for searching for data. Path replication [4] creates replicas of the data and allocates them to all nodes along the path from the requesting node to the providing node in a search query for it. This protocol reduces the search traffic (i.e., the number of messages) by

Cost and Performance Analysis of Cuckoo Search Based File Replication

89

a factor of three compared to Owner replication. On the other hand, it imposes a higher storage cost than Owner replication because of the more significant number of replicated data. Choi et al. proposed a novel replica allocation technique taking into account selfish nodes that cooperate partially or not at all with other nodes [3]. Those selfish nodes could reduce the overall data accessibility in the network. The proposed technique improves data accessibility, the cost of communication, and the average query delay by detecting selfish nodes. 2.2

Replication for Low Demand Information

Kageyama and Shibusawa proposed a replication protocol (KS protocol) for improving the availability of low demand files in super-node based P2P systems [7]. It replicates the high demand files using Owner replication. In addition, it computes a demand forecast for low demand files and decides the number of replicas to be created. It assumes that the access distribution to files obeys Poison distribution as follows. λt −λ e (1) t! λ is the average number of requests on certain data in the Eq. 1. So, the probability of requests to the data at time t is represented as the equation. Every node measures the number of requests to data that are stored in the node. The prediction of data requests is computed based on the exponential smoothing. Let M Rt and P Rt be the number of requests which is measured, and the predicted number of requests. So, P Rt+1 will be represented as follows. P (X = t) =

P Rt+1 = δ · M Rt + (1 − δ) · P Rt

(2)

Data is distinguished as low demand if P Rt+1 for the data becomes lower than the bottom five percent of the probability distribution P (t + 1). 2.3

Cuckoo Search

Cuckoo search proposed by Yang and Deb [13] is a metaheuristic algorithm inspired by the brood parasitic behavior of the common cuckoo species. They lay their eggs in the nest of other species of birds. A cuckoo egg closely mimics the eggs of the host ones. The host birds may incubate and hatch the cuckoo egg, or they may recognize the intruding egg and abandon the nest.

3

System Model

We assume that a MANET consists of a set of nodes N = {n1 , n2 , ..., n|N | }, which are mobile devices (e.g., smartphones, laptop computers, wearable devices). Each node has a unique identifier. Each pair of nodes can communicate with each other. It means that a node in the network has a multi-hop path to any other

90

T. Kurokawa and N. Hayashibara

node. We can obtain the remaining amount of battery RBit and storage RSit of a node Ni at time t. We also define the capacity of battery CBi and storage CSi of a node Ni . Moreover, Ati is defined as a duration (minutes) that Ni joins the network at time t. We conducted simulations by a cycle-based simulator. Thus, Ati is represented as the number of cycles. A MANET we assume is a dynamic distributed system. Nodes leave from the network by going out of the communication range of any other node, turning off the devices, and so on. In addition, nodes can join the network as new nodes. Once a node leaves from the network, it joins it again with another identifier without any data and logs in the previous execution. It means that we assume fail-stop failure. We also assume a membership service such as [2,9,11] to manage the status of each node. Both CSRP protocol and KS protocol determine target files that should be replicated based on the access frequency of each file. Basically, it is monitored while in execution. In our simulation, we assume that it obeys Pareto distribution. The topology of a MANET is assumed to be a random geometric graph.

4

Cuckoo Search Based Replication Protocol

This work aims to improve the availability of low demand files in MANETs. Kurokawa and Hayashibara proposed the Cuckoo Search based Replication Protocol (CSRP) [8]. It is a modified version of KS protocol [7] with a replica allocation mechanism using Cuckoo Search [13]. CSRP consists of two phases; 1) labeling phase, 2) replica placement phase. In the labeling phase, each node affixes a label of high demand or low demand to each file in its storage in CSRP. Then it executes the replica allocation protocol, which consists of two procedures, the replication procedure for high demand files and for low demand files in the replica placement phase. CSRP simply uses Owner replication for high demand files. Owner replication does not ensure high file availability, while it is a cost efficient way. However, high demand files are frequently replicated upon file requests. Thus, it is a suitable protocol for replicating this type of files in MANET. On the other hand, CSRP uses a replica allocation protocol based on Cuckoo search [13] for low demand files. Replica allocation is important for this type of files because they are not frequently replicated. Cuckoo search is for a maximization problem of the value of the object function. We use this mechanism to select the best nodes (mobile devices in MANET) that hold replicated low demand files for improving the availability of those files. More precisely, we use Cuckoo search to find the destination ni of replication, which has a maximum value of E(i) among candidates.    RSit RBit +β× × Ati (3) E(i) = ln α× CBi CSi We suppose that each node ni ∈ N has two parameters, the capacity of battery CBi and storage CSi . Moreover, it has three variables, the remaining

Cost and Performance Analysis of Cuckoo Search Based File Replication

91

amount of battery RBit and storage RSit , and the duration Ati at time t. The proposed protocol evaluates the nodes in the network based on the following equation and computes the evaluation value E(i) of each node Ni . α and β are weights for the evaluation where α, β = [0, 1]. To simplify the capacity of battery and storage, CBi and CSi are defined as 100.0. Thus, RBit and RSit are variables from 0.0 to 100.0.

5

Performance Evaluation

The purpose of the evaluation is to clarify the following things. • How does the proposed protocol improve file availability in a MANET depending on the protection period P . • Trade-off between file availability and the cost of storage and the network. • Performance comparison of the proposed protocol and other protocols. We discuss the things listed above according to the simulation results and compare the proposed protocol with other replication protocols, Owner replication [10], Path replication [4], KS replication protocol [7]. We implemented these protocols and the proposed one on PeerSim simulator [6] and used it for the evaluation. PeerSim is a cycle-based simulator for Peer-to-Peer systems implemented in Java. 5.1

Environment and Configuration

We assume that the number of nodes |N | = 2000, and they are located uniformly at random in the 1000×1000 field. Each node is interconnected with the neighbor nodes that are within a certain communication range of it, hence the topology of a MANET is a random geometric graph. We executed 500 cycles for each simulation run. In the beginning, the storage residual of each node is set 100, and the battery residual of it is decided randomly from 50 to 100. We assume that 50 unique files are randomly located in nodes in the network at the beginning, and no more file is uploaded after that. The main purpose of the simulation is to observe the availability of these files and the impact of replication of them by different replication protocols on the storage and the network. Files in the network are replicated according to the requests for them. We have carried out simulations with three different configurations P = {100, 200, 300} for CSRP. 5.2

Performance Criteria

We conducted simulation runs to measure the following criteria. File Availability It indicates how many files are accessible on the network. Storage Cost It indicates how much replicated data consumes the storage of devices in the network.

92

T. Kurokawa and N. Hayashibara

Communication Cost It indicates how many packets are required for replication. In MANETs, each node communicates other nodes in a multi-hop fashion. We measure the number of hops that are required for replica allocation. We investigate the trade-off between file availability and the costs for storage and communication, and compare the performance of CSRP with other replication schemes. 5.3

File Availability

The primary purpose of data replication is to improve file availability. It indicates how many files are available compared to the total amount of files stored. File availability At at time t is represented in Eq. 4. At =

|Dt | |Doriginal | + |Dreplicated |

(4)

|Doriginal | and |Dreplicated | are the total number of the original files and that of the replicated ones, respectively. The total number of files available at time t t t | + |Dreplicated |. is denoted as Dt where |Dt | = |Doriginal In MANET, files can be disappeared by nodes’ leave. Thus, file availability is an essential criterion to evaluate replication protocols.

Fig. 1. File availability of different replication protocols.

We conducted simulation executions to measure the file availability of four replication protocols; Owner replication, Path replication, KS replication, and the proposed one. Figure 1a, 1b and 1c show the availability of files of different protocols and CSRP configures P = 100, 200 and 300, respectively. All three protocols have almost the same file availability until 100 cycles. Then, Owner and Path replication gradually drop file availability compared to CSRP and KS replication protocol. KS replication protocol is best regarding file availability among replication protocols that we use in the simulation. It resists a decline in file availability by continuing to create replicas of low demand files regularly. On the other hand, it keeps consuming resources such as storage space and network bandwidth.

Cost and Performance Analysis of Cuckoo Search Based File Replication

93

CSRP is almost the same as KS protocol halfway through and then starts to deteriorate its file availability. This is because CSRP stops replicating a low demand file at P cycles elapsed after the last request to the file for saving storage and communication resources. File availability of CSRP is proportional to the preservation period P according to the results shown in Fig. 1a, 1b and 1c. CSRP deteriorates 43.0%, 8.7% and 1.9% with P = 100, 200 and 300 compared to KS protocol at 300 cycles. CSRP enlarges on the gap with KS protocol over time. CSRP drops more than 80% with P = 100 and 200 and 50% with P = 300 compared to KS protocol. We show the costs for storage space and the network from the next section to analyze the trade-off between the file availability and the costs. 5.4

Storage Cost

The storage cost is crucial for replication protocols. Improvement of file availability usually imposes the cost of storage because file availability is proportional to the number of replicated files. In other words, there exists a tradeoff between them. We measure the average residual of storage of each node and the cumulative storage usage. Figure 2, 3 and 4 show the storage cost required by replication protocols with P = {100, 200, 300}, respectively. Each of them includes the cumulative storage occupancy and the average storage residual of devices in the network. Most of the files are low demand because we assume that the file access frequency obeys Pareto distribution. Figure 2b, 3b and 4b show the average storage residual of mobile devices in MANET. It decreases when a protocol starts replicating files. Then it recovers because of losing files by nodes’ leave and join. Although we deploy 50 files at the beginning of the simulation, files are uploaded continuously on MANET in a real situation. In this case, it repeats going down and up every time files are uploaded. CSRP with P = {100, 200} recovers the average storage residual quickly. Figure 2a, 3a and 4a show the cumulative storage occupancy with P = {100, 200, 300}. KS protocol is worse than others regarding storage ocuppancy because it keeps consuming storage space. It almost same as the result of CSRP with P = ∞. Thus, storage occupancy of KS protocol increases linearly. On the other hand, CSRP, Path, and Owner replication are bounded regarding storage occupancy. This property is very important in MANET because mobile devices are limited storage resources. There is a tradeoff between storage cost and file availability. According to the simulation result, we observed that CSRP requires 1.5 times storage cost to improve file availability by 46%. 5.5

Communication Cost

We analyze the cost for communication on replication in CSRP and other replication protocols.

94

T. Kurokawa and N. Hayashibara

Fig. 2. Storage cost of different replication protocols. P = 100 for CSRP.

Fig. 3. Storage cost of different replication protocols P = 200 for CSRP.

Mobile devices mostly consume their battery power through communication. Thus, this cost is also crucial in MANETs in terms of energy consumption. Figure 5 shows the communication cost in CSRP with P = {100, 200, 300}, KS, Owner and Path replication protocols. Owner and Path replication protocols are very efficient in communication because they determine the nodes to which replicated files are allocated every time a file is requested for download. Thus, they do not need extra communication to find the nodes. On the other hand, CSRP and KS protocol impose communication cost for finding eligible nodes to allocate replicated files. A comparison of the results of the proposed protocol and KS protocol shows that CSRP is better on the communication cost than KS replication in every configuration on P . Although both protocols have a substantial cost for replicating low demand files, KS protocol aggressively replicates such files in a whole execution, whereas CSRP stops replicating if there is no request for P cycles from the latest request for saving the cost for storage and communication.

Cost and Performance Analysis of Cuckoo Search Based File Replication

95

Fig. 4. Storage cost of different replication protocols P = 100 for CSRP.

Energy consumption is a crucial problem in MANETs. The communication cost of KS replication is not acceptable in terms of energy consumption even though the availability of low demand files is better than others. CSRP improves its file availability by 46% if P increases from 100 to 300 (see Fig. 1a and 1c). On the other hand, it requires 2.4 times communication cost for replication. However, CSRP is more than 60% efficient than KS protocol regarding communication cost.

Fig. 5. Communication cost of different replication protocols.

6

Conclusion

We measured file availability, storage cost, and communication cost of CSRP with several configurations of the preservation period P . Then we discuss the tradeoff between file availability and the costs of storage and communication. We also compared CSRP with other replication protocols regarding those criteria. According to the simulation results, file availability and costs for storage and communication can be compatible with CSRP by configuring P . CSRP improves its file availability by 46% if P increases from 100 to 300. On the other hand, it requires 1.5 times storage cost and 2.4 times communication cost. As a

96

T. Kurokawa and N. Hayashibara

result, the gap between CSRP and KS protocol regarding file availability reduces with preserving the superiority of CSRP regarding communication efficiency. Although the storage efficiency of CSRP reduces for improving file availability, the cumulative storage occupancy of CSRP is bounded.

References 1. Ali, A., Latiff, L.A., Fisal, N.: GPS-free indoor location tracking in mobile ad hoc network (MANET) using RSSI. In: 2004 RF and Microwave Conference (IEEE Cat. No. 04EX924), pp. 251–255 (2004). https://doi.org/10.1109/RFM.2004.1411119 2. Briesemeister, L., Hommel, G.: Localized group membership service for ad hoc networks. In: Proceedings of the International Conference on Parallel Processing Workshop, pp. 94–100 (2002) 3. Choi, J., Shim, K., Lee, S., Wu, K.: Handling selfishness in replica allocation over a mobile ad hoc network. IEEE Trans. Mob. Comput. 11(2), 278–291 (2012) 4. Cohen, E., Shenker, S.: Replication strategies in unstructured peer-to-peer networks. SIGCOMM Comput. Commun. Rev. 32(4), 177–190 (2002). https://doi. org/10.1145/964725.633043 5. He, X., Wang, F., Wang, Y., Yang, X.S.: Nature-Inspired Algorithms and Applied Optimization. Studies in Computational Intelligence, vol. 744. Springer, Heidelberg (2018) 6. Jelasity, M., et al.: Peersim (2016). http://peersim.sourceforge.net/ 7. Kageyama, J., Shibusawa, S.: Replication that prevents low-demand files from disappearing in P2P network. In: Proceedings of the First Forum on Data Engineering and Information Management (2009). in Japanese 8. Kurokawa, T., Hayashibara, N.: Data replication based on cuckoo search in mobile ad-hoc networks. In: Barolli, L., Hellinckx, P., Enokido, T. (eds.) Proceedings of the International Conference on Advances on Broad-Band Wireless Computing, Communication and Applications (BWCCA 2019). Lecture Notes in Networks and Systems, vol. 97, pp. 199–209 (2019). https://doi.org/10.1007/978-3-030-335069 18 9. Liu, J., Sacchetti, D., Sailhan, F., Issarny, V.: Group management for mobile ad hoc networks: design, implementation and experiment. In: Proceedings of the 6th International Conference on Mobile Data Management. Association for Computing Machinery, New York (2005). https://doi.org/10.1145/1071246.1071276 10. Lv, Q., Cao, P., Cohen, E., Li, K., Shenker, S.: Search and replication in unstructured peer-to-peer networks. In: Proceedings of the 16th International Conference on Supercomputing, ICS 2002, pp. 84–95. ACM, New York (2002). https://doi. org/10.1145/514191.514206 11. Osman, H., Taylor, H.: Managing group membership in ad hoc m-commerce trading systems. In: 2010 10th Annual International Conference on New Technologies of Distributed Systems (NOTERE), pp. 173–180 (2010) 12. Sakano, T., Kotabe, S., Komukai, T., Kumagai, T., Shimizu, Y., Takahara, A., Ngo, T., Fadlullah, Z.M., Nishiyama, H., Kato, N.: Bringing movable and deployable networks to disaster areas: development and field test of MDRU. IEEE Netw. 30(1), 86–91 (2016). https://doi.org/10.1109/MNET.2016.7389836 13. Yang, X., Suash Deb: Cuckoo search via l´evy flights. In: 2009 World Congress on Nature Biologically Inspired Computing (NaBIC), pp. 210–214 (2009). https:// doi.org/10.1109/NABIC.2009.5393690

A New DTN Relay Method Reducing Number of Transmissions Under Existence of Obstacles by Large-Scale Disaster Qiang Gao1(B) and Tetsuya Shigeyasu2 1

Graduate School of Comprehensive Scientific Research, Prefectural University of Hiroshima, Hiroshima, Japan [email protected] 2 Department of Management and Information System, Prefectural University of Hiroshima, Hiroshima, Japan [email protected]

Abstract. Efficient and precise relief activities in disaster affected areas are required immediately after disasters. In order to deal with disaster situations, delay/disruption-tolerant network is a good way to deliver messages to destination by use of multihop forwarding by mobile nodes. Hence, a method called MRD [6], Message Relay Decision has been proposed to reduce the number of transmissions. However, it can not cope with the existence of obstacles and this will degrades message delivery rate. This paper proposes a new routing protocol, MRDAI (MRD with Area Increase) which increases message relay area to avoid negative effect of obstacle. The evaluation results confirm that the MRDAI achieves higher performance under the existence of obstacle.

1

Introduction

Different from the traditional Internet which requires reliable end-to-end connection, a new type of network is born out of Inter Planetary Networks (IPNs)[1] which is one of Intermittently Connected Networks (ICNs). ICN has intermittent connection, low delivery rate, but high latency. As a new network architecture taking over the concept from ICNs, DTN (Delay/Disruption Networking) [2] has been proposed. In addition to the characteristics of IPN, DTNs can also be applied into terrestrial wireless networks: mobile wireless networks, Vehicular Ad hoc NETworks (VANETs) [5]. Those wireless networks are useful in case of disaster relief activities even in the situation that permanent network connections have lost. The existing communication infrastructures in such areas might have been damaged by the disaster and thus might not work. Moreover, even if their hardware is not directly damaged by the disaster, public communication infrastructures are likely to be congested because of attempts by the victims’ families to c The Editor(s) (if applicable) and The Author(s), under exclusive license  to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 97–107, 2021. https://doi.org/10.1007/978-3-030-61108-8_10

98

Q. Gao and T. Shigeyasu

seek information about them. In such a situation, a large number of victims lose their communication method. Hence, a method called MRD [6], Message Relay Decision has been proposed to deal with these situations by reducing the number of transmissions. It incorporates DTN-based message delivery into an autonomous wireless network construction package with intelligence, which is a wireless local area network-shelter network for sharing disaster information proposed in [7–9]. However, MRD just allows message forwarding only for nodes within message relay area. Message relay area consists of field within dth from idealroute1 . Hence, message delivery by MRD might be affected if obstacles exist within the message relay area. So, this paper calls this problem an obstacle blocking (See Fig. 1).

Fig. 1. Example of obstacle blocking

Since original MRD can not cope with the existence of obstacles and this will degrades message delivery rate. Therefore, we propose our new method on the basis of MRD. The routing protocol of the method is based on the Epidemic [10] which does not affect the basis of the core strategy of message forwarding. By new added obstacle detection procedure, the method can deliver messages to destinations while bypassing the obstacles. The rests of this paper are organized as follows. In Sect. 2, we summarize related work. In Sect. 3, we describe detailed procedure of our proposal. Then, results of performance evaluation are reported in Sect. 4, and Sect. 5 summarizes this paper.

2

Related Works

In this section, we describe previous works about DTN. Researchers have proposed an Autonomous wireless Network construction Package with Intelligence 1

idealroute is a straight line from a sender to a receiver.

A New DTN Relay Method

99

(ANPI) which uses independend wireless links established after a disaster [7– 9]. However, although ANPI can facilitate the sharing of disaster information among shelters, it lacks a unique process to improve the probability of message delivery to shelters in disaster areas. Moreover, given the instability of connections in the disaster network, the literature [3] introduced cognitive wireless network for a disaster information system based on dynamic network reconfiguration method to improve the connectivity. In the literature [11] the contact window for selecting the optimal message to forward at any opportunity was estimated. In the special DTN of the star network in [12], the overall message delay of star network intra-cell can be minimized by finding the different frequencies of the message sender to all nodes. [13] is based on publish-subscribe system by controlling buffer management and determining to select the right infostation or carrier on which to publish content based on establishment of temporal profile, which greatly improves delivery rate. By deploying MRD [6] in the disaster network, the number of mobile nodes relayed between transceiver nodes is limited, which also limits the number of transmissions in the entire network, reduces message redundancy and reduces unnecessary transmissions. None of the above agreements consider the special situation if there are obstacles in the narrow environment, but this is extremely bad influence.

3

Proposal

In this section, we proposed a new routing protocol, MRDAI (MRD with Area Increase), which increase message relay area to avoid effect of obstacle. In our proposal, relay nodes in message relay area can automatically detect if obstacles appears in the relay area, which is called as “obstacle blocking”. Then nodes adopt corresponding measures to deal with this special case, which is to increase range of original message relay area even if the obstacles completely block the area. Through the experiments we found that the proposed scheme not only resists obstacles, but also improves performance compared to the previous schemes, but the number of transmissions also has increased a lot, so we again proposed Sub-Relay Stations (SRSs) to reduce unnecessary connection exhaustion. It is based on the establishment of inferior transfer coordinates on the basis of bypassing obstacles in node communication, and re-establishing the message relay area to reduce the number of transmissions. 3.1

MRDAI

In order to avoid completely message blocking by obstacles, this paper proposes MRDAI. MRDAI enlarges the message relay area by increasing dth, to bypass obstacle.

100

Q. Gao and T. Shigeyasu

3.1.1 How to Find Obstacles MRDAI judges the existence of obstacle by Eq. (1). By a nature of DTN, any message will be duplicated at relay nodes. Then, multiple copies of original message arrive at destination. In Eq. (1), H(i, j) is a set of hop count of each delivered copy of ith message from node j. Here, Hmax is an expected maximum hop count of H(i, j), and it is a larger value of two, Hs1 , Hs2 as Eq. (2). max{x | x ∈ H(i, j)} ≥ Hmax

(1)

Hmax = max(Hs1 , Hs2 )

(2)

By Eq. (2), MRDAI decides that there is an obstacle in message relay area when a max hop count of the delivered copies are larger than both of estimation (Hs1 ) and actual measurement (Hs2 ).

Fig. 2. Calculation of Nof f set

1. Hs1 : Hop count estimation Let us consider the condition that the nodes are uniformly distributed in EntireArea appeared in Eq. (3). The model is shown in Fig. 2. i and j are the number of nodes in the vertical and horizontal distribution of message relay area, respectively. By using these two values to compare with the length and width of message relay area (similarity principle), the value of j can be obtained. i will keep increasing from zero, and the same time j becomes smaller. When β becomes the minimum value, we get the optimal value of i and j, so the number of nodes in the horizontal distribution is j. j i − | Wrange Lrange M essageRelayArea s.t. i × j = Ntotal × EntireArea min β =|

(3)

A New DTN Relay Method

101

β is a coefficient to decide number of j. Wrange is a width of message relay area. Lrange is a length of message relay area. Ntotal is a total number of mobile nodes. We use j to get the distance between distributed nodes Nof f set as shown in Eq.(4). Lrange (4) Nof f set = j Following the above part, we calculate Fwd in Eq. (5), and Hs1 in Eq. (6). Here, D(S,R) means distance between sender and receiver.   Rtransmit , Nof f set < Rtransmit Nof f set Fwd = (5) 1, Nof f set ≥ Rtransmit  Hs1 =

D(S,R) Fwd × Nof f set

 (6)

2. Hs2 : Realistic threshold In an actual operation, the average number of successful delivery hops obtained by continuous statistics as shown in Eq. (7).  x Hs2 =

x∈H(i,j)

| H(i, j) |

(7)

Hs2 is hop count of strategy 2, average hops of actual total hops the message delivered successfully. Hactual is actual total hops. If Eq. (1) holds, it is considered that there is an obstacle in this network range, the node can increase the length of dth, and propagate the message of new length of dth to the surrounding neighbors through the signal, and finally spread to the entire network, so that all nodes know the existence of obstacles. 3.1.2 Limitation of Forwarding Direction The message carried by the current node will be forwarded only to the next node within the range of plus or minus 90◦ to avoid redundancy. 3.2

SRS (Sub-Relay Station)

SRS is a second proposal based on MRDAI. SRS consists of two steps. First, SRS decides Relay Coordinates (RC), and second, the sub-relay area is calculated based on RC. After that, the nodes abandon the message relay area calculated by MRDAI to transmit messages, and instead transmits messages within the sub-relay area.

102

Q. Gao and T. Shigeyasu

Fig. 3. Relay coordinates: K = 2

3.2.1 Getting Coordinates of RC We only heuristically select a series of coordinates included by top 10 messages arriving at a destination. Then we can calculate the cluster center point by using these coordinates with K-means as shown in Fig. 3. The number of cluster center point, K, is determined by total number of dth increment spontaneously. 1. K-means introduction The constant K means the final number of clustering categories. First, a method randomly select the initial point as the centroid, and identifies the sample points to the most similar classes according to Euclidean distance. Then, recalculate the centroid of each class (i.e., the class center), the centroid of each class finally can be determined by repeating above process until the centroid does not change. 2. Number of RC Based on scheme of dth increment in MRDAI, it alternatively increases both sides of relay area. We assume one dth increment after the length of both sides is increased once. Equation (8) shows that total number of dth increment, DItotal , as total number of classes into which we need to divide a series of coordinates. (8) K = DItotal 3.2.2 Establish Sub-relay Area By use of a series of obtained RC, the method sets up the sub-relay area in order from the sending node to the destination node. After the sub-relay area is established, there is no possibility of obstacles in the middle of forwarding route as shown in Fig. 4.

A New DTN Relay Method

103

Fig. 4. Example of sub-relay area

4

Performance Evaluation

In this section, we evaluate the performance of proposed method by computer simulation. 4.1

Comparison with 3 Schemes

Evaluation results have been conducted with the parameters as shown in Table 1. Table 1. Simulation parameters Simulation period

10,000 (s)

Number of mobile nodes

400

EntireArea

15001500 (m)

dth

200 (m)

D(S,R)

400 (m)

Buffer size

12.5 (MB)

Message creation interval 30 (s) Transmit speed

250 (kB/s)

Transmit range

100 (m)

Movement model

RandomWalk

From Fig. 5, we can see the different delivery ratio among SRS, MRDAI and MRD. All of which keep decreasing trends with increasing obstacle ratio. However, even if the obstacle ratio is 100%, SRS and MRDAI can still deliver messages to destination. In contrast, MRD can not send the messages to the destination when obstacle ratio is greater than 70%. Moreover, in case of that obstacle ratio is 100%, the delivery ratio of improved method SRS is 2 times higher than MRDAI.

104

Q. Gao and T. Shigeyasu

Fig. 5. Delivery ratio – obstacle ratio

Fig. 6. Average delay – obstacle ratio

According to Fig. 6, because of no-arrival of MRD when obstacle ratio is greater than 70%, the average delay of MRD becomes approx. 10,000s. As a result, average delay of MRD becomes larger than MRDAI and SRS. We can see that number of transmissions of MRD is much smaller than MRDAI. This phenomenon does not imply that MRD outperforms MRDAI and SRS. It just cause no successful message delivery on MRD. It is worth mentioning that when the obstacle rate is 100%, compared with MRDAI, SRS has only approx. 60% of its number of transmissions which greatly saves energy consumption.

A New DTN Relay Method

105

Fig. 7. Number of transmissions – obstacle ratio

4.2

Delivery Ratio of SRS Among Different Buffer Sizes

Through the previous chapter, we clarified that SRS reduces both latency and saves energy consumption. The most gratifying thing is that it did not come at the cost of lowering the delivery rate, on the contrary, it maintains a higher delivery rate compared to MRD. We have high expectations for SRS. Therefore, this chapter discusses the changes in delivery rate of SRS under different buffer sizes. Table 2 is the detailed data about the delivery rate from 12.5M to 50M. Table 2. Delivery ratio of SRS among different buffer sizes Obstacle ratio Delivery ratio 12.5M 18.75M 25M 31.25M 37.5M 43.75M 50M 0.0

0.94

0.94

0.93 0.90

0.91

0.92

0.91

0.1

0.89

0.89

0.88 0.91

0.90

0.88

0.90

0.2

0.88

0.88

0.88 0.89

0.89

0.87

0.87

0.3

0.85

0.84

0.88 0.85

0.85

0.84

0.85

0.4

0.86

0.91

0.91 0.83

0.84

0.89

0.87

0.5

0.68

0.73

0.79 0.83

0.81

0.81

0.82

0.6

0.67

0.71

0.82 0.84

0.79

0.79

0.77

0.7

0.70

0.78

0.79 0.78

0.78

0.77

0.78

0.8

0.67

0.67

0.81 0.78

0.81

0.75

0.76

0.9

0.70

0.77

0.74 0.75

0.79

0.76

0.77

1.0

0.66

0.76

0.63 0.67

0.71

0.72

0.75

106

Q. Gao and T. Shigeyasu

It can be seen that as the size of buffer increases, the delivery rate does not change significantly, but on the other side, the delivery rate gradually stabilized with obstacle rate increasing, indicating that the scheme has better robustness.

5

Conclusion

In this paper, we discussed how to overcome the weakness of MRD for emergency message delivery. The evaluation results confirm that our new proposal, SRS and MRDAI well educes performance of DTN message delivery under the existence of obstacles, but SRS has a better performance than MRDAI. For evaluating the availabilities and improvement of our proposal, we conducted the results by computer simulation comparing our proposals and the conventional MRD. By the results conducted the computer simulations, we clarified the advantages of our proposal than the conventional in terms of delivery ratio, average delay and number of transmissions. In particular, our proposal well improved the delivery ratio under the existence of obstacles. In the future, we will modified our method for achieving higher delivery ratio.

References 1. Caini, C., Cruickshank, H., Farrell, S., Marchese, M.: Delay and disruption-tolerant networking (DTN): an alternative solution for future satellite networking applications. Proc. IEEE 99(11), 1980–1997 (2011) 2. Fall, K., Farrell, S.: DTN: an architectural retrospective. IEEE J. Sel. Areas Commun. 26(5), 828–836 (2008) 3. Sekin, Y., Uchida, N., Shibata, Y., Shiratori, N.: Disaster information network based on software defined network framework. In: 2013 27th International Conference on Advanced Information Networking and Applications Workshops, Barcelona, pp. 237–242 (2013). https://doi.org/10.1109/WAINA.2013.226. 4. Hui, P., Chaintreau, A., Scott, J., Gass, R., Crowcroft, J., Diot, C.: Pocket switched networks and human mobility in conference environments. In: ACM WDTN 2005. Pennsylvania, USA, Philadelphia (2005) 5. Pereira, P., Casaca, A., Rodrigues, J., Soares, V., Triay, J., Cervellopastor, C.: From delay-tolerant networks to vehicular delay-tolerant networks. IEEE Commun. Surveys Tuts. PP(99), 1–17 (2011) 6. Kawamoto, M., Shigeyasu, T.: Message Relay Decision Algorithm to Improve Message Delivery Ratio in DTN-Based Wireless Disaster Information Systems. IEEE (2015) https://doi.org/10.1109/AINA.2015.275 7. Kamegawa, M., Kawamoto, M., Shigeyasu, T., Urakami, M., Matsuno, H.: A new wireless networking system for rescue activities in disasters - System overview and evaluation of wireless nodes. In: Proceedings of 19th International Conference on Advanced Information Networking and Applications Workhops, vol.2, pp. 68–71 (2005) 8. Ohtaki, R., Shigeyasu, T., Urakami, M., Matsuno, H.: A computer system for exchanging disaster information inter shelters - an algorithm for wireless node placement with the consideration of physical property of disaster damaged area (In Japanese). IPSJ J. 52(1), 308–318 (2011)

A New DTN Relay Method

107

9. Urakami, M., Inami, S., Kamekgawa, M., Shigeyasu, T., Matsuno, H.: Wireless distributed network system for relief activities after disasters - improvement of ANPI and ADES-plus systems. Int. J. Space-Based Situated Comput. 1(4), 233– 243 (2011) 10. Vahdat, A., Becker, D.: Epidemic routing for partially-connected ad hoc networks. Duke University Technical Report Cs-2000-06. Technical report (2000) 11. Sandulescu, G., Nadjm-Tehrani, S.: Opportunistic DTN routing with windowaware adaptive replication. In: ACM AINTEC 2008, 18–20 November 2008, Bangkok, Thailand (2008) 12. Roy, S., Tomasi, D., Conti, M., Bhusal, S., Roy, A., Li, J.: Optimizing message ferry scheduling in a DTN. In MobiWac 2018: 16th ACM International Symposium on Mobility Management and Wireless Access, 28 October–2 November 2018, Montr´eal, Qu´ebec, Canada, p. 5. ACM, New York (2018). https://doi.org/10.1145/ 3265863.3265884 13. Giuseppe, S., Mirco, M., Cecilia, M.: TACO-DTN: a time-aware content-based dissemination system for delay tolerant networks. In: ACM MobiOpp 2007, 11 June, San Juan, Puerto Rico, USA (2007)

Performance Comparison of Multi-class SVM with Oversampling Methods for Imbalanced Data Classification Seunghyun Park1 and Hyunhee Park2(B) 1

2

Korea University, Seoul, South Korea [email protected] Myongji University, Yongin-si, Gyeonggi-do, South Korea [email protected]

Abstract. Network traffic data generally comprise a major amount of normal traffic data and a minor amount of attack data. Such an imbalance in the amounts of the two types of data leads to issues such as low prediction performance including misclassifications owing to the estimation bias toward minority data and anomalies. To address this problem, several minority data synthesis models based on the synthetic minority oversampling technique algorithm have been developed. However, in recent years, studies have been actively conducted to synthesize minority data using the newly developed generative adversarial network (GAN) model. In this paper, we examine a GAN based oversampling model to address the data imbalance problem associated with intrusion detection data and compares the performance of the oversampling models. Therefore, the GAN based oversampling model can generate data of a class which has a small number of data so that the problem induced by imbalanced class distribution can be mitigated, and classification performance can be improved. Simulation results using KDD Cup 99 dataset show that the oversampling method using GAN algorithm is effective and that it is superior to existing oversampling methods.

1

Introduction

When typical machine learning algorithms are applied, it is assumed that equal amounts of the training dataset is included in each class during model training. However, several current issues related to the use of these algorithms originate from the problem of data imbalance. Furthermore, data that belong to the minority class tend to be incorrectly classified compared to those that belong to the majority class [1]. Such issues occur because the existing machine learning algorithms are designed to optimize the overall performance and therefore does not consider the relative distribution of each class. For this reason, these types of issues are mostly identified in classifiers such as a decision trees and multi-layer perceptron [1,2]. In the case of attack data, which is used to detect attacks in networks, the amount of the data in each class significantly varies based on c The Editor(s) (if applicable) and The Author(s), under exclusive license  to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 108–119, 2021. https://doi.org/10.1007/978-3-030-61108-8_11

Multi-class SVM with Oversampling for Imbalanced Data Classification

109

the properties of the traffic, which is also considered a typical imbalanced data problem. Data sampling methods are the most commonly used methods for addressing the class imbalance problem. Through these methods, the number of samples of the minority class, which accounts for a small proportion in an imbalanced dataset, or the majority class, which accounts for a large proportion in the dataset, is adjusted to obtain a balanced dataset. These methods are classified into two types, undersampling and oversampling method, based on the number of minority and majority classes to be adjusted [3]. Undersampling is used to remove a certain number of samples of the majority class to make the number of samples consistent to that of the majority class. Examples of undersampling method include random undersampling, in which some samples of the majority class are randomly removed, and EasyEnsemble [4], in which the learning process of a classification model is executed based on a subset independently extracted from the samples of the majority class and a set of the minority class. However, as these methods involve the elimination of some data, some information may be lost. Unlike undersampling methods, oversampling methods are used to generate samples of the minority class based on the number of samples in the majority class. Therefore, with these methods, loss of information can be avoided. Some examples of the oversampling method include random oversampling, in which samples of the minority class are randomly generated, the synthetic minority oversampling technique (SMOTE) [5], in which neighbors in the samples of minority class are searched and new samples between the neighbors are synthesized using the k-nearest neighbors (k-NN) algorithm, adaptive synthetic sampling (ADASYN) [6], in which a weight is added to the SMOTE depending on the density of the majority class around the minority class, and the BorderlineSMOTE [7], in which SMOTE synthesis is performed based on the boundary between the samples of minority and majority class. A deep learning-based generative adversarial network (GAN) algorithm [8] comprises a generator and discriminator. The generator produces synthetic data similar to original data to deceive the discriminator. The discriminator generates synthetic images similar to the original images [9]. During GAN-based oversampling, the distribution of minority class is learned to generate similar data. In addition, data of higher quality can be generated as the discriminator examines the similarity between original and generated data. In this regard, the GAN-based models can overcome the limitations of existing models that generate data based on k-NN algorithm. A Conditional Tabular GAN (CTGAN) [10] algorithm has been designed to apply the function of GAN algorithm to structured data. It utilizes the advantages of both conditional GAN (CGAN) [11] and tabular GAN (TGAN) [12] algorithms and supports oversampling packages specialized for structured data. In this paper, the CTGAN algorithm is used to address the class imbalance problem associated with intrusion detection data. Moreover, the performance of oversampling based on GAN algorithm is compared to the existing oversampling

110

S. Park and H. Park

methods such as the SMOTE, Borderline-SMOTE and ADASYN. To conduct simulation results on intrusion detection, the KDD Cup 99 dataset [13], which is often used in intrusion detection system evaluations, is used to evaluate the performance of each algorithm.

2

Oversampling Method

Intrusion detection data are the most common type of imbalanced data. Resampling refers to the process by which such a data imbalance is addressed during preprocessing. As the performance of a prediction model is affected by the resampling process [14], a resampling model should be applied to the data after its properties are analyzed. The GAN [8] algorithm is an unsupervised learning algorithm based on deep learning. In this algorithm, a generator that creates fake data similar to the original data by learning data distribution competes with the discriminator that examines whether the data are original or created by the generator. Particularly, in the GAN model, modeling is performed to generate similar data and synthetic data that are difficult to be distinguished from original data. However, generation of structured data is more challenging as compared to generating images, owing to the following reasons: • First, there are various types of structured data according to the columns. • Second, structured data exhibit a non-Gaussian distribution. Pixel values of images follow a Gaussian distribution and can be normalized via transformation. However, structured data have independent features. • Third, the mode of distribution should be estimated as each column of the structured data has a different distribution shape. • Fourth, the generator might fail to learn the distribution of a significantly imbalanced data owing to the omission of the data of the minority class. • Lastly, a dependent variable is implemented through one-hot encoding. The generator learns the distribution of all the classes and generates a dependent variable using Softmax. However, when a true value is obtained through one-hot encoding, the discriminator distinguishes between the original and fake data based on the status of one-hot encoding performed to a dependent variable instead of the similarity between the data [15]. In contrast, the CTGAN model comprises the properties of both conditional-GAN and Tabular-GAN algorithms and can therefore be optimized to generate structured data. In this algorithm, a variational Gaussian mixture (VGM) distribution is used to estimate the mode of data distribution by column. In the training process based on this algorithm, VGM is applied to estimate the mode and reflect both multi-modal and nonGaussian distributions.

3 3.1

CTGAN Training Based on Intrusion Detection Data Dataset

The KDD Cup 99 dataset was developed, similar to the DARPA 1998 dataset, for the 99 KDD Cup competition [13]. It includes feature extracted through the

Multi-class SVM with Oversampling for Imbalanced Data Classification

111

analysis of time and sequence windows in data traffic as well as attacks such as Neptune-DoS, pod-DoS, Smurf, and buffer overflow. In particular, it is a set of log data that occur in wired networks, and it contains a total of 42 features. Different configurations of this dataset are available with variation in number of instances but the number of attributes in each case is 42. The attribute labeled 42 in the data set is the ‘class’ attribute which indicates whether a given instance is a normal connection instance or an attack. Table 1 and Table 2 give the description of KDD dataset attributes with basic features (i.e. No. 1 to no. 9), content features (i.e. No. 10 to no. 22), traffic features (i.e. No. 23 to no. 31) and host features (i.e. No. 32 to no. 42) included in the KDD Cup 99 dataset, respectively. Out of these 42 attributes, 41 attributes can be classified into four different classes as discussed below: • Basic features are the attributes of individual Transmission Control Protocol connections. • Content features are the attributes within a connection suggested by the domain knowledge. • Traffic features are the attributes computed using a two-second time window. • Host features are the attributes designed to assess attacks which last for more than two seconds.

Table 1. Basic features of individual Transmission Control Protocol connection No. Feature name

Description

1 2 3 4 5 6 7 8 9

Length (number of seconds) of the connection Type of the protocol, e.g. tcp, udp, etc. Network service on the destination, e.g., http, telnet, etc. Number of data bytes from source to destination Number of data bytes from destination to source Normal or error status of the connection 1 if connection is from/to the same host/port; 0 otherwise Number of “wrong” fragments Number of urgent packets

3.2

Duration Protocol type Service src bytes dst bytes Flag Land wrong fragment Urgent

Data Preprocessing

Basically, all the features can be used as independent variables for input in machine learning algorithms for intrusion detection. However, some features have a significant influence on the prediction model compared to others. Moreover, overhead occurs when all the features are extracted; thus, feature selection should be performed to exclude features that are unlikely to affect the prediction model.

112

S. Park and H. Park Table 2. Detail of KDD dataseta attribute from no. 10 to no. 42

No. Feature name 10 hot 11 num failed logins 12 logged in 13 num compromized 14 root shell 15 su attempted 16 num root 17 num file creations 18 num shells 19 num access files 20 num outbound cmds 21 is hot login 22 is guest login a Detail: KDD Cup 99 kddcup99.html)

No. Feature name 23 24 25 26 27 28 29 30 31

count serror rate

No. Feature name

32 33 34 rerror rate 35 same srv rate 36 diff srv rate 37 srv serror rate 38 srv rerror rate 39 sev diff host rate 40 41 42

dst host dst host dst host dst host dst host dst host dst host dst host dst host dst host class

count srv count same srv rate diff srv rate same src port rate srv diff host rate serror rate srv serror rate rerror rate srv rerror rate

(Available: http://kdd.ics.uci.edu/databases/kddcup99/

Accordingly, in this paper, the data classification of Waikato Environment for Knowledge Analysis (WEKA), which is a data mining tool, is used to obtain the independent variables including only the feature variables and intrusion detection prediction is performed. For the extraction of independent variables, feature selection based on a Pearson correlation coefficient is performed. When all the features are used for security attack detection using a multi-class support vector machine (SVM), the level of complexity increased during detection. Further, the number of errors can also increase when determining a hyperplane. For this reason, only basic and content features among the 41 features are extracted and used to detect security attacks efficiently in the simulation conducted in this paper [16]. As presented in Table 3, the KDD Cup 99 dataset contains 22 training attack types. In this paper, mapping of these attack types is performed based on four classes: denial-of-service (DoS), probing, user to root (U2R), and remote to local (R2L) [17]. Specifically speaking, a DoS attack refers to an attack that is performed to prevent the use of service. A probing attack is a security threat attack that attempts to collect information on the status of a system before an actual attack takes place. An U2R attack attempts to acquire administrative privileges. An R2L attack is delivered by an unauthorized user to obtain access privileges from an external channel. In this paper, data preprocessing is conducted using MySQL, which is a relational database. To assign each dependent variable with a unique number, one-hot encoding is applied, in which a categorical variable is represented as binary vectors. In this method, the size of the word set is considered to be the length of a one-hot encoded vector. Subsequently, a value ofone is assigned to the index of the variable to be represented and zero to the

Multi-class SVM with Oversampling for Imbalanced Data Classification

113

other variables. Consequently, DoS, probing, U2R, R2L, which serve as dependent variables, are given unique vector representation values through one-hot encoding. Table 3. Attack types of the KDD Cup 99 dataset Attack class

Type of security attack

Denial of Service (DoS) Smurf, Pod, Land, Teardrop, Back, Neptune Probing

Portsweep, Ipsweep, Nmap, Satan

User to Root (U2R)

Perl, Buffer overflow, Loadmodule, Rootkit

Remote to Local (R2L) Spy, Phf, Gess pass, Multihop, Imap, Warezmaste, Wareclient, FTP write

In addition, feature scales of independent variables, with different ranges, are normalized by preprocessing the dataset. More specifically, linear transformation is applied to the entire dataset and the data distribution is adjusted to obtain a mean of zero and variance of one. Scaling prevents the overflow or underflow of data and decreases the number of conditions of a covariance matrix of an independent variable, thereby increasing stability and the rate of convergence during optimization. In this paper, scaling is performed using StandardScaler of Scikit-learn. Next, from the KDD Cup 99 dataset, the training and testing datasets are created. In the simulation conducted in the present study, 10% of the training dataset, which is equivalent to 86,678 data records, is used. Table 4 presents the security attacks that form the KDD Cup 99 training dataset. To measure the performance of the security attack detection, 10% of the KDD Cup 99 testing dataset, which is equivalent to 21,409 data records, is used. Table 4. Number of flows in class

3.3

Class

Number of flows

Normal DoS Probing U2R R2L Total

78,010 3,712 3,796 35 1,125 86,678

Multi-class Support Vector Machine with Oversampling

The SVM is a set of supervised learning methods used for classification and regression. The SVM is originally developed for binary decision problems, and

114

S. Park and H. Park

its extension to multi-class problems is not straightforward. How to effectively extend it for multi-class classification is still an on-going research issue. Multiclass SVM is usually implemented by combining several binary SVMs [18]. In this paper, the multi-class SVM algorithm is implemented to analyze the accuracy of intrusion detection based on the KDD Cup 99 dataset in which oversampling is performed. A Gaussian function, which generally shows a high level of accuracy, is applied as a kernel function for the multi-class SVM algorithm. Training is conducted for up to 13,000 epochs until a similar distribution could be generated for each class. However, in the case of the U2R attack class, even if the number of learning is increased, data deviating from the distribution of the existing data is generated.

Fig. 1. Simulation results of Multi-class SVM.

Figure 1 depicts results of a confusion matrix based on multi-class SVM without oversampling method for security attacks. As can be seen, normal flows of 44.64% are included in total test flows and 39.41% flows are correctly determined as normal flows in the multi-class SVM algorithm. Meanwhile, normal flows of 5.23% are incorrectly determined as U2R attacks. As for DoS attacks, DoS attack flows of 19.15% are included in test flows and 17.14% flows are correctly determined as DoS attack flows. Additionally, 2.01% out of DoS attack flows are incorrectly determined by the multi-class SVM algorithm as normal flows. Probing attack flows of 20.03% are included in total test flows and 16.32% flows are correctly determined as normal flows in the multi-class SVM algorithm. On the other hand, probing attack flows of 3.71% are incorrectly determined as normal flows. As for R2L attacks, R2L attack flows of 8.34% are included in test flows and 8.34% flows are correctly determined as R2L attack flows. Finally, U2R attack flows of 7.84% are included in test flows and 4.35% flows are correctly

Multi-class SVM with Oversampling for Imbalanced Data Classification

115

determined as U2R flows in the multi-class SVM algorithm. Meanwhile, U2R attack flows of 3.49% are incorrectly determined as normal flows. Based on the aforementioned results, it can be concluded that false positive detections occur due to U2R attacks; in such attacks, the imbalances present in the dataset are highly significant. To address this problem, the KDD Cup 99 dataset normalized through preprocessing is applied to the oversampling algorithm. Before performing prediction using a classification model, the duplication rates of the synthetic data are compared; a high duplication rate indicates an increase in the amount of minority data and no increase in the information included in the data. For this reason, an improved performance must not be expected in such a case. Moreover, duplication of synthetic data leads to overfitting. Therefore, the quality of sampling can be determined by examining the amount of data duplicated by oversampling. The hyperparameter used in the oversampling method and classifier is as follows. k-NN-based SMOTE, Borderline-SMOTE, and ADASYN both used k = 5. In the case of CTGAN, both the generator and classifier consisted of two hidden layers, and the activation function and optimizer used ReLU [19] and Adam [20], respectively. The SVM used as the classifier is set to C = 1.0, the kernel is set to the Radial Basis Function.

Fig. 2. Simulation results of Multi-class SVM with CTGAN oversampling.

Figure 2 shows results of a confusion matrix based on multi-class SVM with CTGAN oversampling method. As can be seen, normal flows of 33.88% are included in total test flows and 32.49% flows are correctly determined as normal flows in the multi-class SVM algorithm. Meanwhile, normal flows of 1.39% are incorrectly determined as U2R attacks. As for DoS attacks, DoS attack flows of 20.83% are included in test flows and 19.34% flows are correctly determined

116

S. Park and H. Park

as DoS attack flows. Additionally, 1.49% out of DoS attack flows are incorrectly determined by the multi-class SVM algorithm as normal flows. Probing attack flows of 20.38% are included in total test flows and 18.55% flows are correctly determined as normal flows in the multi-class SVM algorithm. On the other hand, probing attack flows of 1.83% are incorrectly determined as normal flows. As for R2L attacks, R2L attack flows of 13.23% are included in test flows and 13.21% flows are correctly determined as R2L attack flows. Also 0.02% out of R2L attack flows are incorrectly determined as normal flows. Finally, U2R attack flows of 11.68% are included in test flows and 10.39% flows are correctly determined as U2R flows in the multi-class SVM algorithm. Meanwhile, U2R attack flows of 1.29% are incorrectly determined as normal flows. Eventually, the overall ratio of the dataset is constructed similarly through CTGAN based oversampling, thereby increasing prediction accuracy. The reason is that in the case of U2R with the smallest number of dataset, the prediction accuracy is only 55.48% before oversampling, but the prediction accuracy after oversampling is significantly increased to 88.95%. High rate of data duplication means that the number of small data increases, but the information (e.g., features such as distribution of data) of the data does not increase. Therefore, it is difficult to expect performance improvement due to high rate of data duplication. In addition, overlapping of synthetic data is a factor of overfitting. Therefore, it is possible to determine the quality of each sampling by checking how much duplicate data is generated due to oversampling. Table 5 presents the rates of data duplication due to oversampling. Train refers to existing training data. The SMOTE algorithm exhibits the highest rate of data duplication at 15.783%. The rate of data duplication is calculated to be 9.274% for the Borderline-SMOTE algorithm, 6.893% for the ADASYN algorithm, and 2.489% for the CTGAN algorithm. Table 5. Number of duplicate data per oversampling Data type

Ratio of duplication

Train 0% 15.783% SMOTE Borderline SMOTE 9.274% 6.893% ADASYN 2.489% CTGAN

After all, it can be seen that oversampling method based CTGAN is effective in the multi-class SVM algorithm. For a more detailed comparison, the Wilcoxon signedrank test [21] is performed. The Wilcoxon signed-rank test is used to verify the null hypothesis that there is no significant difference between the two classifiers. If the p-value is less than the significance level, the null hypothesis is rejected and the two classifiers are decided to have significant differences. Table 6

Multi-class SVM with Oversampling for Imbalanced Data Classification

117

shows the results of the Wilcoxon signed-rank test with the significance level set to 0.05 in this paper. Table 6. Result of Wilcoxon signed-rank test Oversampling method

p-value ( P) Once the behavior is determined by Eq. (2), each parameter is updated by Eq. (3). In order to do differential by the variable representing selected behavior by the Eq. (1), only

DTN Routing Protocol Using Reinforcement Learning

193

the parameters associated with the selected behavior are updated. Where α represents the learning rate. θn ← θn + α ∗ δ ∗

∂Q(x, a, θ ) θn

(3)

The δ in Eq. (3) is calculated by Eq. (4). Where r is the reward obtained from the behavior and γ is the decay rate. The obtained reward is a positive if the message is reached. If the message is not reached, obtained reward is a negative. If neither, the reward is calculated as zero. δ = rt+1 + γ Q(st+1 , at+1 , θ ) − Q(st , at , θ )

(4)

The above calculations are repeated multiple times until the training is completed, and the parameters are updated with each behavior during the simulation. 3.3 Judgement of Duplication Priority Control The learning flow of the duplication priority control judgment is the same as the duplication judgment described above. In this section, we show the difference between the judgment of duplication and the judgment of duplication priority control. The first point of difference is the parameters and behaviors to be learned. The behavioral valuation value function is shown in Eq. (5). θ is the same as the judgment of duplication, and is the learning parameters in the behavioral valuation value function. x1 is own storage utilization minus the storage utilization of the encountered node. x2 is the average TTL in the own possession messages minus the average TTL in the encountered node’s messages. x3 is the average number of duplications in own possession messages minus the average number of duplications in the encountered node’s messages. x4 is the average size in the own message minus the average size in the message of the encountered node. a0 is the behavior for duplicating from the message of TTL with big number in order to duplicate from the message of longest validity time. a1 is a behavior for duplicating from messages of the number of duplications with small in order to duplicate from messages not yet spread across the network a2 is a behavior for duplicating from messages with small size in order to allow more messages to be duplicated. Q(x, a, θ) = (θ1 + θ2 x1 + θ3 x2 + θ4 x3 + θ5 x4 )a0 + (θ6 + θ7 x1 + θ8 x2 + θ9 x3 + θ10 x4 )a1 + (θ11 + θ12 x1 + θ13 x2 + θ14 x3 + θ15 x4 )a2

(5)

When the value of the duplicated behavior is judged to be high in the behavioral valuation value function for the duplication judgment, move to the judgment of duplication priority control. In the judgment of duplication priority control, we follow the same procedure as in the determination of duplication, using the ε-greedy method in Eq. (2), select a behavior. Then, by updating the parameters for the result of the behavior by Eq. (3) and Eq. (4), learn the behavior of duplication in response to network states.

194

K. Henmi and A. Koyama

4 Performance Evaluation In this section, we have evaluated the performance of the proposed method, PRoPHETv2, and MaxProp. 4.1 Simulation Condition The simulation conditions are shown in Table 1. The simulation is assumed the case of information sharing using the DTN. In the simulation, there are three different types of nodes with different moving speed and frequency, and each node communicates using IEEE802.11b. Total number of nodes are 180 nodes, and each kinds of nodes are 60 nodes. Messages are generated during the simulation at 20 s intervals randomly determining which nodes to generate, and the message size is randomly determined Table 1. Simulation conditions Parameters

Value

Simulation time [s]

3600

Storage size [GB]

1

transmission speed [Mbps]

10

transmission area [m]

50

Message generation interval [s]

20

Message Destination

Choose randomly from walking nodes

Minimum message size [kB]

500

Maximum message size [MB]

10, 20, 30, 40, 50

TTL [min]

30

Table 2. Node conditions Group

Parameter

Value

Walk

Speed [m/s]

0.5–1.5

Wait time [s]

0–120

Number of nodes 60 Bicycle Speed [m/s] Wait time [s]

2.7–8.3 0–30

Number of nodes 60 Car

Speed [m/s]

2.7–16.6

Wait time [s]

0–30

Number of nodes 60

DTN Routing Protocol Using Reinforcement Learning

195

between the minimum message size and the maximum message size. The destination node is randomly selected from the 60 nodes in the walk group shown in Table 2 per each message. Next, the simulation used two types of movement models: one was a 2 km square area with the random waypoint model. The other model is a model in which each node moves along the path of the map. The used map is shown in Fig. 2.

Fig. 2. Map (Tokyo)

In evaluation of the proposed method, we used the functional approximation of the state behavioral evaluation value function, which is the training results of 1000 iterations of simulations with a maximum message size of 50 MB in the random waypoint model. 4.2 Simulation Results As the evaluation items, we evaluated the message delivery ratio and overhead. First, the message delivery rate and overhead of the random waypoint model are shown in Figs. 3 and 4. The proposed method resulted in the highest message delivery ratio. However, compared to MaxProp, the proposed method improved the message delivery ratio by at most 2% and overhead is degraded by up to 15. Ideally, by judgement the duplication node, it should be possible to reduce overhead and improve message delivery ratio. Examining the selected tendencies at learning in the proposed method, we found that early on, it chooses a behavior that was not duplicated, with the learning progress, it gradually became more active in duplicating. In other words, the proposed method learned that duplicating was a more valuable behavior, no matter what the destination was. This trend can be seen from the fact that the performance of PRoPHETV2, which duplicates only when it has a higher probability of being reached than other node, has a degraded message delivery ratio than MaxProp and the proposed method. For the duplication priority control judgment behaviors, the most valuable behavior was the behavior to duplicate from the smallest message size order. We think that the behavior of duplicating from

196

K. Henmi and A. Koyama

Fig. 3. Delivery ratio (Random Waypoint Model)

Fig. 4. Overhead (Random Waypoint Model)

the smallest message size order was judged to be the highest value because it allows to more duplication. Next, the results of the evaluation of the message delivery ratio and overhead when using the map are shown in Figs. 5 and 6. The results of the evaluation using map information also showed similar trends to those of the random waypoint model. The probability of encountering a node is inevitably reduced compared to the random waypoint model because it moves along the road. As a result, the message delivery ratio is also reduced compared to the random waypoint model. Even in this condition, in the proposed method to duplicate actively is judged to more valuable behavior. and as a result, the overhead is increased compared to MaxProp and PRoPHETV2. As similar to the random waypoint model, with the

DTN Routing Protocol Using Reinforcement Learning

197

Fig. 5. Delivery ratio (Tokyo)

duplication priority control judgment, it was judged that the actively duplicating behavior from the smallest message size order has the highest value.

Fig. 6. Overhead (Tokyo)

5 Conclusion In this paper, we proposed a DTN routing protocol based on reinforcement learning. The proposed method has two functions: to learn whether or not to duplicate the message to

198

K. Henmi and A. Koyama

encountered nodes, and to learn the order in which the messages should be duplicated when duplicating. We compared the proposed method with MaxProp and PRoPHETV2 and found an improvement in message delivery rate, however, the overhead was worse than that of the existing methods. In the future, we would like to improve the proposed method not only trained the message delivery, but also overhead or other parameters.

References 1. Delay Tolerant Networking Research Group. http://www.dtnrg.org/ 2. Vahdat, A., Becker, D.: Epidemic Routing for Partially Connected Ad Hoc Networks. Technical Report CS-2000-06 Duke University (2000) 3. Haas, Z.J., Small, T.: A new networking model for biological applications of ad hoc sensor networks. ACM/IEEE Trans. Netw. 14(1), 27–40 (2006) 4. Spyropoulos, T., Psounis, K., Raghavendram, C.S.: Spray and wait: an efficient routing scheme for intermittently connected mobile networks. In: Proceedings of ACM SIG-COMM Workshop on Delay-Tolerant Networking, pp. 77–90 (2005) 5. Lindgren, A., Doria, A., Schelen, O.: Probabilistic routing in intermittently connected networks. ACM SIGMOBILE Mob. Comput. Commun. Renew 7(3), 19–20 (2003) 6. Grasic, D.S., Lindgren, E.A., Doria, A.: The evolution of a DTN routing protocol PRoPHETv2. In: Proceedings of the 6th ACM Workshop on Challenged Networks, pp. 27–30 (2011) 7. Burgess, J., Jensen, B.D., Levine, B.N.: MaxProp: routing for vehicle-based disruptiontolerant networks. In: Proceedings of IEEE Infocom, pp. 1688–1698 (2006) 8. Dudukovich, R., Hylton., A.: A machine learning concept for DTN routing. In: Proceedings of IEEE International Conference on Wireless for Space and Extreme Environments (2017) 9. Littman, M., Boyan, J.: A distributed reinforcement learning scheme for network routing. In: Proceedings of the First International Workshop on Applications of Neural Networks to Telecommunications, pp. 45–51 (1993) 10. Sogabe, T.: Introduction to Reinforcement Learning Algorithms. Ohmsha, Tokyo (2019)

Data Fusion Protocols for Cloud Infrastructures Lidia Ogiela1(B) , Makoto Takizawa2 , and Urszula Ogiela1 1 Pedagogical University of Krakow, Podchor˛az˙ ych 2 Street, 30-084 Kraków, Poland

[email protected], [email protected] 2 Department of Advanced Sciences, Hosei University, 3-7-2, Kajino-Cho, Koganei-Shi,

Tokyo 184-8584, Japan [email protected]

Abstract. In this paper will be presented a new class of security protocols dedicated to Cloud infrastructures. Fusion technology in different types of data and management processes, will be described. Especially, will be presented an impact of description, interpretation and analysis processes as well as securing processes. Security processes, which are based on fusion methodology also include data sharing technologies. Data fusion protocols can be used in different infrastructures, so the impact of these processes will be analysed and presented. Keywords: Data fusion · Process of fusion security · Cloud infrastructures

1 Introduction Data registration processes obtained from various sources, now are becoming the main aspect of the development of systems for diversified data registration, processing and analysis [2]. Diversification of data sources opens up wide possibilities of analysing data obtained both from identical data sets as well as from various, different sources. Thus, the possibilities of processing data with various structures, forms and records are opened. Their importance begins to play an important role, which, in the context of the complete diversity of the analysed data sets, allows to indicate points that are similar and comparable in the process of in-depth data analysis. The realization of such possibilities is ensured by the process of fusion of data obtained during various registration processes. Data fusion means the process of combining data collected during completely different processes of their acquisition. Extensive data sets are registered independently, and then in the process of their connection, the characteristics of the processed data sets are defined, on the basis of which it will be possible to assess the significance of the analysed sets. In this way, the comparison and combination of data will take place according to their importance for the analysis process being carried out, and not on the basis of the type of data processed. The primary role in the data fusion process is played by the influence and importance of the processed information on the entire analysis process. Therefore, the most important stage of the data fusion process is the meaningful analysis of all obtained information, regardless of the method of its collection. This process is based on the implementation of cognitive algorithms for the meaning of the © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 199–203, 2021. https://doi.org/10.1007/978-3-030-61108-8_19

200

L. Ogiela et al.

analyzed sets and its aim is to extract individual semantic features characterizing only the analyzed data sets. The processes of semantic analysis have been described in [2, 4, 5, 8], where their main advantages both in the processes of automatic data understanding and in-depth interpretation of meaning were indicated.

2 Data Fusion Protocol Data fusion protocols must take into account both various sources of obtaining information, as well as various forms of data recording and the different influence of the obtained information on the course of the analysed occurrence. The data fusion protocol is therefore as follows: Protocol 1. Data fusion Step 1. Definition of the research issue/analysis Step 2. Indicating the sources of data acquisition Step 3. Determining the methods of data registration Step 4. Independent data recording from each indicated sources Step 5. Data pre-processing obtained from a given source in order to: – extract relevant information, and – select irrelevant data Step 6. Determining the characteristic features of the analyzed information sets Step 7. Indication of typical features for the analyzed data sets, clearly identifying the analyzed data Step 8. Assessment of the degree of influence of pre-processed data on the course of the analysis process Step 9. Data fusion all registrant sources Step 10. Definition of the set of all characteristic features identified in step 6 Step 11. Assessment of the importance of the information obtained semantic analysis of the complete data set. The data fusion protocol ensures the implementation of the process of combining information obtained both from various sources and recorded in various forms. The essence of this process is to obtain knowledge about the analyzed phenomenon stored in the semantic layers of the processed data sets [2, 4]. Extensive knowledge obtained from various sources gives the opportunity to understand the meaning of a given situation and its changes, which are determined of independent processes created this phenomenon.

Data Fusion Protocols for Cloud Infrastructures

201

3 Fusion Processes in Cloud Computing Data fusion has a variety of applications, especially in the case of the implementation of processes aimed at deep evaluation of the analysed phenomena, searching for the reasons for their occurrence, understanding the basis for the occurrence of the situations, assessing the significance and impact of certain phenomena on other situation. Data fusion can also play an important role in data protection processes by providing access to information through cryptographic threshold protocols that separate shadows of a secret can be dedicated to trustees who know the relevance of the set of characteristics accumulated in the fusion process [1, 3, 6, 10]. The essence of such an approach will therefore be the proper selection of participants in the data sharing protocol [1, 10], whose knowledge of the analyzed situation is varied and depends on the degree of expert analyzes. In this context, the separation of a part of the shared secret will be based on the degree of knowledge of the analyzed phenomenon, with the simultaneous allocation of the appropriate number of shadows taking into account this degree of knowledge. The knowledge obtained in the process of data fusion will therefore allow to generate a part of the secret, taking into account the characteristics of the shared information. Another example of the possible application of data fusion are information management support processes [7–9]. In this case, the knowledge obtained at the merger stage allows to significantly reduce the processes that have no influence (or are of residual importance) on the shape and course of decision-making processes, management and its improvement. The elimination of the stages of processing irrelevant data as well as the implementation of unnecessary stages of the data management process significantly improves and accelerates the implementation of the proper information management process. Of course, an information management process also includes the previously discussed data protection process, during which it is possible to protect important information against disclosure. It is also part of the data management process, where secret information is fully protected. In this case, both of the above-mentioned processes are carried out simultaneously, giving the possibility of both full data protection and optimal data management. Data protection and support processes are carried out at various levels of data processing and analysis. The most expected level for modern entities is the level of cloud computing. In this approach, the implementation of the data fusion process allows the possibility of simple collection of information at the level of a given entity, but also creates the possibility of transferring and managing large data collections based on their understanding and meaningful assessment. In the case of data fusion dedicated to cloud computing, a data fusion protocol is the following: Protocol 2. Data fusion in Cloud Step 1–11. Implementation of the protocol 1 (Data fusion) Step 12. Identification of the entity for the implementation of the protocol Step 13. Choice of implementation levels Step 14. Determining the relationship between the levels and the possibility of data transfer between them

202

L. Ogiela et al.

Step 15. Determining the degree of data protection during the transfer of information Step 16. Definition of a cryptographic protocol for data protection at each selected process level Step 17. Selection of protocol participants and allocation of rights to each of them Step 18. Data distribution between protocol participants based on their knowledge Step 19. Implementation of the information management process, taking into account the allocated parts. Implementation of the above protocol ensures efficient and safe management of data obtained from various sources and of different importance. The use of universal relationships in the field of data evaluation possible on the basis of the extraction of semantic layers determining the meaning of the recorded information, allows for a significant optimization of the processes of analysis, evaluation and data management whit fully them protection.

4 Conclusions In this paper was presented a new protocols based on fusion methodology, dedicated to Cloud infrastructures. Fusion technology in different types of data and management processes, will be described. Especially, will be presented an impact of description, interpretation and analysis processes as well as securing processes. Security processes which are based on fusion methodology also include data sharing technologies. Data fusion protocols can be used in different infrastructures, so the impact of these processes will be analysed and presented. Data fusion greatly enriches the data acquisition process. It is not limited to a single source of data acquisition, but allows to collect data recorded by various systems, units, entities, etc. It can be used both in the processes of simple information acquisition and analysis, as well as in complex data interpretation processes. The essence of the solution presented in this paper is the possibility of classifying data recorded by various recorders in order to determine their real significance and influence on the course of the analyzed phenomena using cognitive data evaluation algorithms. Acknowledgments. This work has been supported by the National Science Centre, Poland, under project number DEC-2016/23/B/HS4/00616. This work has been supported by Pedagogical University of Krakow research Grant No BN.610-28/PBU/2020.

References 1. Gregg, M., Schneier, B.: Security Practitioner and Cryptography Handbook and Study Guide Set. Wiley, Hoboken (2014) 2. Ogiela, L.: Semantic analysis in cognitive UBIAS & E-UBIAS systems. Comput. Math Appl. 63(2), 378–390 (2012) 3. Ogiela, L.: Cryptographic techniques of strategic data splitting and secure information management. Pervasive Mob. Comput. 29, 130–141 (2016)

Data Fusion Protocols for Cloud Infrastructures

203

4. Ogiela, L., Ogiela, M.R.: Bio-inspired cryptographic techniques in information management applications. In: IEEE 30th International Conference on Advanced Information Networking and Applications (IEEE AINA), Switzerland, 23–25 March 2016, pp. 1059–1063 (2016) 5. Ogiela, M.R., Ogiela, L.: Cognitive keys in personalized cryptography. In: 31st IEEE International Conference on Advanced Information Networking and Applications (IEEE AINA), Taiwan, 27–29 March 2017, pp. 1050–1054 (2017) 6. Ogiela, M.R., Ogiela, U.: Secure information splitting using grammar schemes. New Challenges Comput. Collective Intell. Stud. Comput. Intell. 244, 327–336 (2009) 7. Ogiela, M.R., Ogiela, U.: Secure information management in hierarchical structures. In: Kim, T., Adeli, H., Robles, R.J., Balitanas, M. (eds.) AST 2011. CCIS, vol. 195, pp. 31–35. Springer, Heidelberg (2011). 8. Ogiela, U., Takizawa, M., Ogiela, L.: Classification of cognitive service management systems in cloud computing. In: Barolli, L., Xhafa, F., Conesa, J. (eds.) BWCCA 2017. LNDECT, vol. 12, pp. 309–313. Springer, Cham (2018). https://doi.org/10.1007/978-3-319-69811-3_28 9. Wei, Y., Blake, M.B.: Service-oriented computing and cloud computing: challenges and opportunities. IEEE Internet Comput. 14(6), 72–75 (2010) 10. Yan, S.Y.: Computational Number Theory and Modern Cryptography. Wiley, Hoboken (2013)

Implementation of Process Migration Method for PC-FPGA Hybrid System Keisuke Takano1(B) , Tetsuya Oda2 , Ryo Ozaki2 , Akira Uejima2 , and Masaki Kohata1 1

Graduate School of Engineering, Okayama University of Science, 1-1 Ridaicho, Kita-ku, Okayama-shi, Okayama, Japan [email protected] 2 Faculty of Engineering, Okayama University of Science, 1-1 Ridaicho, Kita-ku, Okayama-shi, Okayama, Japan {oda,ozaki,uejima,kohata}@ice.ous.ac.jp

Abstract. This paper describes the distribution of processing and process migration in a PC-FPGA hybrid system. Here, the proposed hybrid system is constructed by a bus network using Ethernet, and communication between the nodes of the system is achieved by a proprietary protocol that is defined on the basis of Ethernet frames. For performance evaluation, we compare the processing time by distributed processing with that in the previous work. In addition, the execution process is migrated to FPGA. Experimental results show that the proposed system is capable of distributed processing using both PC and FPGA, and has the ability to migrate PC processing to FPGA.

1 Introduction Cluster-based systems are often used to implement high-performance systems, and there are implementation examples of systems such as [1–3] in FPGA-based systems. To build a system that can be used more flexibly by combining PC and FPGA, we are proposing a PC-FPGA hybrid system (PFH system) [4] in which PC and FPGA are equally connected. However, in recent years, FPGAs have been installed in data centers, which are one of the key elements in broadband services [5, 6]. In such systems, load balancing of services is an important problem. In our previous work, we showed an example of migration in a PFH system that implemented a communication path between FPGAs in a ring network [7]. In this study, we propose a PFH system that implements a communication path between FPGAs in Ethernet, and distributes processing. This study shows that the system can be operated by mounting s communication circuit on the FPGA board and executing the same processing as the previous system. The rest of the paper is organized as follows. Section 2 presents the PFH system. In Sect. 3, we show the description of the proposed method and FPGA transactions. Performance evaluations are presented in Sect. 4. Finally, conclusions and future work are presented in Sect. 5.

c The Editor(s) (if applicable) and The Author(s), under exclusive license  to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 204–210, 2021. https://doi.org/10.1007/978-3-030-61108-8_20

Implementation of Process Migration Method for PC-FPGA Hybrid System

205

Fig. 1. FPGA internal circuit configuration

2 Implementation of the PFH System 2.1 Basic Configuration Figure 1 shows the FPGA circuit configuration and PFH system. The PFH system is a mixed system of PC and FPGA. The PFH system allows PCs and FPGAs to equally participate in the network, and the system is designed to allow not only PCs but also FPGAs to be the source of communication. The PFH system includes embedded FPGA nodes (PC and FPGA connected via PCI Express) and an independent FPGA node (a single FPGA board that operates independently). Figure 1 shows that the PFH system consists of four PCs and FPGAs. Each FPGA is built into PC that is connected via PCI Express, and FPGAs are connected to each other through Ethernet. The FPGA circuits are connected by an internal bus and form a network that can communicate with any module. Table 1 lists each address space. The bus has an address space, and PC and APP module can access each module of FPGA from this address space. This functionality can be easily extended to any circuit that complies with bus standards. The application circuit can be mounted on the FPGA circuit of the proposed system. The application circuit corresponds to the application software on PC. The user can dynamically recon the application circuit using HWICAP.

206

K. Takano et al. Table 1. Memory map Internal Bus Address

2.2

Module (Unit)

Start Address

ICAP

0x00000000 0x00000FFF

End Address

Size 4 KB

Network Interface Controle port 0x00010000 0x00011FFF

8 KB

Network Interface Data port

0x00012000 0x00013FFF

8 KB

APP module

0x00600000 0x0060FFFF 64 KB

DRAM

0x40000000 0x7FFFFFFF

1 GB

FPGA Network

FPGA boards are installed in the PCI Express slot of each PC, and these FPGA boards constitute Ethernet. The SFP+ module of FPGAs is connected by an optical fiber cable and uses 10 Gigabit Ethernet PCS/PMA IP for physical layer control [8]. The data transfer rate of the physical layer is set to 10 Gbps. FPGAs that use an L2 switch are compatible with 10 GBASE-R and can build a network. An Ethernet II+ packet is defined by expanding the Ethernet II frame. Ethernet II+ packet is shown in Fig. 2. Ethernet II+ packets are used as communication units

Fig. 2. Ethernet II+ packet

in proposed system networks. The destination MAC address of the Ethernet header is used as the FPGA destination address, and the source MAC address is used as the source FPGA address. There are seven types of packets: register write, register read request, register read data, DMA address, DMA write data, DMA read request, and DMA read data. Control information, including these types of packets, is specified in a header consisting of five fields. The details are (2 octet) packet type, (6 octet) source address and (6 octet) destination address, (1 octet) source port and destination port, and (1 octet) frame size. The source and destination addresses specify the source and destination MAC addresses, respectively. The source and destination ports specify the source and destination within the node, which can be either PC or FPGA. The packet type field is identified by the packet type field. The user uses five types: register write (0 × 0101),

Implementation of Process Migration Method for PC-FPGA Hybrid System

207

register read request (0 × 0103), DMA address (0 × 110F), DMA write data (0 × 0111), and DMA read request (0 × 0121). Register read data (0 × 0102) and DMA read data (0 × 0112) are used in system-generated packets. The frame size field is used for DMA write data, DMA read requests, and DMA read data packets. The DMA write data and DMA read data packets indicate the size of the data stored in the packet. For example, 0 × 8 is specified in the frame size field when generating a packet in which 8 KB of data is stored. These header configurations allow flexible data transfer in mixed PC and FPGA systems. Table 2. API types Function

API

Data transfer

send data recv data

Register control register read register write FPGA control

partial reconfig batch start

2.3 Application Programming Interface (API) Implementation For distributed processing, we use Spark in host PC, but we implemented Python API for Spark in FPGA, because Spark does not support FPGA implementation. Table 2 shows the list of APIs that transfer data for distributed processing. send data transfers data in host PC memory to other PC memory or other FPGA DRAMs, recv data transfers data in other FPGA DRAM or other PC memory to host PC memory. register read and register write control register values for FPGAs locally and on the network. partial reconfig runs a dynamic partial reconfiguration and batch start activates the APP module. These APIs call modules in the PFH system for any processing. 2.4 APP Module and HWICAP To perform dynamic partial reconfiguration, an application circuit (binary file) for reconfiguration is transmitted to HWICAP. From PC, all modules of FPGA can be accessed from the character device driver address space. This binary file is stored on host PC and typically performs a dynamic partial reconfiguration. This driver enables control register read/write and DMA transfer execution. It is also possible to accept a request from the APP module with a character device driver interruption.

208

2.5

K. Takano et al.

Dynamic Partial Reconfiguration

Dynamic partial reconfiguration is a feature that allows to modify the circuitry running on FPGA. The APP module needs to change its functionality even when simultaneously running other circuits. Therefore, in a PFH system, can modify the circuit using dynamic partial reconfiguration. partial reconfig transfers the APP module partial bit file to HWICAP. It is possible to change the function of the APP module circuit without stopping PCIe I/F and network interface circuit.

3 Proposed Method The migration method of the process using the proposed system is explained. Prepare a bitfile of the application circuit that can perform the same processing as the PC process. Transfer the bitfile to FPGA at the timing when it is needed to migrate the PC processing and make FPGA executable. PC uses DMA to transfer the data to migrating FPGA. Then, FPGA processing is started using the API of batch start. The end of processing can be determined by checking the register of the APP module. The host process checks the register of the APP module of FPGA that delegated the process at regular intervals, and as soon as it can confirm the end of the process, it collects the data from the DRAM of FPGA to which the process is delegated. A previous study has shown that FPGA can continue processing even if the power of PC is turned off by separating the power of FPGA from PC. It is also possible to collect data after turning off the power supply of PC and waiting for a sufficient time to finish the process. Table 3. Resource utilization in the FPGA [%] Basic circuit APP module Total FF

15.13

5.72

20.85

LUT

27.32

15.18

42.50

Memory LUT

9.97

0.16

10.13

BRAM

25.51

21.68

47.19

DSP48

0.00

33.81

33.81

Table 4. Execution time for each distribution ratio. Distribution ratio of PC:FPGA

Previous work

Number of images 0 : 10 1 : 4 1 : 1 4 : 1 7 : 1 9 : 1 10 : 0 0 : 10 1 : 4 1 : 1 4 : 1 9 : 1 10 : 0 1000

17.0

14.1 8.6

3.5 3.3

3.2

3.5

19.7

16.6 12.6

4.0 3.2

3.7

2000

33.2

27.6 17.8

6.7 5.3

5.5

6.1

39.9

38.5 20.5

8.2 5.0

5.6

3000

50.1

40.0 25.9 11.2 6.7

6.9

8.5

59.2

46.2 29.9 12.2 6.8

7.7

4000

72.2

59.7 36.3 14.4 8.9

9.4

10.8

80.1

71.2 38.3 16.3 8.0

9.0

Implementation of Process Migration Method for PC-FPGA Hybrid System

209

4 Performance Evaluation To measure the communication and distributed processing performance, we measured the time required to complete the execution of an image compression application implemented using Spark and the FPGA control API. The CPU used in each PC was an Intel Core i5-9600, and Ubuntu 18.04 was used as the OS. The Xilinx KC705 evaluation board was adopted for FPGA. DDR3-1600 was used as DRAM on the FPGA board. In the PC network, each PC was connected via Ethernet (1000BASE-T), and the router (HUB) was WZR-900DHP (Buffalo Technology, Nagoya, Japan). In the FPGA network, each FPGA was connected via Ethernet (10 GBASE-R), and the L2 switch was QSW-804-4C (QNAP Systems, New Taipei City, Taiwan). The DRAM used on the FPGA board was DDR3-1600, and PCIe I/F was used on a Gen2 × 4 interface on XDMA (Xilinx, California, USA). For the experiment, we used a 320×240×24 bit image. High-level synthesis (HLS) is used in the JPEG encoder Vivado HLS to perform the same processing on PC and FPGA. To unify the processing between PC and FPGA, we employed a JPEG encoder written in C. The PFH system circuit was generated using Vivado 2019.2. Table 3 shows the resource utilization of the PFH System basic circuit and APP module in FPGA. The application binaries compiled by gcc on PC and the circuit with HLS in Vivado HLS on FPGA were executed. The method of distributed processing was based on reference [4]. Table 4 shows the experimental results of JPEG-encoder processing for each distributed ratio. The results confirm that the process can be distributed in the same way as in the example in [4]. In the FPGA system, that is the case of the distribution ratio is 0:10 in Fig. 2, the proposed system is faster than that in the previous work. The execution times of the proposed system were 10.8 s and 72.2 s for 4000 images at a PC-FPGA ratio of 10:0 and 0:10. The PC-FPGA performance ratio was 6.7:1. The execution time was shortest when the ratio of PC to FPGA was 7:1, which was close to the ratio of 6.7:1. This allows us to find the best distribution ratio from the time ratio as in the previous work. Next, we prepared an APP module that could perform Laplacian filtering of multiple images at a time for migration experiments. The resources used by this APP module were 1, 802 (0.44%) for FF, 7, 235 (3.55%) for LUT, 0 (0.0%) for Memory LUT, and 6 (1.35%) for BRAM. Initially, PC0 and FPGA0-3 were started and PC1-3 ware was turned off. of note, the power supply of each FPGA is independent from PC. The filtering software is run on PC0 to filter 1000 images. After filtering 600 images out of 1000 images with PC0, and after transferring the filter circuit data and the execution data to FPGA0-3, batch start was used to delegate the processing, and the power of PC0 was turned off. After a period of time, PC0 was turned on and the data were retrieved from FPGA0-3. The experiment showed that all images were correctly filtered. Overall, it was determined that the intended behavior of process migration was realized.

5 Conclusions In this study, we implemented a migration method in a system that combines PC and FPGA. The FPGA network of the proposed system was implemented with a configuration using Ethernet. For initial evaluation, we implemented and evaluated distributed

210

K. Takano et al.

processing in image compression algorithm. We demonstrated that the system can be operated by migrating the process from PC to FPGA using the proposed system. In the future, we will implement applications for efficient processing by migration in the proposed system.

References 1. Sano, K., Yamamoto, S.: FPGA-based scalable and power-efficient fluid simulation using floating-point DSP blocks. IEEE Trans. Parallel Distrib. Syst. 28(10), 2823–2837 (2017) 2. Andrew, P., et. al.: A reconfigurable fabric for accelerating large-scale datacenter services. In: Proceedings 41th Annual International Symposium on Computer Architecture (ISCA), pp. 13–24 (2014). https://doi.org/10.1145/2678373.2665678 3. Ghasemi, E., Chow, P.: Accelerating apache spark big data analysis with FPGAs. In: 2016 International IEEE Conferences on Ubiquitous Intelligence Computing, Advanced and Trusted Computing, Scalable Computing and Communications, Cloud and Big Data Computing, Internet of People, and Smart World Congress (UIC/ATC/ScalCom/CBDCom/IoP/ SmartWorld), pp. 737–744 (2016). https://doi.org/10.1109/UIC-ATC-ScalCom-CBDComIoP-SmartWorld.2016.0119 4. Takano, K., et al.: Implementation of distributed processing using a PC-FPGA hybrid system. In: Proceedings of FPT-2019, pp. 387–390 (2019) 5. Weerasinghe, J., et al.: Network-attached FPGAs for data center applications. In: 2016 International Conference on Field-Programmable Technology (FPT), pp. 36–43 (2016). https://doi. org/10.1109/FPT.2016.7929186 6. Huang, M., et. al.: Programming and runtime support to blaze FPGA accelerator deployment at datacenter scale. In: Proceedings the Seventh ACM Symposium on Cloud Computing (SoCC 2016), pp. 456–469 (2016). https://doi.org/10.1145/2987550.2987569 7. Takano, K., et al.: PC process migration using FPGAs in ring networks. IEICE Commun. Express 9(5), 141–145 (2020). https://doi.org/10.1587/comex.2019XBL0163 8. 10 Gigabit Ethernet PCS/PMA v6.0 LogiCORE IP Product Guide (PG068). https://japan. xilinx.com/support/documentation/ip documentation/ten gig eth pcs pma/v6 0/pg068-tengig-eth-pcs-pma.pdf. Accessed 5 Oct 2016

Speeding-Up of Construction Algorithms for the Graph Coloring Problem Kazuho Kanahara1(B) , Kengo Katayama1 , Takafumi Miyake1 , and Etsuji Tomita2 1

2

Graduate School of Engineering, Okayama University of Science, 1-1 Ridaicho, Kita-ku, Okayama-shi 700-0005, Japan [email protected], [email protected] The Advanced Algorithms Research Laboratory, The University of Electro-Communications, 1-5-1 Chofugaoka, Chofu, Tokyo 182-8585, Japan

Abstract. The graph coloring problem (GCP) is one of the most important combinatorial optimization problems that has many practical applications. DSATUR and RLF are well known as typical solution construction algorithms for GCP. It is necessary to update the vertex degree in the subgraph when selecting vertices to be colored in both DSATUR and RLF. There is an issue that the higher the edge density of a given graph, the longer the processing time. In this paper, we propose a subgraph degree updating method to improve this issue. Experimental results show that the proposed method is faster than the conventional method and LazyRLF, especially for graphs with higher edge density.

1 Introduction The construction algorithm for the combinatorial optimization problem is an algorithm for constructing a solution that satisfies constraints on the problem. The graph coloring problem (GCP) is one of the classical combinatorial optimization problems that has important practical applications such as scheduling [8], register allocation [4], frequency assignment [7], telecommunications [2], and spectrum resources assignment in LTE Networks [6]. DSATUR [3]1 and RLF [8] are well known as construction algorithms for GCP. DSATUR is an algorithm that colors vertex in descending order of degree of saturation. RLF is an algorithm that colors recursive searched independent uncolored vertex set. Much research has been conducted on algorithms that improve the accuracy and processing time [1, 5]. It is necessary to update the vertex degree in the subgraph when selecting vertices to be colored in both DSATUR and RLF. The updating method is a common process in algorithms for graph problems. There is an issue that the higher the edge density of a given graph, the longer the processing time. LazyRLF [5] was proposed by Chiarandini et al. LazyRLF is an improved algorithm of RLF that speeds up the method by avoiding unnecessary vertex degree computations. 1

In this paper, we consider a DSATUR based heuristic algorithm rather than a DSATUR based branch and bound algorithm shown in [6].

c The Editor(s) (if applicable) and The Author(s), under exclusive license  to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 211–222, 2021. https://doi.org/10.1007/978-3-030-61108-8_21

212

K. Kanahara et al.

In this paper, we propose a subgraph degree updating method to improve this issue. The proposed method can be applied to general graph problems that do not depend on specific algorithms and graph peculiarity. It is evaluated on DIMACS graphs and random graphs. Experimental results show that the proposed method is faster than the conventional methods and LazyRLF, especially for graphs with higher edge density.

2 Graph Coloring Problem Let G = (V, E) be an arbitrary undirected graph where V is the set of n vertices and E ⊆ V × V is the set of edges in G. A coloring of G is an assignment of colors to the vertices V such that no two adjacent vertices share the same color. A stable set is a set of pairwise non-adjacent vertices. Hence, a vertex coloring of G is a partition of its vertex set into k stable sets called the colored vertex set (s = (C1 ,C2 , . . . ,Ck )). The objective of the vertex coloring problem (VCP) is to find a solution s that minimizes the number k of colored vertex set. The main types of graph coloring problem are vertex coloring, edge coloring and surface coloring. The vertex coloring problem is generally called graph coloring problem, since other coloring problems can be transformed into a vertex coloring instance. The GCP includes important applications in different domains: scheduling [8], register allocation [4], frequency assignment [7], telecommunications [2], spectrum assignment [6]. We next give several notations used in this paper.

C

: the colored vertex set. C ⊆ V

Ck

: the proper subset of colored vertex set that is colored with color k. Ck ⊆ C,C1 ∪C2 . . .Ck−1 ∪Ck = C

U

: the uncolored vertex set. U ⊆ V,C ∪U = V,C ∩ U = 0/

U

: the proper subset of the uncolored vertex set U. U  ⊆ U

Uk

: the proper subset of the uncolored vertex set U that can be colored with color k. U k ⊆ U

 Uk

: the proper subset of the uncolored vertex set U k . U k ⊆ U k

Unk



: the proper subset of the uncolored vertex set U that can not be colored with color k. U k ∪Unk = U,U k ∩Unk = 0/

c(v)

: the color number for a vertex v

Γ (v)

: the set of adjacent vertices to vertex v

Λ (v)

: the set of not adjacent vertices to vertex v. v ∈ / Λ (v), Γ (v) ∪ Λ (v) ∪ {v} = V

degG(S) (v)

: the degree of a vertex v in the subgraph G(S) of the graph G(V )

dsaturG(C) (v) : the number of different colors adjacent to a vertex v in a colored subgraph G(C), that is called the degree of saturation.

Speeding-Up of Construction Algorithms for the Graph Coloring Problem

213

3 DSATUR DSATUR [3] is a construction algorithm that colors vertex in descending order of degree of saturation. The pseudo code of DSATUR is shown in Fig. 1. At first, we initialize colored set C, uncolored set U and the degree for a subgraph G(U) (Line 1). We select one vertex v with the max degree degG(U) (v) for the uncolored subgraph G(U). The selected vertex v is colored with the smallest color 1. We update the degree for the uncolored subgraph G(U) and the degree of saturation for the colored subgraph G(C) (Lines 2–4). Next, a vertex v is found with the max degree of saturation dsaturG(C) (v) for the colored subgraph G(C). If we find a subset U  of multiple vertices with the same max degree of saturation for the colored subgraph G(C), we select one vertex v ∈ U  randomly with the max degree degG(U) (v) for the uncolored subgraph G(U) (Lines 6– 8). We find the least possible color k that can color the selected vertex v. The selected vertex v is colored with the color k. Then, update the degree for the uncolored subgraph G(C) and the degree of saturation for the colored subgraph G(C) (Lines 9–10). This process is repeated until the uncolored vertex set U is empty (Lines 5–12).

Fig. 1. The pseudo code of DSATUR

4 Recursive Largest First (RLF) RLF [8] is a construction algorithm that colors recursive searched independent uncolored vertex set. The pseudo code of RLF is shown in Fig. 2. At first, we initialize color k, colored set C, uncolored set U and the degree for a subgraph G(U) (Line 1). We select one vertex v with the max degree degG(U k ) (v) for the uncolored subgraph G(U k ) that can be colored with color k. The selected vertex v is colored with the color k. We drop the set of adjacent vertices Γ (v) for vertex v from the vertex set U k . Then, update the degree for the colored subgraph G(U) and the degree for the colored subgraph G(U k ) (Lines 4–6). Next, a vertex v is found with

214

K. Kanahara et al.

the max difference between the degree degG(U) (v) for the colored subgraph G(U) and the degree degG(U k ) (v) for the colored subgraph G(U k ) (the degree degG(Unk ) for the 

uncolored subgraph G(Unk )). If we find a subset U k of multiple vertices with the same  max difference of degrees, we select one vertex v ∈ U k randomly with the min degree degG(U k ) (v) for the uncolored subgraph G(U k ). The selected vertex v is colored with the color k (Lines 9–11). Then, update the degree for the uncolored subgraph G(U) and the degree of saturation for the colored subgraph G(U k ). If the set of U k is empty, color it with color k + 1 (Lines 7–14). This process is repeated until the uncolored vertex set U is empty (Lines 2–15).

Fig. 2. The pseudo code of recursive largest first

5 Novel Updating Subgraph Degree Method In this section, we first describe standard subgraph degree updating methods performed by DSATUR and RLF. Next, we describe novel subgraph degree updating method. 5.1

Standard Subgraph Degree Updating Methods

It is necessary to updating the vertex degree in the subgraph when selecting vertices to be colored in both DSATUR and RLF. There are two standard degree updating methods A and B with different sets of loop targets (Fig. 3, Fig. 4). These methods decrement it by one the degree of vertex v, that is adjacent to vd and contains a subset S every time vertex vd of a subset S is dropped.

Speeding-Up of Construction Algorithms for the Graph Coloring Problem

215

Fig. 3. The pseudo code of standard subgraph degree updating method A

Fig. 4. The pseudo code of standard subgraph degree updating method B

5.2 Proposed Updating Subgraph Degree Method The standard degree updating method described above has an issue that the higher the edge density of a given graph, the longer the processing time. We propose a subgraph degree updating method to improve this issue. The standard degree updating method updates the degree of all vertices every time a single vertex is dropped from the subset S. It is inefficient because it even the degrees of the multiple vertices to be dropped every vertex dropping. The proposed method updates the degree degG(S) (v) of vertex v in the subset S after dropping multiple vertices D from the subset S. In the main loop of the proposed method, we adaptively switch between three degree updating processes (Lines 2, 6, 10). As the three processes, we introduce the standard degree updating methods A, B and a new degree updating method. The new degree updating method performs a degree update process by a set Λ (v) that is not adjacent to vertex v. In the new method , we update the degree by incrementing by one the degree of vertex v ∈ S ∩ Λ (vd ) that are not adjacent to the deleted vertex vd in the subset S. We need to keep the difference bias from the actual degree because the degree of all the vertices in the subset S is one greater than the actual degree when degree updating method is performed (Line 11). The actual degree of a vertex vi can be calculated by degG(S) (vi ) − bias. The relationship between the size of the actual degrees of two vertices vi and vi can be represented by degG(S) (vi ) < degG(S) (v j ) instead of degG(S) (vi ) − bias < degG(S) (v j ) − bias (Fig. 5).

216

K. Kanahara et al.

Fig. 5. The pseudo code of proposed subgraph degree updating method

6 Experimental Results To evaluate the effectiveness of the proposed method, we performed computational experiments on the 119 instances of DIMACS benchmark graphs and the 3230 instances of the random graphs. The random graphs are identified by the following vertex sizes (17 sizes): n = 100, 200, 300, 400, 500, 600, 700, 800, 900, 1000, 2000, 3000, 4000, 5000, 10000, 15000, 20000 and density ρ (19 types) from 0.05 to 0.95 separated by 0.05. We generated 3230 types random graphs that are composed of 10 different random graphs generated based on each combination of 17 types of each size and 19 types of each density. The codes of conventional DSATUR2 and RLF methods and LazyRLF3 were used. The proposed algorithm was implemented in C++. All experiments were performed on a machine with a 3.6GHz Intel Core i7 CPU and 22.4GiB RAM under Ubuntu 18.04, using the g++ compiler 8.3.0 with ‘−O3’ option. To execute the DIMACS Machine Benchmark, this machine required 0.13 (s) CPU seconds for r300.5, 0.79 (s) for r400.5 and 3.04 (s) for r500.5. In the experiments, the performance of DSATUR and RLF with proposed method are compared with that of the conventional DSATUR and conventional RLF and LazyRLF. Table 1 shows the results of running time of DSATUR and RLF with proposed method and the conventional DSATUR and RLF methods, LazyRLF for the DIMACS graphs. Each algorithm was run independently for 100 times on each DIMACS graph. In the table, we show, the instance names, vertex size |V | = n, graph density ρ and for DSATUR and RLF with proposed method, the conventional DSATUR and RLF with methods, LazyRLF, the average running time “Time(ms)” in milliseconds. The bold-faced data indicate a better result in each of the columns “Time(ms)” 2 3

https://imada.sdu.dk/∼marco/gcp-study/. https://imada.sdu.dk/∼marco/gcp/rlf/.

Speeding-Up of Construction Algorithms for the Graph Coloring Problem

217

Table 1. Results of DSATUR and RLF with proposed method and the conventional DSATUR and conventional RLF and LazyRLF on the DIMACS Graphs instance

DSATUR

RLF

Proposed Conventional Proposed Conventional LazyRLF Name

n

ρ

Time (ms) Time (ms)

Time (ms)

DSJC125.1

125 0.10 0.067

Time (ms) Time (ms) 0.141

0.073

0.091

0.169

DSJC125.5

125 0.50 0.150

0.349

0.379

0.269

0.498

DSJC125.9

125 0.90 0.129

0.413

0.158

0.412

0.690

DSJC250.1

250 0.10 0.223

0.531

0.286

0.389

0.550

DSJC250.5

250 0.50 0.415

1.084

1.342

1.349

2.345

DSJC250.9

250 0.90 0.337

1.404

0.561

2.140

2.549

DSJC500.1

500 0.10 0.610

1.392

1.080

1.355

2.397

DSJC500.5

500 0.50 1.601

3.649

8.083

5.140

10.333

DSJC500.9

500 0.90 1.407

6.572

3.118

10.669

8.491

DSJR500.1

500 0.03 0.296

1.027

0.525

1.090

1.510

DSJR500.1c

500 0.97 1.379

7.134

0.696

8.128

7.526

DSJR500.5

500 0.47 1.480

3.693

5.973

6.016

4.030

DSJC1000.1

1000 0.10 2.258

4.437

5.186

5.258

16.373

DSJC1000.5

1000 0.50 7.413

18.902

53.353

21.362

59.986

DSJC1000.9

1000 0.90 9.827

30.470

18.995

62.076

63.387

fpsol2.i.1

496 0.10 0.588

1.371

0.689

1.381

1.840

fpsol2.i.2

451 0.09 0.358

0.910

0.585

0.903

1.317

fpsol2.i.3

425 0.10 0.343

0.921

0.563

0.861

0.864

inithx.i.1

864 0.05 1.143

2.817

1.480

2.426

2.710

inithx.i.2

645 0.07 0.646

1.781

0.965

1.809

2.156

inithx.i.3

621 0.07 0.617

1.698

0.916

1.490

1.620

le450 15a

450 0.08 0.414

0.985

0.730

0.952

1.441

le450 15b

450 0.08 0.413

0.954

0.736

1.139

1.315

le450 15c

450 0.17 0.604

1.711

1.580

1.703

2.420

le450 15d

450 0.17 0.612

1.656

1.679

1.591

1.997

le450 25a

450 0.08 0.412

1.076

0.810

0.926

1.137

le450 25b

450 0.08 0.413

0.901

0.881

1.370

1.386

le450 25c

450 0.17 0.632

1.459

1.815

1.600

2.586

le450 25d

450 0.17 0.662

1.497

1.776

1.721

2.089

le450 5a

450 0.06 0.431

0.979

0.714

0.943

1.754

le450 5b

450 0.06 0.462

1.114

0.434

0.996

1.460

le450 5c

450 0.10 0.560

1.143

0.524

1.080

1.687

le450 5d

450 0.10 0.505

1.206

0.541

1.146

1.654

mulsol.i.1

197 0.20 0.121

0.379

0.186

0.394

0.427

mulsol.i.2

188 0.22 0.128

0.381

0.213

0.291

0.461

mulsol.i.3

184 0.23 0.125

0.365

0.221

0.316

0.407

mulsol.i.4

185 0.23 0.147

0.350

0.214

0.304

0.365

mulsol.i.5

186 0.23 0.138

0.352

0.214

0.316

0.392 (continued)

218

K. Kanahara et al. Table 1. (continued)

instance

DSATUR

RLF

Proposed Conventional Proposed Conventional LazyRLF Name

n

ρ

Time (ms) Time (ms)

Time (ms)

zeroin.i.1

211 0.19 0.175

Time (ms) Time (ms) 0.407

0.184

0.339

0.511

zeroin.i.2

211 0.16 0.170

0.364

0.200

0.341

0.422

zeroin.i.3

206 0.17 0.149

0.382

0.198

0.322

0.418

miles1000

128 0.40 0.137

0.299

0.191

0.236

0.254

miles1500

128 0.64 0.105

0.364

0.160

0.478

0.453

miles250

128 0.05 0.051

0.153

0.064

0.098

0.132

miles500

128 0.14 0.063

0.195

0.105

0.159

0.151

miles750

128 0.26 0.109

0.258

0.147

0.216

0.215

queen5 5

25 0.53 0.013

0.029

0.012

0.032

0.035

queen6 6

36 0.46 0.019

0.051

0.023

0.047

0.048

queen7 7

49 0.41 0.028

0.073

0.037

0.056

0.071

queen8 12

96 0.30 0.093

0.162

0.098

0.069

0.145

queen8 8

64 0.36 0.054

0.108

0.051

0.067

0.078

queen9 9

81 0.33 0.065

0.149

0.080

0.081

0.115

queen10 10

100 0.30 0.150

0.143

0.126

0.129

0.210

queen11 11

121 0.27 0.080

0.202

0.176

0.157

0.270

queen12 12

144 0.25 0.102

0.292

0.215

0.245

0.365

queen13 13

169 0.23 0.146

0.355

0.252

0.328

0.542

queen14 14

196 0.22 0.157

0.392

0.373

0.423

0.987

queen15 15

225 0.21 0.206

0.479

0.432

0.480

1.107

queen16 16

256 0.19 0.257

0.520

0.537

0.549

1.292

anna

138 0.05 0.064

0.146

0.059

0.099

0.141

david

87 0.11 0.030

0.096

0.034

0.055

0.075

homer

561 0.01 0.476

1.178

0.534

0.934

1.307

huck

74 0.11 0.023

0.075

0.034

0.058

0.066

jean

80 0.08 0.030

0.073

0.029

0.059

0.059

games120

120 0.09 0.049

0.146

0.071

0.110

0.141

latin square 10

900 0.76 5.489

18.994

18.669

21.545

43.753

school1

385 0.26 0.591

1.428

1.334

1.593

2.308

school1 nsh

352 0.24 0.490

1.141

1.066

1.279

1.748

myciel3

11 0.36 0.004

0.016

0.004

0.019

0.026

myciel4

23 0.28 0.007

0.023

0.008

0.022

0.031

myciel5

47 0.22 0.017

0.060

0.020

0.035

0.063

myciel6

95 0.17 0.047

0.138

0.052

0.064

0.119

myciel7

191 0.13 0.147

0.362

0.145

0.224

0.424

mugg88 1

88 0.04 0.034

0.092

0.037

0.051

0.105

mugg88 25

88 0.04 0.033

0.076

0.040

0.051

0.090

mugg100 1

100 0.03 0.060

0.095

0.045

0.051

0.071

mugg100 25

100 0.03 0.064

0.088

0.044

0.063

0.118 (continued)

Speeding-Up of Construction Algorithms for the Graph Coloring Problem Table 1. (continued) instance Name

n

ρ

DSATUR RLF Proposed Conventional Proposed Conventional LazyRLF Time (ms) Time (ms) Time (ms) Time (ms) Time (ms)

abb313GPIA ash331GPIA ash608GPIA ash958GPIA will199GPIA 1-Insertions 1-Insertions 1-Insertions 2-Insertions 2-Insertions 2-Insertions 3-Insertions 3-Insertions 3-Insertions 4-Insertions 4-Insertions 1-FullIns 3 1-FullIns 4 1-FullIns 5 2-FullIns 3 2-FullIns 4 2-FullIns 5 3-FullIns 3 3-FullIns 4 3-FullIns 5 4-FullIns 3 4-FullIns 4 4-FullIns 5 5-FullIns 3 5-FullIns 4 wap01a wap02a wap03a wap04a wap05a wap06a wap07a wap08a qg.order30 qg.order40 qg.order60 qg.order100

1557 662 1216 1916 701 67 202 607 37 149 597 56 281 1406 79 475 30 93 282 52 212 852 80 405 2030 114 690 4146 154 1085 2368 2464 4730 5231 905 947 1809 1870 900 1600 3600 10000

0.04 0.02 0.01 0.01 0.03 0.11 0.06 0.03 0.11 0.05 0.02 0.07 0.03 0.01 0.05 0.02 0.23 0.14 0.08 0.15 0.07 0.03 0.11 0.04 0.02 0.08 0.03 0.01 0.07 0.02 0.04 0.04 0.03 0.02 0.11 0.10 0.06 0.06 0.07 0.05 0.03 0.02

3.337 0.569 1.867 4.099 0.694 0.029 0.186 0.629 0.011 0.066 0.639 0.018 0.218 2.568 0.029 0.472 0.009 0.055 0.241 0.021 0.129 1.200 0.028 0.398 6.229 0.047 0.781 25.549 0.078 1.681 7.558 7.652 32.682 36.951 1.733 1.815 5.547 5.953 1.489 4.296 21.680 155.271

4 5 6 3 4 5 3 4 5 3 4

7.437 1.644 5.244 10.365 1.840 0.071 0.298 1.448 0.033 0.163 1.416 0.047 0.425 5.415 0.068 0.948 0.025 0.104 0.494 0.051 0.304 2.595 0.088 0.756 9.943 0.116 1.738 40.136 0.208 3.118 16.905 18.270 59.426 70.438 3.733 4.281 12.707 12.897 2.972 7.414 36.032 243.277

4.762 0.820 2.222 5.032 0.895 0.025 0.098 0.589 0.012 0.062 0.624 0.016 0.151 2.331 0.020 0.353 0.010 0.035 0.219 0.019 0.122 0.972 0.028 0.300 5.007 0.075 0.671 17.849 0.067 1.735 18.478 18.509 56.287 60.486 5.895 5.924 15.700 16.036 3.447 10.268 46.928 405.712

6.173 1.270 3.795 7.630 1.348 0.047 0.210 1.049 0.029 0.109 1.125 0.025 0.290 4.413 0.038 0.557 0.022 0.060 0.273 0.038 0.182 1.858 0.056 0.521 7.847 0.098 1.409 23.742 0.112 2.733 15.928 16.992 52.191 60.487 4.684 4.000 12.995 12.451 2.766 7.544 31.105 233.041

18.186 2.548 8.302 19.928 2.313 0.068 0.378 1.949 0.040 0.189 2.418 0.055 0.812 10.784 0.073 1.442 0.033 0.113 0.596 0.060 0.353 3.982 0.076 1.328 18.426 0.173 3.361 54.621 0.254 6.066 25.653 28.641 112.630 128.870 5.337 5.189 17.312 18.895 10.860 38.654 254.257 4896.978

219

220

K. Kanahara et al.

Fig. 6. Comparison of execution times of the proposed RLF, conventional RLF and Lazy RLF for vertex size 100 random graphs

Fig. 7. Comparison of execution times of the proposed RLF, conventional RLF and Lazy RLF for vertex size 500 random graphs

in both algorithms for each graph4 . The computational results in Table 1 showed that the proposed method is effective in that the performance of DSATUR and RLF can be improved for most of the DIMACS graphs. For example the proposed method with RLF is about 3.26 times faster than the conventional RLF in graph of DSJC1000.9, the graph is of large vertex size and higher edge density (0.9) than the other graphs. Figures 6, 7, 8, 9, show the results of RLF with proposed method, conventional RLF and LazyRLF for the random graphs. Each algorithm was run independently for 10 times on each random graph. The proposed method with RLF is faster than the conventional method for dense random graphs with 20,000 vertex size and edge densities greater than 0.8. The time complexity of the degree update is O(|ΓG(V ) |) for the conventional RLF, O(|U k | · |ΓG(U k ) |) for the Lazy RLF, and O(min(|U k |, |ΓG(V ) |, |ΛG(V ) |)) for the proposed RLF. However, the conventional RLF is faster than the proposed RLF in some case, the reason reported by Chiarandini et al. is a higher cache hit ratio of the conventional RLF. These results show that the proposed method is more effective than the conventional method for high-density graphs. In the 3349 graph examples used in the experiments. We have shown that the proposed method has comparable or better performance than 3337 instances (99%) in the conventional DSATUR, 1358 instances (41%) in the conventional RLF, 3309 instances (99%) in Lazy RLF. The computational results showed that the proposed method is a fast degree updating method for high dense graphs compared to the conventional method (Figs. 10, 11, 12, 13).

4

The implementation of the random vertex selection process in the vertex tie-break selection is different for each solution method. Although the accuracy of the solutions is slightly different, the accuracy is comparable with the average.

Speeding-Up of Construction Algorithms for the Graph Coloring Problem

221

Fig. 8. Comparison of execution times of the proposed RLF, conventional RLF and Lazy RLF for vertex size 1000 random graphs

Fig. 9. Comparison of execution times of the proposed RLF, conventional RLF and Lazy RLF for vertex size 2000 random graphs

Fig. 10. Comparison of execution times of the proposed RLF, conventional RLF and Lazy RLF for vertex size 5000 random graphs

Fig. 11. Comparison of execution times of the proposed RLF, conventional RLF and Lazy RLF for vertex size 10000 random graphs

Fig. 12. Comparison of execution times of the proposed RLF, conventional RLF and Lazy RLF for vertex size 15000 random graphs

Fig. 13. Comparison of execution times of the proposed RLF, conventional RLF and Lazy RLF for vertex size 20000 random graphs

222

K. Kanahara et al.

7 Conclusion We proposed a subgraph degree updating method to improve the issue that the higher the edge density of a given graph, the longer the processing time. The computational results showed that the proposed method is effective in that the running times of DSATUR and RLF can be improved for highly dense graphs particularly. It was shown to be more effective than LazyRLF for high dense graphs. Future work is to apply the proposed method to graph problems such as the maximum clique problem. Acknowledgement. This work was supported by JSPS KAKENHI Grant Numbers JP19K12166, JP17K00006.

References 1. Adegbindin, M., Hertz, A., Bella¨ıche, M.: A new efficient RLF-like algorithm for the vertex coloring problem. Yugoslav J. Oper. Res. 26(4), 441–456 (2016) 2. Balabhaskar, B., Sergiy, B.: Graph domination, coloring and cliques in telecommunications. In: Handbook of Optimization in Telecommunications, pp. 865–890. Springer (2006) 3. Br´elaz, D.: New methods to color the vertices of a graph. Commun. ACM 22, 251–256 (1979) 4. Chaitin,G.J.: Register allocation and spilling via graph coloring. In: Proceedings of the 1982 SIGPLAN Symposium on Compiler Construction, SIGPLAN 1982, pp. 98–105 (1982) 5. Chiarandini, M., Galbiati, G., Gualandi, S.: Efficiency issues in the RLF heuristic for graph coloring. In: Proceedings of the 9th Metaheuristics International Conference, MIC 2011, pp. 461–469 (2011) 6. Dimitris, T., Eirini, L., Nikos, P., Lazaros, M.: A graph-coloring secondary resource allocation for D2Dcommunications in LTE networks. In: 2012 IEEE 17th International Workshop on Computer Aided Modeling and Design of Communication Links and Networks (CAMAD), pp. 56–60. IEEE (2012) 7. Hale, W.K.: Frequency assignment: theory and applications. Proc. IEEE 68(12), 1497–1514 (1980) 8. Leighton, F.T.: A graph coloring algorithm for large scheduling problems. J. Res. Nat. Bur. Stand. 84(6), 489–506 (1979)

An On-Board Equipment and Blockchain-Based Automobile Insurance and Maintenance Platform Wen-Yao Lin(B) , Frank Yeong-Sung Lin, Ting-Huan Wu, and Kuang-Yen Tai Department of Information Management, National Taiwan University, Taipei, Taiwan [email protected]

Abstract. This study provides a forward-looking usage-based automobile blockchain platform on the Internet of Vehicles network, describing the possible stakeholders, services, and their interaction modes. Most existing UBI (UsageBased Insurance) products use the drivers’ driving distance, driving time, or driving area as premium calculating standard. If driving data are recorded on the blockchain to take usage-based automobile insurance, it could bring more additional services such as automatic claiming, improvement of traffic adjudication’s quality, and preventive maintenance. The feasibility of this platform will discuss the configuration of the blockchain and related management or privacy issues to explain how this research can improve the existing limitation of usage-based automobile insurance. Keywords: Blockchain · Internet of vehicles · UBI (Usage-Based Insurance) · Usage-based automobile insurance

1 Introduction With the development of information technology, the nature and procedure of financial products are changing. Take insurance for example: in the past, when a customer wanted to submit a claim, costumers needed to sign a contract for his/her solicitor. Then, the solicitor would ask costumers to fill out the claim form and attach some hard copy certificates related to the insured peril. However, we can observe that insurers have made a great change in recent years. Now, they take advantage of the Internet and required information technology, and some insurers can automatically pay compensation to their clients in a short time after the insured peril happened. What is more, the premium calculating basis has changed because it was easier to collect the insurance data. For instance, researchers adopt automobile insurance service [1] with on-board equipment to collect drivers’ driving data to judge the driver’s risk level. This can calculate the premium through actual driving sensed data—that is, PHYD (Pay How You Drive). PHYD can calculate higher precise premiums than the traditional way. Most insurance products only adopt records to handle some business logics. A shared repository is maintained by peers to provide trust between parties, immutability, decentralization, and automation [2]. If driving data are recorded on the blockchain © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 223–232, 2021. https://doi.org/10.1007/978-3-030-61108-8_22

224

W.-Y. Lin et al.

to take usage-based automobile insurance, it could bring more additional service such as automatic claiming, improvement of traffic adjudication’s quality, and preventive maintenance. When a usage-based automobile insurance can operate on the blockchain system and smart contract, it is a forward-looking insurance product to break the existing operation of the automobile insurance ecosystem. The research scope of this work is as follows: (1) Research related works about blockchain application of insurance or finance service and their architecture. (2) Design an insurance product and its maintenance platform of usage-based automobile insurance based on blockchain and on-board equipment. (3) Design the driving data synchronization method to assure transmission successfully and efficiently. (4) Analyze its architecture and blockchain in detail. (5) Discuss the management issues of the architecture. (6) Explore the additional service if use this platform.

2 Related Works A. UBI Products First, insurance company announced that its automobile insurance product will calculate the premium base on drivers’ driving behavior. Its usage-based automobile insurance will collect driving data through on-board equipment, and it will analyze the driving data to give an insured bonus or penalty on the platform [1]. It collects specific driving data, e.g., average speed, times of breaking and total driving time. Different to the architecture this work brings out, it stores the driving data centrally; that is, no one knows what data the company will collect, and it limits the availability of driving data. Also, the data collected is too insufficient to bring more values to cover the loss of sacrificing by giving detailed driving data to insurance company. Second, compared to the UBI product of the Taian insurance company, the UBI product of Cathay Century is easier to use. The driver does not need to install on-board equipment to collect driving data; he/she only need to install an application on his/her mobile device. However, despite its convenience, the data it collects is not precise as that of the Taian insurance company. Therefore, besides to the limitation of the Taian insurance company’s UBI product, this insurance product has lacked attractiveness. It is critical for insurance companies to use a more precise data-collection method when they want to release a usage-based automobile insurance product [2]. B. Patent of Synchronization of Vehicle Sensor Information It is important to synchronize the driving data from on-board equipment to the blockchain stably. U.S. patent no. 9,959,764 [3] introduced the data synchronizing architecture in detail, as shown in Fig. 1. In this architecture, the on-board equipment will collect all the sensors on the car by data collection unit and initially analyze the driving data via a data analysis unit of the on-board system. Then, the on-board system will upload the initially analyzed data to the interface of insurance company’s system through unspecific

An On-Board Equipment and Blockchain-Based Automobile Insurance

225

communication technology [4]. After the data collection unit receives the data, it precedes many analyses to identify the vehicle, determine the risk, generate the risk judgement model, and determine the habits of the driver. However, this study aims to record all the data to the blockchain. Hence, this architecture needs to be modified appropriately to meet the requirements of the blockchain architecture.

Fig. 1. Architecture of U.S. Patent no. 9.959.764

C. Blockchain Design (1) Delegated proof of stake (dPoS) [5] The dPoS consensus mechanism can be easily considered as that the coin holders needs vote for delegates, who are responsible for validating transactions and maintaining the blockchain. The fault tolerance is evolved from practical Byzantine Fault Tolerance, pBFT. Although its concept has improved many issues of PoW and PoS [6], it still faces some problems, e.g., enough decentralization can never be achieved practically because a network cannot have an excessive amount validator or else it risks slowing down. (2) Tendermint [7] Different from dPoS, Tendermint is asynchronous. Its objective is to improve the efficiency of the consensus mechanism of BitCoin. By its asynchronous feature, its fault tolerance can be better than dPoS. For example, when facing DDOS attack, dPoS will result in many forks, but Tendermint will not. It will stop operation to ensure consistency [8]. It is based on a Proof of Authority consensus mechanism to record the origin of seafood as a supply chain. Although this case is like the platform this essay issued, there are some reasons that PoA is not the best choice for this platform. First, because this platform is for business use, the cost of operation is the primary consideration. Furthermore, due to the driving record conditions of this platform, it will record to the blockchain periodically or by critical events. To make it efficient, recording on different sub-blockchain may be reasonable [9]. Therefore, PoW is both unsuitable and uneconomic. Secondly, because this platform must be as decentralized as possible to ensure

226

W.-Y. Lin et al.

fair and mutual trust, neither PoA nor pure PoS will be the first choice. The other two choices, dPoS and Tendermint, are both suitable. Taking the mean time of block generation into consideration, Tendermint is a better choice. However, the scalability of the Tendermint is better [10]. To sum up, although every consensus has its own pros, cons, and ideal application scenario, Tendermint may be the most suitable choice for this platform based on the existing demands and background information. A comparison of the models is shown in Table 1. Table 1. Comparison of Pow, Pos, dPos, and Tendermint Items

PoW

PoA

Pure-PoS

dPos (EOS)

Tender mint

Consensus basis

Computing power

Authority nodes

Stakes

Stakes

Stakes

Mean time of block generation

20 s–10 min

By assignment 1–10 min

3–40 s

1–3 s

Lower

Lower

Cost

Higher

Fair

Advantages

Widely used

Configuration Use fewer is very flexible computing power

Higher

Continuous Faster block service when generation being attacked time, higher consistency, greater horizonal extension ability

Disadvantages

Waste computing power, lower efficiency

It’s not so suitable for union and public blockchain

Will generate more forks when being attacked

May be more centralized during specific situations

Service will stop when being attacked

3 Model and Blockchain Architecture A. Model In this model, when on-board sensors collect the driving data, they undergo standardization and other necessary processes, as shown in Fig. 2. Then, update the block through Internet of Vehicles (IoV). After a period, the system will automatically calculate the risk level of driving behavior and do a premium adjustment or other applications by smart contract or API. This model modified the patent of synchronization of vehicle sensor information to meet the requirements to sync the data to the blockchain architecture. Because the data’s volume is very large, the application

An On-Board Equipment and Blockchain-Based Automobile Insurance

227

of 5G technology is required to assure that the synchronization is successful and computation is distributed to analyze the data efficiently [11]. The collected data will be used to establish the risk model to make the premium calculation more precise and increase the profitability. Also, collecting data by this model can develop additional services, which will be discussed in the following section.

Fig. 2. Model of usage-based automobile insurance based on blockchain

B. Blockchain Architecture This system will be built with reference to the Tendermint architecture. The most important member of the UBI blockchain is due to the driver’s driving behavior will be recorded to the blockchain in detail [12]. This differs from the existing usage-based automobile insurance product. Under this architecture, there are many third-party members; for example, not only the driver and insurance company but also the policy agency, vehicle supervisor, and vehicle manufacture can access the blockchain data. However, all the data are not public; any member who wants to access the data must have a legitimate reason and adhere to the strict personal data access policy to ensure that data will not be abused. The blockchain architecture is shown in Fig. 3. To establish a complete ecosystem and provide additional services, this platform collects not only driving data but also vehicle diagnosis data and uploads them to the blockchain after preprocessing. Initially, on-board equipment may collect data on driving behavior and vehicle condition, such as maximum speed, average speed differential, driving habits, driving path, and vehicle diagnosis data [13]. Also, when analyzing data, this platform will collect path condition and weather based on the driving path and time to consider all conditions as much as possible. C. The Advantages of Using Blockchain Compared to the existing usage-based automobile insurance, this model can collect all of the data from on-board sensors, not only specific data. Using blockchain as a basis, the data cannot be modified easily; that is, the data is trustable. Therefore, the application of data can be much more than before. In the past, the data can only be used for analysis and to adjust the risk level of the driver. However, in the current model, the data can be considered as evidence when accidents occur and, by syncing to the car repair shops periodically, they can recommend drivers to maintain the car in advance. Because vehicle manufactures join the blockchain as a member, they can help confirm that the collected data is correct and has not been manipulated illegally [14]. Therefore, all the data is highly trustable.

228

W.-Y. Lin et al.

Fig. 3. Architecture of blockchain usage-based insurance

4 Solution A. The Stakeholders of Blockchain (1) Driver The driver is one of the most important stakeholders of this blockchain because the driver’s driving data will be collected in detail. This is because driving data is the basis of the following services: the driver’s willingness to join the blockchain and share driving data depends on the assurance that his/her privacy can be well protected. To achieve this objective, de-identification and access control are necessary. This system can provide drivers with driving risk assessment to recommend them how to drive safer and be the basis of insurance fare adjustment. (2) Vehicle Owner The vehicle owner has right to install on-board equipment. Therefore, the practicality of the following services is important. If the vehicle owner permits onboard equipment collection and uploads driving data and vehicle diagnosis data continuously, this platform can offer the driver’s basic vehicle diagnosis, vehicle maintenance record, advance maintenance suggestion, and automobile insurance fare adjustment. (3) Insurer According to the blockchain and platform’s structure, the insurer will collect and analyze all the data to precisely analyze the insured automobile and driver’s risk, optimize the automobile insurance product, and offer insurance with an immediate automatic claim. Also, the insurer, vehicle manufacture, and vehicle repair shop can form an alliance to reduce the overall risk and raise incomes. (4) Vehicle Manufacturer The vehicle manufacturer is an indispensable participant of this platform because it can decide whether data can be accessed by on-board equipment. It can provide the vehicle’s parameters to better provide maintenance suggestions and localize the vehicles’ design by analyzing the data on the blockchain.

An On-Board Equipment and Blockchain-Based Automobile Insurance

229

B. Service Interaction Modes (1) Driving Risk Assessment This is one of the basic applications of this platform. The car’s on-board equipment will keep continue to upload the driving data to the blockchain periodically or when critical events occur, e.g., engine start and turn off, dangerous driving behaviors, accidents, or components breakdowns, because uploading driving data every second is not economic. This platform will periodically analyze these data to divide the driver into his risk group. When calculating the driving risk, the platform will automatically select the driving data of a period and use some criteria, e.g., average driving speed, times of braking hard, times of turning without flashing the light, to calculate the driving risk. (2) Basic Vehicle Diagnosis The existing on-board equipment product can provide a very simple vehicle diagnosis. However, due to the participation of the vehicle manufactures, this platform can provide a more powerful vehicle diagnosis initiatively. (3) Vehicle Maintenance History Because this platform contains not only vehicle manufacturers but also vehicle repair shops, the blockchain will record the history of the vehicle maintenance, and the vehicle owner can query them at any time. This record can be very useful when trading and understanding a vehicle. (4) Insured Automobile’s Risk Analysis By recording the vehicle maintenance history and analyzing its situation, this platform can analyze both driver’ risk and automobile’s risk separately. This platform can develop the insurance product differently than any existing platform. On the other hand, the vehicle owner can decide the best portfolio of maintenance fee and insurance fee, as shown in Fig. 4. C. Currency Establishment Mechanism The services this platform provided may cause currency flows; specifically, if this platform uses the cryptocurrency as the currency, it will need a currency clearing mechanism. We proposed a mechanism such as the existing electronic payment clearing mechanism, using a virtual account to record every user’s balance. Also, to make this platform selfsufficient, this platform may charge the user for some services. Figure 5 shows the cash flow of this platform.

230

W.-Y. Lin et al.

Fig. 4. Flowchart of insured automobile’s risk analysis

Fig. 5. Cash flowchart of the platform

5 Discussion The privacy issue is that all of the drivers who want to buy this type of usage-based automobile insurance product care. However, the usage-based automobile insurance product needs as many as drivers provide their driving data to establish a precise risk prediction model. Therefore, it is important to minimize the drivers’ hesitation and maximize the benefits, not only in consideration of a possible premium bonus but also other additional services.

An On-Board Equipment and Blockchain-Based Automobile Insurance

231

To attract enough drivers to buy this product, the benefits of buying this product must be higher than the sacrifice of privacy. After buying this UBI product, the driver’s premium might be much more than before. Both will make the driver want to drop out of the blockchain [15]. To prevent this situation, the fare adjustment formula must be public to let every driver easily know why the driver’s premium is decreased or increased. Also, this platform can let the driver see the driver’s driving habit clearly. That is, the insurance company can encourage the driver as much as possible. By taking this strategy, not only might the total compensation be decreased, but the customer pool can also be increased. The privacy issue is not only a difficultly but also an important management issue.

6 Conclusion and Future Works In this study, we have provided a potential solution for establishing a UBI platform based on blockchain and on-board equipment. In this solution, we have explained why blockchain should be used as this platform’s basis, what the possible stakeholders will do, how they interact in possible applications, how this platform can break the limitation of the usage of driving data, and how to preliminarily resolve the privacy and access control issues caused by adopted blockchain as basis. By adopting this solution, the maintenance platform and UBI product can reach multiple goals, e.g., ensuring that collected driving data is trustable, pricing more precisely, providing more offline functions, helping to establish a related ecosystem, and providing more complete services by analyzing the driving data. Furthermore, by establishing an online and offline ecosystem, the UBI products can provide the insured with more benefits to make them more willing to buy UBI products and provide their driving data for insurers to analysis. Thus, not only can the insured benefit from this platform and additional services, but the insurer can also sell more UBI policies and improve the pricing and services. The ecosystem established by adopting this platform can also benefit. It is too complicated to precisely determine the costs and the benefits of using this platform. However, we can say that the most expensive cost is to establish the blockchain to store the data and maintain it, and the biggest benefit is that every stakeholder can benefit from the provision of services to other stakeholders. To sum up, we believe that the structure of this platform would be helpful for P&C insurance companies, financial technology companies, or anyone who wants to develop new UBI products and services or research this area. Nevertheless, there are still some issues limited by the scope of our research. First, due to this system and platform, which focused on the automobiles, the other objects’ data on the road may not be completely collected by this platform. This system has its limitations. It cannot replace the existing road supervision systems, but it can help improve them. The most important value this system may bring is how the driving data this platform collected will help the development of the car industry. In the past, the car manufacture was unable to collect the driving data widely. However, by this system, how will car manufacture use the data to develop their new product. This will be a very potential value. By analyzing the numerous driving data, this platform can develop a more precise model for drivers’ behaviors. Compared to the existing solution, by gender or age, this model may change the driving behavior evaluation system in the future.

232

W.-Y. Lin et al.

In the future, considering the development of the self-driving car, it may make a great influence of the automobile insurance. Therefore, the model and the UBI products that depend on the platform structure proposed by this study should consider that if necessary.

References 1. Andem, I.E.: Perceived influence of relationship communication on insurance service quality and customer satisfaction. Indep. J. Manag. Prod. 10(1), 234–258 (2019) 2. Taian Insurance Co. Ltd.: UBI smart automobile insurance 2020. https://www.taian.com.tw/ web_ubi/index.asp. Accessed: 20 Apr 2020 3. Binion, T., Harr, J., Fields, B., Cielocha, B., Balbach, S.J.: Synchronization of vehicle sensor information, U.S. Patent 9,959,764, 1 May 2018 4. Federico, F., Shane, K.H.W., Emmanuel, T., Jun, E.L., Luke, S.T.W.: Design and implementation of a compact rotational speed and air flow sensor for unmanned aerial vehicles. IEEE Sens. J. 19(22), 10298–10307 (2019) 5. Jiang, L., Xie, S., Maharjan, S., Zhang, Y.: Joint transaction relaying and block verification optimization for blockchain empowered D2D communication. IEEE Trans. Veh. Technol. 69(1), 828–841 (2020) 6. Pan, X., Song, M., Ai, B., Ming, Y.: Blockchain technology and enterprise operational capabilities: an empirical test 52, 1–9 (2020) 7. di Silvestre, M.L., Gallo, P., Ippolito, M.G., Musca, R., Sanseverino, E.R., Tran, Q.T.T., Zizzo, G.: Ancillary Services in the energy blockchain for microgrids. IEEE Trans. Ind. Appl. 55(6), 7310–7319 (2019) 8. Cook, B.: Blockchain: transforming the seafood supply chain (2018). http://awsassets.wwfnz. panda.org/downloads/draft_blockchai_report_1_4_1.pdf. Accessed 26 Oct 2019 9. Sahoo, S., Fajge, A., Halder, R., Cortesi, A.: A hierarchical and abstraction-based blockchain model. Appl. Sci.-basel 9(11), 1–20 (2019) 10. Hu, W., Fan, Z., Gao, Y.: Research on smart contract optimization method on blockchain. IT Prof. 21(5), 33–38 (2019) 11. Nkenyereye, L., Tama, B., Shahzad, M., Choi, Y.: Secure and blockchain-based emergency driven message protocol for 5G enabled vehicular edge computing. Sensors 1(1), 1–16 (2020) 12. Zhang, Y., He, D., Choo, K.: BaDS: blockchain-based architecture for data sharing with ABS and CP-ABE in IoT. Wirel. Commun. Mob. Comput. 2018, 1–10 (2018) 13. Mazzei, D., Baldi, G., Fantoni, G., Montelisciani, G., Pitasi, A., Ricci, L., Rizzello, L.S.: A blockchain tokenizer for industrial IOT trustless applications. Future Gener. Comput. Syst. Int. J. Sci. 105, 432–445 (2020) 14. Singh, M., Kim, S.: Branch based blockchain technology in intelligent vehicle. Comput. Netw. 145, 219–231 (2018) 15. Benarous., L., Kadri, B., Bouridane, A.: Blockchain-based privacy-aware pseudonym management framework for vehicular networks. Arab. J. Sci. Eng. 15, 1–34 (2020)

An Integrated Fuzzy-Based Simulation System for Driver Risk Management in VANETs Considering Relative Humidity as a New Parameter Kevin Bylykbashi1(B) , Ermioni Qafzezi1 , Makoto Ikeda2 , Keita Matsuo2 , Leonard Barolli2 , and Makoto Takizawa3 1

Graduate School of Engineering, Fukuoka Institute of Technology (FIT), 3-30-1 Wajiro-Higashi, Higashi-Ku, Fukuoka 811-0295, Japan [email protected], [email protected] 2 Department of Information and Communication Engineering, Fukuoka Institute of Technology (FIT), 3-30-1 Wajiro-Higashi, Higashi-Ku, Fukuoka 811-0295, Japan [email protected], {kt-matsuo,barolli}@fit.ac.jp 3 Department of Advanced Sciences, Faculty of Science and Engineering, Hosei University, 3-7-2, Kajino-machi, Koganei-shi, Tokyo 184-8584, Japan [email protected]

Abstract. In this paper, we present an integrated Fuzzy-based Simulation System for Driver Risk Management (FSSDRM) in Vehicular Ad hoc Networks (VANETs). FSSDRM considers the current condition of different parameters which have an impact on the driver and vehicle performance to assess the risk level. The considered parameters include vehicle’s Environment Temperature (ET), Relative Humidity (RH), Noise Level (NL), Driver’s Health Condition (DHC), Weather Condition (WC), Road Condition (RC) and Vehicle Speed (VS). FSSDRM is composed of three Fuzzy Logic Controllers (FLCs): FLC1, FLC2 and FLC3. FLC1 has the following inputs: RH, NL and ET, while WC, RC and VS are the inputs of FLC2. Both outputs of these two FLCs together with DHC, serve as input parameters for FLC3. The input parameters’ data can come from different sources, such as on-board and on-road sensors and cameras, sensors and cameras in the infrastructure and from the communications between the vehicles. Based on the system’s final output i.e., driving risk level, a smart box informs the driver for a potential risk/danger and provides assistance. We show through simulations the effect of the considered parameters on the determination of the driving risk and demonstrate a few actions that can be performed accordingly.

1

Introduction

Traffic accidents, road congestion and environmental pollution are persistent problems faced by both developed and developing countries, which have made c The Editor(s) (if applicable) and The Author(s), under exclusive license  to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 233–243, 2021. https://doi.org/10.1007/978-3-030-61108-8_23

234

K. Bylykbashi et al.

people live in difficult situations. Among these, the traffic incidents are the most serious ones because they result in huge loss of life and property. For decades, we have seen governments and car manufacturers struggle for safer roads and car accident prevention. The development in wireless communications has allowed companies, researchers and institutions to design communication systems that provide new solutions for these issues. Therefore, new types of networks, such as Vehicular Ad hoc Networks (VANETs) have been created. VANET consists of a network of vehicles, in which vehicles are capable of communicating among themselves in order to deliver valuable information such as safety warnings and traffic information. Nowadays, every car is likely to be equipped with various forms of smart sensors, wireless communication modules, storage and computational resources. The sensors gather information about the road and environment conditions and share it with neighboring vehicles and adjacent roadside units (RSU) via vehicleto-vehicle (V2V) or vehicle-to-infrastructure (V2I) communication. However, the difficulty lies on how to understand the sensed data and how to make intelligent decisions based on the provided information. As a result, various intelligent computational technologies and systems such as fuzzy logic, machine learning, neural networks, adaptive computing and others, are being or already deployed by many car manufacturers [7]. They focus on these auxiliary technologies to launch and fully support the driverless vehicles. Fully autonomous vehicles still have a long way to go but driving support technologies are becoming widespread, even in everyday cars. The goal is to improve both driving safety and performance relying on the measurement and recognition of the outside environment and their reflection on driving operation. On the other hand, we are focused not only on the outside information but also on the in-car information and driver’s health information to detect a potential accident or a risky situation, and alert the driver about the danger, or take over the steering if it is necessary. We aim to realize a new intelligent driver support system which can provide an output in real-time by combining information from many sources. In this work, we implement a fuzzy-based simulation system for driving risk management considering different types of parameters that can affect the driving performance in a direct and indirect way. The model of our proposed system is given in Fig. 1. The considered parameters include environmental factors that have an impact on the driver and factors that could impact the vehicle performance. The parameters that affect the driver and we consider for the implementation of our system are the vehicle’s environment condition such as temperature, relative humidity and noise level, while the considered parameters that affect the driving performance are driver’s health condition, weather condition, road condition and vehicle speed. Different from our previous works we include all these parameters all at once in an integrated fuzzy-based system which is composed of three Fuzzy Logic Controllers (FLCs). Also, we include the relative humidity as a new parameter for a more accurate assessment of the vehicle’s environment condition. Based on the proposed system’s output value,

An Integrated Fuzzy-Based System for Driving Risk Management

235

Vehicle Speed

Actor Device Weather Condition Road Condition Vehicle’s Environment Condition

Driver’s Health Condition

Fig. 1. Proposed system architecture.

it can be decided if an action is needed, and if so, which is the appropriate task to be performed in order to provide a better driving support. The structure of the paper is as follows. Section 2 presents a brief overview of VANETs. Section 3 describes the proposed fuzzy-based simulation system and its implementation. Section 4 discusses the simulation results. Finally, conclusions and future work are given in Sect. 5.

2

Overview of VANETs

VANETs are a special case of Mobile Ad hoc Networks (MANETs) in which the mobile nodes are vehicles. In VANETs, nodes (vehicles) have high mobility and tend to follow organized routes instead of moving at random. Moreover, vehicles offer attractive features such as higher computational capability and localization through GPS. VANETs have huge potential to enable applications ranging from road safety, traffic optimization, infotainment, commercial to rural and disaster scenario connectivity. Among these, the road safety and traffic optimization are considered the most important ones as they have the goal to reduce the dramatically high number of accidents, guarantee road safety, make traffic management and create new forms of inter-vehicle communications in Intelligent Transportation Systems (ITS). The ITS manages the vehicle traffic, support drivers with safety and other information, and provide some services such as automated toll collection and driver assist systems [8]. Despite the attractive features, VANETs are characterized by very large and dynamic topologies, variable capacity wireless links, bandwidth and hard delay constrains, and by short contact durations which are caused by the high mobility, high speed and low density of vehicles. In addition, limited transmission ranges, physical obstacles and interferences, make these networks characterized by disruptive and intermittent connectivity.

236

K. Bylykbashi et al.

To make VANETs applications possible, it is necessary to design proper networking mechanisms that can overcome relevant problems that arise from vehicular environments.

3

Proposed Fuzzy-Based Simulation System

Although the developments in autonomous vehicle design indicate that this type of technology is not that far away from deployment, the current advances fall only into the Level 2 of the Society of Automotive Engineers (SAE) levels [14]. However, the automotive industry is very competitive and there might be many other new advances in the autonomous vehicle design that are not launched yet. Thus, it is only a matter of time before driverless cars are on the road. On the other side, there will be many people who will still be driving even on the era of autonomous cars. The high cost of driverless cars, lack of trust and not wanting to give up driving might be among the reasons why those people will continue to drive their cars. Hence, many researchers and automotive engineers keep working on Advanced Driver Assistance Systems (ADASs) as a primary safety feature, required in order to achieve full marks in safety. ADASs are intelligent systems that reside inside the vehicle and help the driver in a variety of ways. These systems rely on a comprehensive sensing network and artificial intelligence techniques, and have made it possible to commence the era of connected cars. They can invoke action to maintain driver’s attention in both manual and autonomous driving. While the sensors are used to gather data regarding the inside/outside environment, vehicle’s technical status, driving performance and driver’s condition, the intelligent systems task is to make decisions based on these data. If the vehicle measurements are combined with those of the surrounding vehicles and infrastructure, a better environment perception can be achieved. In addition, with different intelligent systems located at these vehicles as well as at geographically distributed servers, more efficient decisions can be attained. Our research work focuses on developing an intelligent non-complex driving support system which determines the driving risk level in real-time by considering different types of parameters. In the previous works, we have considered different parameters including in-car environment parameters such as the ambient temperature and noise, and driver’s vital signs, i.e. heart and respiratory rate for which we implemented a testbed and conducted experiments in a real scenario [2,5]. The considered parameters include environmental factors and driver’s health condition which can affect the driver capability and vehicle performance. In [4], we included vehicle speed in our intelligent system for its crucial impact on the determination of risk level and in [3,6] the weather condition and road condition were added. In this work, we not only consider relative humidity as a new parameter but also include all the aforementioned parameters all at once in an integrated fuzzy-based system. We use fuzzy logic to implement the proposed system as it can make a realtime decision based on the uncertainty and vagueness of the provided information [1,9–13,15,16].

An Integrated Fuzzy-Based System for Driving Risk Management

237

Fig. 2. Proposed system structure.

The proposed system called Fuzzy-based Simulation System for Driving Risk Management (FSSDRM) is shown in Fig. 2. It consists of three FLCs, two of which are used to lower the overall complexity of the system. While it seems more complex by the way our system is built, it is far better than having seven input parameters in a single FLC because this would result in a very complex Fuzzy Rule Base (FRB) composed of hundreds of rules, which, in turn, would increase the overall complexity of the system. FSSDRM has the following inputs: vehicle’s Environment Temperature (ET), Relative Humidity (RH), Noise Level (NL), Driver’s Health Condition (DHC), Weather Condition (WC), Road Condition (RC) and Vehicle Speed (VS), and its output is the Driving Risk Management (DRM). FLC1 makes use of RH, NL and ET, whereas the inputs of FLC2 are WC, RC and VS. The outputs of these two FLCs–Vehicle’s Environment Condition (VEC) and Weather-Road-Speed (WRS), serve alongside DHC, as input parameters for FLC3. FLC2 is presented in a previous work, therefore in this work, we describe in detail only FLC1 and FLC3. The term sets of linguistic parameters of FLC1 and FLC3 are defined respectively as: T (RH) = {Low (L), M oderate (M ), High (H)}; T (N L) = {Quiet (Q), N oisy (N ), V ery N oisy (V N )}; T (ET ) = {Low (Lo), M edium (M e), High (Hi)}; T (V EC) = {Extremely U ncomf ortable (EU C), V ery U ncomf ortable (V U C), U ncomf ortable (U C), M oderate (M od), Comf ortable (C)}; T (W RS) = {N o/M inor Danger (N/M D), M oderate Danger(M D), Considerable Danger(CD), High Danger(HD), V ery High Danger(V HD)}; T (DHC) = {V ery Bad (V B), Bad (B), Good (G)}. T (DRM ) = {Saf e (Sf ), V ery Low (V L), Low (Lw), M oderate (M d), Considerable (Co), High (Hg), V ery High (V H), Severe (Sv), Danger (D)}.

Based on the linguistic description of input and output parameters, we make the Fuzzy Rule Bases (FRBs) of the FLCs. The FRB forms a fuzzy set of dimen-

238

K. Bylykbashi et al. Table 1. FRB of FLC1. No RH NL ET VEC No RH NL ET VEC No RH NL ET VEC 1

L

Q

Lo UC

10 M

Q

Lo Mod 19 H

Q

Lo UC

2

L

Q

Me Mod 11 M

Q

Me C

20 H

Q

Me Mod

3

L

Q

Hi

12 M

Q

Hi

Mod 21 H

Q

Hi

4

L

N

Lo VUC 13 M

N

Lo UC

22 H

N

Lo VUC

5

L

N

Me UC

14 M

N

Me Mod 23 H

N

Me UC

6

L

N

Hi

VUC 15 M

N

Hi

N

Hi

7

L

VN Lo EUC 16 M

VN Lo VUC 25 H

VN Lo EUC

8

L

VN Me VUC 17 M

VN Me UC

VN Me VUC

9

L

VN Hi

VN Hi

UC

EUC 18 M

UC

24 H 26 H

VUC 27 H

VN Hi

UC

VUC

EUC

Table 2. FRB of FLC3. No VEC WRS

DHC DRM No VEC WRS

DHC DRM

1

EUC N/MD VB

DHC DRM No VEC WRS VH

26 VUC HD

B

Sv

51 Mod MD

G

Lw

2

EUC N/MD B

Hg

27 VUC HD

G

VH

52 Mod CD

VB

Hg

3

EUC N/MD G

Co

28 VUC VHD

VB

D

53 Mod CD

B

Co

4

EUC MD

VB

Sv

29 VUC VHD

B

D

54 Mod CD

G

Md

5

EUC MD

B

VH

30 VUC VHD

G

Sv

55 Mod HD

VB

VH

6

EUC MD

G

Hg

31 UC

N/MD VB

Co

56 Mod HD

B

Hg

7

EUC CD

VB

D

32 UC

N/MD B

Md

57 Mod HD

G

Co

8

EUC CD

B

Sv

33 UC

N/MD G

Lw

58 Mod VHD

VB

Sv

9

EUC CD

G

VH

34 UC

MD

VB

Hg

59 Mod VHD

B

VH

10 EUC HD

VB

D

35 UC

MD

B

Co

60 Mod VHD

G

Hg

11 EUC HD

B

D

36 UC

MD

G

Md

61 C

N/MD VB

Lw

12 EUC HD

G

Sv

37 UC

CD

VB

VH

62 C

N/MD B

VL

13 EUC VHD

VB

D

38 UC

CD

B

Hg

63 C

N/MD G

Sf

14 EUC VHD

B

D

39 UC

CD

G

Co

64 C

MD

VB

Md

15 EUC VHD

G

Lw

D

40 UC

HD

VB

Sv

65 C

MD

B

16 VUC N/MD VB

Hg

41 UC

HD

B

VH

66 C

MD

G

VL

17 VUC N/MD B

Co

42 UC

HD

G

Hg

67 C

CD

VB

Co

18 VUC N/MD G

Md

43 UC

VHD

VB

D

68 C

CD

B

Md

19 VUC MD

VB

VH

44 UC

VHD

B

Sv

69 C

CD

G

Lw

20 VUC MD

B

Hg

45 UC

VHD

G

VH

70 C

HD

VB

Hg

21 VUC MD

G

Co

46 Mod N/MD VB

Md

71 C

HD

B

Co

22 VUC CD

VB

Sv

47 Mod N/MD B

L3

72 C

HD

G

Md

23 VUC CD

B

VH

48 Mod N/MD G

VL

73 C

VHD

VB

VH

24 VUC CD

G

Hg

49 Mod MD

VB

Co

74 C

VHD

B

Hg

25 VUC HD

VB

D

50 Mod MD

B

Md

75 C

VHD

G

Co

An Integrated Fuzzy-Based System for Driving Risk Management

239

sions | T (x1 ) | × | T (x2 ) | × · · · × | T (xn ) |, where | T (xi ) | is the number of terms on T (xi ) and n is the number of FLC input parameters. FLC1 has three input parameters with three linguistic terms each, therefore, there are 27 rules in the FRB1, which is shown in Table 1. The FRB of FLC3 is shown in Table 2. Since FLC3 has three input parameters, with two parameters having five linguistic terms each and one parameter having three, it consists of 75 rules. The control rules of FRB have the form: IF “conditions” THEN “control action”. The membership functions used for fuzzification and defuzzification are given in Fig. 3. We use triangular and trapezoidal membership functions because they are suitable for real-time operation.

Fig. 3. Membership functions.

240

K. Bylykbashi et al.

45

Fig. 4. Simulation results for FLC1

4

Simulation Results

In this section, we present the simulation results for our proposed system. The simulation results are presented in Fig. 4 and Fig. 5. In Fig. 4 is shown the relation between VEC and ET for different RH and NL values. Regarding RH, we consider the scenarios with a dry, normal and humid inside environment represented by a 10%, 45% and 90% relative humidity, whereas for the NL parameter we consider the values 45, 70 and 85 dB which simulate a quiet, noisy and very noisy environment, respectively. Relative humidity plays an important role in the comfort levels and we can see this fact also in our results. When the inside ambient is dry or humid (see Fig. 4(a) and Fig. 4(c)), we can see that there is not any situation that the vehicle environment condition is decided as comfortable. The only scenario when the environment is comfortable is when the humidity is in normal levels. However, regardless the normal levels of RH, if driver is driving in low/high temperatures or in a noisy environment, VEC is never decided as comfortable. Anyway, a moderate level of comfort is still acceptable but any other level may have an impact on the driver, which in turn may affect the driving performance. In Fig. 5 is shown the relation between DRM and DHC for different VEC and WRS values. We consider all the possible levels of VEC induced by the different RH, NL and ET values. To be specific, we consider the values 0.1, 0.3, 0.5, 0.7 and

An Integrated Fuzzy-Based System for Driving Risk Management

241

Fig. 5. Simulation results for FLC3.

0.9 which indicate scenarios with extremely uncomfortable, very uncomfortable, uncomfortable, moderate and comfortable environment, respectively. While for WRS, we consider No or Minor, Moderate, Considerable, High and Very High Danger represented by the 0.1, 0.3, 0.5, 0.7 and 0.9 values. In Fig. 5(a), we consider the VEC value 0.1 and change the WRS from 0.1 to 0.9. We can see that all the DRM values are equal or greater than 0.5 which indicate potential risks. This is due to the fact that the VEC is extremely uncomfortable i.e., dry/humid, cold/hot and very noisy environment, which could affect the driver’s ability to focus on the driving process. In addition, if the weather

242

K. Bylykbashi et al.

and road condition are not good, or the driver is driving at a high speed, the risk levels are decided even up to the “Danger” level. If driver’s health condition parameter indicates a very good status of his/her health, we can see a slight improvement on the risk levels but still not enough for a situation with a moderate risk. By comparing the DRM values for all the considered scenarios, we can see the significant impact of the VEC on the determination of the driving risk. As we can see from Fig. 5(c), Fig. 5(d) and Fig. 5(e), when VEC is equal or greater than 0.5, situations with low risk are now present and happen when the driver is in good health condition while WRS remains low. When VEC is less than 0.5, the driver cannot manage to drive with low risk even if he/she is in good health condition and WRS is No or Minor Danger. In the cases when the risk level is above the moderate level for a number of consecutive decided DRM values, the system can perform a certain action. For example, when the DRM value is slightly above the moderate level the system may take action to lift the driver’s mood, or improve the conditions of the inside environment by adjusting the temperature and humidty levels. On the other hand, when the DRM value is very high, the system could even decide to limit the vehicle’s operating speed to a speed that the risk level is decreased significantly.

5

Conclusions

In this paper, we presented an integrated fuzzy-based system to decide the driving risk management. We considered seven parameters: the vehicle’s environment condition decided by the relative humidity, noise level and temperature, the road condition, weather condition, vehicle speed, and driver’s health condition. We showed through simulations the effect of the considered parameters on the determination of the driving risk level. In addition, we demonstrated a few actions that can be performed based on the output of our system. However, it may occur that the system provides an output which determines a low risk, when actually the chances for an accident to happen are high, or the opposite scenario, which is the case when the system’s output implies a false alarm. Therefore, we intend to implement the system in a testbed and estimate the system performance by looking into correct detection and false positives/negatives to determine its accuracy.

References 1. Bylykbashi, K., Elmazi, D., Matsuo, K., Ikeda, M., Barolli, L.: Effect of security and trustworthiness for a fuzzy cluster management system in VANETs. Cogn. Syst. Res. 55, 153–163 (2019). https://doi.org/10.1016/j.cogsys.2019.01.008

An Integrated Fuzzy-Based System for Driving Risk Management

243

2. Bylykbashi, K., Elmazi, D., Matsuo, K., Ikeda, M., Barolli, L.: Implementation of a fuzzy-based simulation system and a testbed for improving driving conditions in VANETs. In: International Conference on Complex, Intelligent, and Software Intensive Systems, pp. 3–12. Springer (2019). https://doi.org/10.1007/978-3-03022354-01 3. Bylykbashi, K., Qafzezi, E., Ampririt, P., Matsuo, K., Barolli, L., Takizawa, M.: A fuzzy-based simulation system for driving risk management in VANETs considering road condition as a new parameter. In: International Conference on Intelligent Networking and Collaborative Systems. Springer (2020) 4. Bylykbashi, K., Qafzezi, E., Ikeda, M., Matsuo, K., Barolli, L.: A fuzzy-based system for driving risk measurement (FSDRM) in VANETs: a comparison study of simulation and experimental results. In: International Conference on P2P, Parallel, Grid, Cloud and Internet Computing, pp. 14–25. Springer (2019) 5. Bylykbashi, K., Qafzezi, E., Ikeda, M., Matsuo, K., Barolli, L.: Fuzzy-based Driver Monitoring System (FDMS): implementation of two intelligent FDMSs and a testbed for safe driving in VANETs. Future Gener. Comput. Syst. 105, 665–674 (2020). https://doi.org/10.1016/j.future.2019.12.030 6. Bylykbashi, K., Qafzezi, E., Ikeda, M., Matsuo, K., Barolli, L., Takizawa, M.: A fuzzy-based simulation system for driving risk management in VANETs considering weather condition as a new parameter. In: International Conference on Innovative Mobile and Internet Services in Ubiquitous Computing, pp. 23–32. Springer (2020) 7. Gusikhin, O., Filev, D., Rychtyckyj, N.: Intelligent vehicle systems: applications and new trends. In: Informatics in Control Automation and Robotics, pp. 3–14. Springer (2008). https://doi.org/10.1007/978-3-540-79142-31 8. Hartenstein, H., Laberteaux, L.: A tutorial survey on vehicular ad hoc networks. IEEE Commun. Mag. 46(6), 164–171 (2008) 9. Kandel, A.: Fuzzy Expert Systems. CRC Press, Boca Raton (1991) 10. Klir, G.J., Folger, T.A.: Fuzzy Sets, Uncertainty, and Information. Prentice Hall Inc., Upper Saddle River (1987) 11. McNeill, F.M., Thro, E.: Fuzzy Logic: A Practical Approach. Academic Press, Cambridge (1994) 12. Munakata, T., Jani, Y.: Fuzzy systems: an overview. Commun. ACM 37(3), 69–77 (1994). https://doi.org/10.1145/175247.175254 13. Qafzezi, E., Bylykbashi, K., Ikeda, M., Matsuo, K., Barolli, L.: Coordination and management of cloud, fog and edge resources in SDN-VANETs using fuzzy logic: a comparison study for two fuzzy-based systems. Internet Things 11, 100169 (2020) 14. SAE On-Road Automated Driving (ORAD) Committee: Taxonomy and definitions for terms related to driving automation systems for on-road motor vehicles. Technical report, Society of Automotive Engineers (SAE) (2018). https://doi.org/10. 4271/J3016201806 15. Zadeh, L.A., Kacprzyk, J.: Fuzzy Logic for the Management of Uncertainty. Wiley, New York (1992) 16. Zimmermann, H.J.: Fuzzy Set Theory and Its Applications. Springer, New York (1996). https://doi.org/10.1007/978-94-015-8702-0

IoT Device Power Management Based on PSM and EDRX Mechanisms Kun-Lin Tsai1,3 , Fang-Yie Leu2(B) , Tz-Yuan Huang1 , and Hao-En Yang1 1 Department of Electrical Engineering, Tunghai University,

Taichung City, Taiwan [email protected], [email protected], [email protected] 2 Department of Computer Science, Tunghai University, Taichung City, Taiwan [email protected] 3 Research Center for Smart Sustainable Circular Economy, Tunghai University, Taichung City, Taiwan

Abstract. Recently, many Internet of Things (IoT) related applications have been designed to assist industrial development and human beings’ lives. Among these IoT applications, many of them require small amount of data but long distant communication with the features of high-reliability and high level of security. Moreover, since many IoT user equipments (UEs) are powered by batteries, their energy management is an important design consideration. In order to support low-power communication, the 3GPP has proposed two energy-saving mechanisms, i.e., extended Discontinuous Reception (eDRX) and Power Save Mode (PSM). However, how to effectively integrate these two mechanisms to adjust UE’s parameters so that the UE’s energy consumption can be optimal is not an easy task. In this paper, the UE’s energy consumption model by combining eDRX with PSM is investigated and the energy consumption of a UE is also analyzed. According to the Poisson distribution, the uplink and downlink data transmission rates are fused in the energy consumption model. Our simulation results show that for an IoT with small amount of data transmission, the energy consumption can be significantly reduced when setting proper eDRX and PSM parameters. Keywords: Energy management · PSM · eDRX · Internet-of-Things (IoT) · NB-IoT

1 Introduction In the past decade, many Internet of Things (IoT) applications have been developed due to the fast progress of information and communication technologies. IoT connects many devices, sensors, and computers via conventional networks or mobile networks so that the connected objects can communicate with each other. According to the Ericsson Mobility Report 2020 [1], the number of Massive IoT connections increased by a factor of 3 during 2019, reaching close to 100 million. Moreover, a total number of IoT connections, © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 244–253, 2021. https://doi.org/10.1007/978-3-030-61108-8_24

IoT Device Power Management Based on PSM and EDRX Mechanisms

245

including wide-area IoT, cellular IoT, and short-range IoT, will approach 24.6 billion in 2025. The communication technology plays an important role in an IoT system, and it can be mobile network, 802.11 [2], Bluetooth [3], Zigbee [4], LoRa [5], NB-IoT [6], etc. Among them, LoRa and NB-IoT have attracted many users’ attention due to their low power and the features of wide range communication. The NB-IoT related specifications developed by 3GPP (3rd Generation Partnership Project) utilize existing mobile networks which have many features, e.g., massive connectivity, enhanced coverage, and ultra-low power consumption. There are three operation modes in an NB-IoT system, i.e., In Band, Guard Band, and Stand Alone, which can be easily equipped inside the mobile devices by upgrading software so that the transmission quality and data security can be guaranteed. Since most IoT devices are powered by batteries and deployed on a large scale, the costs of battery replacement are high. In order to effectively implement the NB-IoT based system, the energy consumption of IoT user equipment (UE) must reduce as much as possible. The 3GPP 23.682-4.5.4 UE [7] defines Power Saving Mode (PSM) which is similar to power-off, but the UE remains registered within the network. As a result, the UE need not to re-attach or re-establish the Packet Data Network (PDN) connections. In 3GPP Release-13, the Extended Discontinuous Reception (eDRX) [8] mechanism is proposed to extend the sleeping duration of traditional DRX so that more energy can be saved. Detailed descriptions of PSM and DRX will be discussed in next section. Both PSM and eDRX reduce energy consumption at the cost of increasing communication latency. In order to seek the balance between energy consumption and transmission performance, in this paper, the impact of eDRX and PSM parameters on energy consumption under different transmission requirements are analyzed. We discuss the PSM and eDRX models and consider various influencing factors, including packet arrival rate, number of paging cycles, energy consumption of each state, etc. Then, the parameters that affect energy consumption and transmission performance are defined by utilizing mathematical models as so to formally explore the relationship and influence among parameters. Our simulation results by using the MATLAB® [9] show that the UEs’ energy consumption can be effectively lowered when PSM and eDRX parameters are configured appropriately. The rest of this paper is organized as follows. Section 2 briefly introduces the PSM and eDRX mechanisms. Besides, some related studies are also investigated in Sect. 2. Section 3 presents the IoT UE power model by using PSM and eDRX mechanisms. Section 4 shows simulation results. Section 5 concludes this paper and descripts some future studies.

2 Preliminary 2.1 PSM Mechanism PSM is a power saving method provided for UE in 3GPP Release-12 [7]. When UE completes data transmission or the Active Timer expires, it enters PSM mode, in which, UE is in a sleep state with ultra-low power consumption. Once UE is in PSM mode, UE is not reachable from mobile core network, but UE is still in a registered state in the mobile core network. Such PSM characteristic can be applied to some IoT devices

246

K.-L. Tsai et al.

which infrequently upload data, only receiving messages from server side. In this paper, we use “UE” to represent an IoT device of an NB-IoT system. The PSM operation and its corresponding power consumption is shown in Fig. 1, in which T3412 is a periodic Tracking Area Update (TAU) timer used to periodically notify the availability of UE to network. T3324 is an active timer which determines how long the UE will monitor paging before entering PSM. When UE finishes TAU procedure and transfers to idle state, both T3412 and T3324 timer will be trigger. Before T3324 timer expires, UE periodically executes Discontinuous Reception (DRX) to listen paging messages sent by core network. If no data transmission is required within T3324 period, UE switches itself from idle to PSM. Power T3412

Tx TAU Rx

T3324

PSM

TAU

DRX

Sleep Deep Sleep

Time

Fig. 1. PSM operation and its corresponding power consumption.

In PSM, T3412 can be set up to 9,920 h, while the longest T3324 defined in the 3GPP standard can only be set up to 186 min, and the shortest can be 0 min. When T3324 is set to 0, it means that UE will immediately switch into the PSM after a TAU procedure. Both T3412 and T3324 can be adjusted by Mobility Management Entity (MME), but MME can also negotiate with UE about the settings of T3324. 2.2 eDRX Mechanism eDRX is an extended version of DRX which enables UE to discontinuously receive the Physical Downlink Control CHannel (PDCCH). According to periodic configuration, there are two stages in each DRX cycle. In the first stage, UE monitors the PDCCH in a short period. In the second stage, the UE suspends monitoring of the PDCCH. DRX truly reduces the energy consumption of UE but still allows the core network to communicate with UE through paging messages. Based on DRX, eDRX controls the Paging Time Window (PTW) by configuring eDRX-Config. In the connected state, the paging cycle is extended to a maximum of 10.24 s, and in the IDLE state, it is up to 40 min. The UE’s sleep cycle, i.e., eDRX cycle, can be as long as 175 min. The eDRX operation and corresponding power consumption is shown in Fig. 2, in which a period of time before entering long-term sleep is so-called

IoT Device Power Management Based on PSM and EDRX Mechanisms

247

Paging Time Window (PTW). During this period, the UE executes the DRX so as to receive paging messages from MME. Power eDRX cycle length duration

Rx

DRX

Sleep Deep Sleep Paging Time Window

Time

Fig. 2. eDRX operation and corresponding power consumption.

2.3 Related Studies In recent years, many researches on NB-IoT energy control have been proposed. Oh and Shin [10] proposed a small data transmission scheme for NB-IoT to effectively use radio resources. In [10], an idle UE is able to deliver small amount of data without Radio Resource Control (RRC) connection setup process, aiming to increase the number of UEs in an NB-IoT system which has insufficient radio resources. Lauridsen et al. [11] proposed two NB-IoT IoT battery power consumption models. Their experimental results claimed that using DRX will shorten the battery life and the reference models provided by 3GPP cannot truly meet the energy requirement in 3GPP standard. Yeoh et al. [12] utilized Non-Access Stratum (NAS) for NB-IoT data transmission and introduced the NAS model to calculate the uplink and downlink RRC power consumption. They combined the PSM mode, T3324 timer with T3412 timer in their model, and their experimental results showed that when UE is registered to the NB-IoT network and the Evolved Packet System (EPS) is triggered each time, the battery life of UE will be greatly reduced. They concluded that a longer PSM will greatly increase the data transmission time, so it is not suitable for those applications with frequent data transmission. Oh et al. [13] also modeled and analyzed the battery consumption rate of NB-IoT devices. In order to explore the efficiency of a RRC connection state in an NB-IoT system, in their battery consumption analysis, they focused on the duration of the RRC connection state. According to the experimental results, they claimed that decreasing the duration of the RRC connection state is helpful to increase the battery life. Bello et al. [14] introduced two sub-states, i.e., random access state and transmission state, in the connected state and then combined these two sub-states with PSM and idle states to form a semi-Markov model which was utilized to evaluate the energy consumption and delay of periodic uplink transmission, so that the user can evaluate the trade-off between energy saving and transmission delay. Sultania et al. [15] also proposed an energy consumption model

248

K.-L. Tsai et al.

for NB-IoT devices by using PSM and eDRX, and simulated the proposed model by NS-3 simulator. They analyzed the NB-IoT energy consumption in different packet arrival time. Their experimental results indicated that more than 12 years battery life can be achieved when transmitting only one packet per day. Although the abovementioned studies investigated PSM and eDRX in an NB-IoT system, what seems to be lacking is the relation among energy consumption, data arrival rate, T3412 duration, and eDRX timer. Thus, in this paper, the UE energy consumption model by combining eDRX with PSM is investigated and then the energy consumption of a UE is analyzed.

3 PSM and eDRX Based IoT Energy Model Since RRC controls the communication between UE and eNB, NB-IoT supports two RRC states, i.e., RRC_Idle state and RRC_Connected state, which are called idle state and connected state for short, respectively. In an idle state, the UE is still registered in but not connected to the mobile core network. During this period, the UE keeps listening paging messages from eNB and measures the signal strength of neighboring cells. When some data require transmitting to/from UE, the UE starts the RRC_Connection_Establishment procedure and transfers to its connected state to obtain radio resources. In a connected state, UE not only measures signal strength of neighboring cells, but also listens to messages on the control channel and reports the measured channel quality. Most importantly, in a connected state, UE transmits data and monitors all downlink channels. The Narrowband Physical Downlink Control Channel (NPDCCH) has three search spaces, which are used for scheduling general data transmission, RRC procedure related information transmission, and paging information transmission. Consequently, the UE in the connected state consumes more energy than that in the idle state. The state diagram of the UE in an NB-IoT system is shown in Fig. 3, which has three major states, namely Connected, Idle, and PSM, and two sub-states, i.e., Uplink and Downlink. T3324 expires

PSM T3412 expires or UL Packet Generated

Idle eDRX Ɵmer expires or UL Packet Generatd

RRC InacƟvity Timer expires or DL data unavailable

Connected UL Data Transfer

Uplink

DL Data Transfer

UL Data Transfer Available

DL Data Transfer Available

Downlink

Fig. 3. State diagram of the UE in an NB-IoT system.

IoT Device Power Management Based on PSM and EDRX Mechanisms

249

As shown in Fig. 3, when UE stays in the connected state and does not receive downlink or uplink data, the RRC Inactivity Timer will be activated. The Inactivity Timer represents the duration for the UE to trigger T3324 and T3412 times after receiving NPDCCH data. That is when no NPDCCH data is received, the RRC Inactivity Timer will start. If NPDDCH data is received during this period, the RRC Inactivity Timer will reset to zero and recalculate. When RRC Inactivity Timer expires, the UE will turn to its idle state, in which, as abovementioned, the UE only receives paging data. In the idle state, if no data needs to be transmitted before T3324 timer expires, the UE turns to its PSM state. Except T3412 timer, the UE in PSM state does not receive or transmit data. When T3412 time is up, UE returns to its connected state, and the RRC Inactivity Timer is also reset. As shown in the bottom half of Fig. 3, when UE stays in the connected state and needs to transmit data, it switches to the Uplink or Downlink sub-state. Thus, according to this state diagram, we define three cases of UE’s operating cycle, and assume that each cycle starts from idle state. The following explains these cases. Case 1: Idle  Connected  Idle When UE receives uplink or downlink data before T3324 timer expires, it switches itself to the connected state. The probability is defined as PC1 = 1 − e−(λUL +λDL )·Tidle ,

(1)

where λUL and λDL are Poisson arrive rates of the uplink and downlink messages, respectively, and Tidle is the period that UE stays in idle state which is equal to the period of T3324 timer. In case 1, the total time that UE stays in idle state is defined as TC1 Idle , which is equals to   C1 C1 C1 , (2) TIdle = α1 · TeDRX · PDL_i + α2 · TeDRX · PUL_i C1 where α1 and α2 are constants, TeDRX is eDRX cycle time, and PC1 UL_i and PDL_i are the occurence probabilities of uplink and downlink messages in case 1 idle state, respectively. Thus, the energy consumption in idle state of case 1 is    C1 C1 C1 . = β · (EP + α1 · TeDRX · Pwidle ) · PDL + (EP + α2 · TeDRX · Pwidle ) · PUL EIdle

(3) where β is a constant, EP is the energy consumption of paging operation, and Pwidle is the power consumption when UE stays in its Idle state. C1 In connected state of case 1, assuming that TRRC is the RRC inactivity timer and PDL_c C1 and PUL_c are the probabilities of downlink and uplink messages in case 1 connected state, respectively, the time and energy consumption are 1 − e−(λUL +λDL )·TRRC (λUL + λDL ) · e−(λUL +λDL )·TRRC   C1 C1 = PDL_c · EDL + PUL_c · EUL · e−(λUL +λDL )·TRRC

C1 TConnect = C1 EConnect

C1 + (1 − e−(λUL +λDL )·TRRC ) · EC + PConnect · EC

(4)

.

(5)

250

K.-L. Tsai et al.

According to Eq. (2) to Eq. (5), the total time and energy spent in case 1 are C1 C1 TC1 = TIdle + TConnect

(6)

C1 C1 and EC1 = EIdle + EConnect .

(7)

Case 2: Idle  PSM  Connected  Idle If UE does not receive uplink or downlink data before the T3324 timer expires, case 2 will occur. The probability of occurrence is   PC2 = e−λUL ·Tidle · 1 − e−(λUL +λDL )·(T3412 −Tidle ) (8) where T3412 represents the PSM timer, and Tidle is the period that UE stays in its idle state. The energy consumption of idle state in case 2 is defined as C2 EIdle = η · EP + Tidle · EIdle

(9)

in which, η is the number of paging cycles, EP represents the energy consumption of paging operation, and EIdle is the energy consumption of idle state in case 2. The uplink packet arrival rate in PSM state is λUL · e−λUL ·t 1 − e−λUL ·(T3412 −Tidle ) Accordingly, the total time and energy spent in case 2 are  T3412 −Tidle C2 TC2 = Tidle + t · PPSM (t)dt + TConn PPSM (t) =

 C2 EC2 = Eidle +

0

(10)

(11)

0

T3412 −Tidle

C2 t · PPSM (t)dt · EPSM + EConn

(12)

C2 represents the where EPSM is the energy consumption of UE in PSM mode, and EConn energy consumption of UE in connected state of case 2. Case 3: Idle  PSM  Idle If UE does not receive uplink or downlink data before the expiration of T3324 and T3412, case 3 will occur. The probability of occurrence is

PC3 = e(λUL +λDL )T3412

(13)

the total time and energy spent in case 3 are TC3 = T3412

(14)

EC3 = η · EP + Tidle · Eidle + (T3412 − Tidle ) · EPSM + EP

(15)

Combining the above three cases, when given a period of time Ttotal , the UE’s total energy consumption Etotal can be formularized as Etotal = PC1 · EC1 + PC2 · EC2 + PC3 · EC3

(16)

Ttotal = PC1 · TC1 + PC2 · TC2 + PC3 · TC3 .

(17)

and

IoT Device Power Management Based on PSM and EDRX Mechanisms

251

4 Simulation Result In order to investigate the trade-off between energy consumption and transmission efficiency, the simulation is performed by using MATLAB with LTE module. The related parameters are listed in Table 1. Table 1. Simulation parameters. Parameter

Symbol

Value

PSM state power consumption

EPSM

10.8 nW

Idle state power consumption

Eidle

21.6 nW

Energy needed to transmit a UL packet

EUL

735.4 uW

Energy needed to transmit a DL packet

EDL

153.8 uW

Paging state power consumption

EP

35.5 uW

Energy per unit time during connected state

EConn

246.3 uW

RRC inactivity Timer

TRRC

10 s

Number of paging cycles

η = Tidle /TeDRX

1

The impact of T3412 timer and uplink (UL) frequency on energy consumption is shown in Fig. 4. As shown, the smaller the T3412, the more the energy consumption, i.e., the longer the time in the PSM, the less energy consumption. As the UL data interval increases (i.e., the UL data frequency decreases), the average energy consumption decreases accordingly, and the impact of the T3412 timer is greater.

Fig. 4. The impact of T3412 timer and uplink frequency on energy consumption.

Further considering the impact of T3412 and eDRX on energy consumption, the results are shown in Fig. 5. It can be seen that the longer the T3412 duration and the longer

252

K.-L. Tsai et al.

the eDRX duration, the lower the energy consumption. However, compared with the T3412 duration, the eDRX has no significant effect in energy saving, and the difference between 10 s and 100 s is only 0.3 J/h.

Fig. 5. The impact of T3412 and eDRX on energy consumption.

5 Conclusion and Future Studies In this paper, the PSM and eDRX based energy model is discussed, and MATLAB simulation tool with LTE module is also utilized to analyze and simulate energy saving mechanisms under different transmission requirements. The results show that for an IoT system with small amount and long-distance transmission, appropriately configuring the T3412 timers greatly reduces UE’s energy consumption. In the near future, we would investigate other power saving mechanisms for IoT related communication protocols so that the IoT user can easily control the energy consumption of whole IoT evnironment. Besides, the edge computing effect on IoT power saving will also be studied. These constitute our future studies.

References 1. Cerwall, P., et al.: Ericsson mobility report, technique report published by ericsson, June 2020. https://www.ericsson.com/en/mobility-report/reports/june-2020. Accessed 01 Aug 2020 2. IEEE 802.11 Wireless Local Area Networks. http://www.ieee802.org/11Accessed 03 Aug 2020 3. Bluetooth. https://www.bluetooth.com. Accessed 03 Aug 2020 4. Zigbee. https://zigbeealliance.orgAccessed 03 Aug 2020 5. Lora-alliance. https://www.lora-alliance.org. Accessed 03 Aug 2020 6. Grant, S.: 3GPP low power wide area technologies - GSMA White Paper, gsma.com. GSMA: 49, September 2016 7. GPP TS 23.682 V15.0.0. Architecture Enhancements to Facilitate Communications with Packet Data Networks and Applications, March 2017

IoT Device Power Management Based on PSM and EDRX Mechanisms

253

8. GPP TS 36.321 V14.2.0. Evolved Universal Terrestrial Radio Access (E-UTRA), Medium Access Control (MAC) Protocol Specification, March 2017 9. MATLAB. https://www.mathworks.com/products/matlab.html. Accessed 03 Aug 2020 10. Oh, S.M., Shin, J.: An efficient small data transmission scheme in the 3GPP NB-IoT system. IEEE Commun. Lett. 21(3), 660–663 (2017) 11. Lauridsen, M., Krigslund1, R., Rohr, M., Maduenon, G.: An empirical NB-IoT power consumption model for battery lifetime estimation. In: Proceedings of 2018 IEEE 87th Vehicular Technology Conference (VTC Spring), June 2018 12. Yeoh, C.Y., Bin Man, A., Ashraf, Q.M., Samingan, A.K.: Experimental assessment of battery lifetime for commercial off-the-shelf NB-IoT module. In: Proceedings of 2018 20th International Conference on Advanced Communication Technology (ICACT), February 2018 13. Oh, S.M., Jung, K.R., Bae, M., Shin, J.: Performance analysis for the battery consumption of the 3GPP NB-IoT device. In: Proceedings of 2017 International Conference on Information and Communication Technology Convergence (ICTC), October 2017 14. Bello, H., Jian, X., Wei, Y., Chen, M.: Energy-delay evaluation and optimization for NB-IoT PSM with periodic uplink reporting. IEEE Access 7, 3074–3081 (2018) 15. Sultania, A.K., Zand, P., Blondia, C., Famaey, J.: Energy modeling and evaluation of NBIoT with PSM and eDRX. In: Proceedings of IEEE Global Communications Conference, December 2018

Combining Agile with Traditional Software Development for Improvement Maintenance Efficiency and Quality Sen-Tarng Lai1(B) and Fang-Yie Leu2 1 Department of Information Technology and Management, Shih Chien University,

Taipei 10462, Taiwan [email protected] 2 Department of Computer Science, Tunghai University, Taichung 40704, Taiwan [email protected]

Abstract. The continuous changes environment and the rapid evolution of technology make that information system needs continuously and effectively maintain. The research report of the Standish group showed that the success rate of agile projects far exceeds the traditional projects. However, the agile software project has ignored many maintenance issues and characteristics. This paper combines the agile software development advantages and the maintenance benefits of traditional development, and proposes the CMQM model and CMQI process to strengthen the four critical quality characteristics which are documentation quality, test management quality, CM quality, and continuous operations quality characteristics. High critical quality characteristics can concretely improve the efficiency and quality of maintenance operations, and effectively extend the life of information systems.

1 Introduction According to the IEEE 1219 [1], a definition of software maintenance is the modification of a software product after delivery. Delivered information system must have high efficiency and high quality maintenance operations in order to meet the requirements of changing environment and rapid evolution technology. The online banking system provides the transaction services that are not restricted by time and region, greatly improves the service efficiency and transaction convenience. However, for keeping good services quality, the online banking need continuous maintains, and no service can be provided in maintenance period. The maintenance operations always cause extremely inconvenient and impact the market competitiveness. In the age of network and information, various industries have applied information systems to improve their service quality and enhance customer’s market competitiveness. These information systems also face maintenance issues. In information systems development, injecting the critical maintenance quality characteristics can effectively improve the maintainability of the system, and improve the efficiency and quality of maintenance operations and reduce use impact.

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 254–264, 2021. https://doi.org/10.1007/978-3-030-61108-8_25

Combining Agile with Traditional Software Development

255

Traditional waterfall development is a methodology that takes great importance to documentation and process quality, which can effectively improve the maintainability of software system [2]. However, the waterfall development method ignores the personal interaction, continuous communicates and refactoring ability. Cause the waterfall method inability to accept requirement changes in any time, and it is difficult to make timely maintenance in line with environmental changes and technological evolution, resulting in high failure rates for projects of all sizes. Agile method attaches importance to the interaction and communications between the stakeholders and the developers. In addition, continuously release workable software and accept timely feedback from users, and accept the requirement changes in any time. Agile methods can timely refactor in accordance with environmental changes and technological evolution. According to the research report of the Standish Group [3], the success rate of waterfall projects is much lower than that of agile projects. Therefore, in recent years, many software development projects have adopted agile methods to increase the success rate of the projects. However, the agile project neglects many maintenance issues and quality characteristics to make the system existing the high maintenance risk. This paper combines the development advantages of agile methods with the maintenance benefits of the waterfall method are used to improve maintenance operations efficiency and quality, and to extend the life cycle of the information system. Based on maintenance evaluation factors, this paper proposes the Critical Maintenance Quality Measurement (CMQM) model and designs the Critical Maintenance Quality Improvement (CMQI) process to increase the critical maintenance quality of software systems, improve the efficiency and quality of maintenance operations, and reduce use impact. Section 2 discusses the maintenance requirements and explain the importance of software maintenance. Section 3 combines of the development advantages of agile methods and the maintenance characteristics of waterfall method, and lists four critical maintenance quality characteristics. Section 4 proposes the CMQM model based on linear combination. Section 5 based on the CMQM model to design the CMQI process. Section 6 re-emphasizes the importance of maintenance and describes contributions of this paper. And, make a conclusion on this topic.

2 The Importance of Software Maintenance 2.1 Three Successful Factors of Software Maintenance Post-delivery maintenance is a necessary mission for the continued operation of the software system. Unless the system has been retired, maintenance is inevitable. Current information system (ex. online banking system) provides that operates around the clock (24 h), regardless of region, and does not require personnel intervention, which greatly improves service quality and user convenience, and increase the customers values and market competitively. However, the online banking system in Taiwan, system maintenance notification and several hours temporary stop services are announced about one month a time. It is very inconvenience for the customers, and may affects the customers service quality and market competitive. How to evaluate the effectiveness of maintenance operations to increase user satisfaction is the important maintenance issues. According to the maintenance values, the

256

S.-T. Lai and F.-Y. Leu

maintenance requirements, and the maintenance goals of enterprises and organizations, the maintenance operations effectiveness are collected and analyzed. In this paper, we define three important factors which are high efficiency, high quality and reduce use impact to evaluate the maintenance operations effectiveness (shown as Fig. 1). High efficiency maintenance operations can improve information system usability and convenience, and increase the enterprise and organization market competitive. High quality maintenance operations can improve information system trust ability and acceptability, and increase the enterprise and organization reputation and good impression. Reduce use impact intent to short system stop services time and to reduce frequency of the system maintenance to increase the enterprises and organizations uninterrupted service and reliable image. High efficiency

Tree Successful Factors of Software Maintenance

High quality

Reduce use impact

Fig. 1. Three successful factors of software maintenance

2.2 Agile Project Development Advantages and Maintenance Risk Agile software development improves many shortcomings of traditional software development, which effectively increases the success rate of agile projects [4, 5]. The agile manifesto emphasizes the four established values [6]. These values also drive many advantages of agile software development [4]: • The agile development team is composed of people with different skills. From the beginning to the end of the project, the team members are completely dedicated to the project’s cross-functional team, and are empowered to manage the team. Through daily meetings and work together achieve the goals. • Face-to-face interaction and communication are simpler, faster and more reliable than detailed documents. • Agile methods values collaboration with customers. Customers become part of the development process. Customers are also members of the development team. Customer feedback is the key to develop the right system. • Adopt top down evaluation and rolling style, adjust the priority of the work list according to the value of the customer or enterprise. Agile project accepts changes in requirements at any time, even in the later stages of development.

Combining Agile with Traditional Software Development

257

• Adopt Iterative and Incremental Development (IID) and CI/CD processes [5], and continue to deliver valuable software as early as possible to meet customer values and improve market competitiveness. From software maintenance operations, discuss the maintenance risk brought about by the four values of the agile manifesto: (1) Lack of detailed documents: The maintenance personnel is not a team member of agile project and has not participated in the development work. They must understand the software system architecture and critical attributes through detailed development documents. It is difficult for maintenance personnel without correct, complete and consistent development documents to effectively and timely complete various maintenance missions in accordance with maintenance requirements. (2) Not adopting the CM (Configuration Management) process [7] and version control (VC) tools: Agile projects ignore a standardized CM process and VC tools. Therefore, agile project lacks complete of change records and version difference contents. Unsound system architecture, incomplete change information and bad version control tools, it is difficult for maintenance personnel to confirm the differences between the versions, to obtain revision contents and document versions for maintenance operations. (3) Not adopting formal test procedures: For maintenance operations, any changes, expansions and adjustments of maintenance items need to undergo complete testing, including unit testing, integration testing, feature testing, acceptance testing and regression testing. It is impossible without complete test procedure, automatic test process and test cases management to have high quality maintenance operations. In addition, lack of automation testing, it requires a lot of manpower and time to take manual testing, making it difficult for maintenance personnel to achieve high efficiency maintenance operations. (4) The maintenance risk of CI/CD: In maintenance operations, for reducing use impact, continuous integration and continuous deployment process should be adjusted. The iterative period and frequency must be adjusted to shorten the continuous processes and accelerate the deployment time.

3 Critical Quality Characteristics of Software Maintenance 3.1 Major Maintenance Advantages of Waterfall Methodology The waterfall methodology is an important software development method that was proposed in the 1970s [8]. It almost became the standard development methodology for software projects in follow 20 years. The new released methods almost are based on the waterfall method, and do not make major adjustments [2, 8]. The development standard and rigorous processes of waterfall could not cope with the environmental changes and technological evolution, and caused the high failure rate of software project. In 2001, the Agile Software Alliance was established by 17 software professional and the Agile Manifesto was issued [9]. The Agile Software Alliance subverted multiple development

258

S.-T. Lai and F.-Y. Leu

concepts of waterfall methodology. However, waterfall methodology have many processes and activities that inject the critical maintenance quality characteristics for maintenance operations. The critical quality characteristics are the maintenance advantages of waterfall methodology, describe as follows: (1) Documentation advantages: Waterfall model is oriented on plan and emphasizes the phase development approach. Each development phase must complete detailed phase documents in accordance with development specifications. Detailed phase documents are the basis to the interaction and communication of developers at all phases, as well as a important resource to improve maintenance operations efficiency and quality. (2) CM and VC advantages: Waterfall development method attaches great importance to the CM system and VC process. The CM workflow specification can record the complete system architecture, the revision record of the version, identify the differences between the versions, establish the traceability and cross-reference relationship between the document and the source code [7]. In the features addition, expansion and adjustment maintenance requests, CM and VC can effectively improve the maintenance operations efficiency. (3) Test management advantages: In order to avoid the propagation and expansion of errors and defects, the waterfall development method adopts a rigorous review of phase documents, timely identifies the errors and defects of the documents, and immediately takes correction manners. For controlling the software quality, source codes are required to adopt the test process of V model that from unit test to acceptance test. Using test management procedure effectively manage the test cases, steps, process, results and related items and build a basis of automation test. Test management procedure is key resource to improve the maintenance operations efficiency and quality. 3.2 Critical Maintenance Quality Characteristics In order to reduce the maintenance risk of agile projects, this paper lists documentation, CM, test management and continuous operations four critical quality characteristics that must be improved and strengthened. Four critical quality characteristics are highly correlated with maintenance evaluation factors (shown as Table 1). Each critical quality characteristic is composed of related quality factors, and each quality factor has its corresponding quality items. The multi-layer quality characteristics architecture (shown as Fig. 2) can effectively identify quality problems and deficiencies by quality measurement, and can concretely find the related quality activities. The following describes the relationship between the quality factors that constitute the four critical quality characteristics, and the quality items that affect the quality factors: • Documentation quality characteristic is combined correctness, completeness and consistency quality factors: – Correctness quality factor should consider the correctness of user stories, features, tasks and the related documentation quality items.

Combining Agile with Traditional Software Development

259

– Completeness quality factor should consider the completeness of user stories, features, tasks and the related documentation quality items. – Consistency quality factor should consider the consistency of user stories, features, tasks and the related documentation quality items. • CM quality characteristic is combined CM process, VC guidelines and VC tools quality factors: – CM process quality factor should consider the quality items of specification, availability and completeness. – VC guidelines quality factor should consider the quality items of operation flow, availability and completeness. – VC tools quality factors should consider the quality items of basic functions, identifiability and availability. Table 1. Quality characteristics and maintenance evaluation factors relation table Successful factors

Maintenance efficiency

Maintenance quality

Reduce use impact

Quality characteristics Documentation CM Test Management Continuous Operations : High influential

: Influential

• Test Management quality characteristic is combined test processes, test case management and automation test quality factors: – Test processes quality factor should consider the quality items of specification, availability and completeness. – Test case management quality factor should consider the quality items of access functions, availability and identifiability. – Automation test quality factor should consider the quality items of testing tools, features and completeness. • Continuous operations quality characteristic is combined CI process, CD process and CI/CD pipeline quality factors: – CI process quality factor should consider the quality items of guidelines, availability and completeness. – CD process quality factor should consider the quality items of guidelines, availability and completeness.

260

S.-T. Lai and F.-Y. Leu

– CI/CD pipeline quality factor [10] should consider the quality items of stages planning, availability and automation. Software professionals and experienced maintainers can establish quantitative values of basic quality items through inspection and review activities. Combining the related quantified quality items, the quality factors of quality characteristics can be produced. The quality factors is a basis of CMQM model. Critical Maintenance Quality Documentation Quality Characteristic Correctness Quality Factor Completeness Quality Factor Consistency Quality Factor CM Quality Characteristic CM System Quality Factor VC Tool Quality Factor VC Process Quality Factor Test Management Quality Characteristic. Test Processes Quality Factor -layer quality measurement model TC Management Quality Factor quality problems Automation Test Quality Factor improvements and corrective measures for Continuous Characteristic quality itemsOperations and qualityQuality activities. ality CI Process Quality Factor CD Process Quality Factor CI/CD Pipeline Quality Factor

Fig. 2. Multi-layer critical quality characteristics architecture

4 Critical Maintenance Quality Measurement Model Based on the linear combination model [11], this paper proposes a CMQM model that includes four quality characteristics measurements and one CMQM. The senior software engineers assign the weighted value between 0 and 1. The weighted value close to 1 indicates that the quality item is important for quality measurement. The CMQM model describes as follows:

Combining Agile with Traditional Software Development

261

(1) Documentation quality characteristic need consider the documentation correctness, completeness and correctness three quality factors. For this, Documentation Quality characteristic Measurement (DQM) is combined correctness, completeness and consistence three documentation quality factors. Combination formula shows as Eq. (1):

(1) (2) CM quality characteristic need consider CM process, VC guidelines and VC tools three quality factors. For this, CM Quality characteristic Measurement (CMQM) is combined CM process, VC guidelines and VC process quality factors. Combination formula shows as Eq. (2):

(2) (3) Test management quality characteristic need consider test processes, test case management and automation test three quality items. For this, Test Management Quality characteristic Measurement (TMQM) is combined test processes, test case management and automation test quality factors. Combination formula shows as Eq. (3):

(3) (4) Continuous operations quality characteristic need consider CI process, CD process and CI/CD pipeline three quality factors. For this, Continuous Operations Quality characteristic Measurement (COQM) is combined CI process, CD process, and CI/CD pipeline quality factors. Combination formula shows as Eq. (4):

(4)

262

S.-T. Lai and F.-Y. Leu

(5) Critical Maintenance Quality Measurement (CMQM) is combined four critical quality characteristics measurements which include documentation quality, CM quality, test management quality, and continuous operations quality characteristics. And according to the influence of quality characteristics, the weighted value of linear combination measurements are assigned. Combination formula shown as Eq. (5):

(5)

5 Critical Maintenance Quality Improvement Process Based on the multi-layer linear combination model, CMQM is composed of four maintenance quality characteristics, and each quality characteristic measurement is composed of several quality factors. The multi-layer quality measurement model can effectively identify quality problems and defects, and find out bad quality items, and then concretely plan improvements and corrective measures for quality items. Before the quality measurement, the acceptable range of quality measurement must be set in advance. Once the up-layer quality measurement isn’t in acceptable range, combination equations of CMQM model and the quality factors of the quality items need be analyzed. And identify the bad quality items by a top-down manner, and propose improvement activities and corrective measures for quality items. In addition, the required missions that need to be completed for each maintenance requirements are different. Therefore, in maintenance process, it is necessary to set an acceptable measurement range for the quality characteristics and maintenance quality measurements. For critical maintenance quality measurement, inspection, improvement and corrective measures must be repeated until the measurement of critical maintenance quality meet the acceptable range. Critical Maintenance Quality Improvement (CMQI) process is composed of two improvement procedures. Quality Characteristics Improvement procedure has five steps, describes as follows: 1. According to the quantified quality factor to produce the quality characteristic measurement. 2. IF all critical quality characteristics have been measured, and the measurement are within the acceptable range THEN enter procedure. 3. IF the measurement is within the acceptable range THEN go to Step 1 for next quality characteristic measurement. ELSE the bad quality factors need be identified, and the defects of quality items need be modified. 4. Plan the quality correction and improvement manners for defect quality items. 5. Re-quantify the modified quality items and go back to Step 1

Combining Agile with Traditional Software Development

263

Quality Improvement procedure has four steps, describes as follows: 1. Combine the four critical quality characteristics measurements to produce critical maintenance quality measurement. 2. IF the measurement is within the acceptable range THEN end of CMQI process. ELSE the bad quality characteristics must be identified, and the acceptable range of bad quality characteristics need be adjusted. 3. Arrange the adjustment activities of quality characteristic measurement acceptable range. 4. Enter procedure. The advantages of the CMQI process are described as follows: • The multi-layer measurement model can effectively identify the bad quality items, and propose improvement activities and corrective manners. • Apply iteratively the quality measurement, identification of quality defects, and improvement activities and corrective manners. • The weighted value of the CMQM model and the acceptable range of measurement can be adjusted in time based on feedbacks, maintenance requirements and maintenance missions.

6 Conclusion In order to improve service performance and get market competitive advantages for enterprises and organizations, the service items of the information system must be continuously new added, adjusted and expanded. The efficiency, quality and reduce use impact of software maintenance are major factors for enhancing values and market competitiveness of enterprises and organizations. Agile software projects have a high success rate, but there are also high maintenance risk. In order to adapt to the continuous changes in the environment and the rapid growth of technology, the software system must inject the critical quality characteristics. This paper combines the development advantages of agile methods with the maintenance benefits of the waterfall method to improve maintenance operations efficiency, quality and reduce use impact. Based on the CMQM Model, the paper designs the CMQI process to identify and improve quality items problems and defects for enhance critical quality characteristics. The contributions of this paper are described as follows: • Multi-layer critical maintenance quality characteristics architecture can concretely identify problems and defects in quality items. • CMOI process can effectively improve the efficiency and quality of maintenance operations and reduce use impact. • High maintenance quality characteristics are sufficient to cope with the continuous changes of the environment and rapid technological evolution. • Agile projects can have a high success rate and own well maintenance results.

264

S.-T. Lai and F.-Y. Leu

References 1. Edelstein, D.V.: Report on the IEEE STD 1219–1993—standard for software maintenance. ACM SIGSOFT Softw. Eng. Notes 18(4), 94–95 (1993) 2. Munassar, N.M.A., Govardhan, A.: A comparison between five models of software engineering. Int. J. Comput. Sci. Issues (IJCSI) 7(5), 94 (2010) 3. Standish Group: The CHAOS report. 2015 (2015) 4. Sliger, M., Broderick, S.: The Software Project Manager’s Bridge to Agility. Addison-Wesley Professional, Boston (2008) 5. Edeki, C.: Agile software development methodology. Eur. J. Math. Comput. Sci. 2(1) (2015) 6. Beck, K., et al.: Manifesto for agile software development. http://www.agilemanifesto.org/ 7. Conradi, R., Westfechtel, B.: Version models for software configuration management. ACM Comput. Surv. (CSUR) 30(2), 232–282 (1998) 8. Royce, W.W.: Managing the development of large software systems. In: Proceedings of IEEE WESCON, Los Angeles, no. 8, pp. 328–338 (1970) 9. Schach, S.R.: Object-Oriented and Classical Software Engineering, 8th edn. McGraw-Hill, New York (2011) 10. Wolf, J., Yoon, S: Automated testing for continuous delivery pipelines. In: Pacific NW Software Quality Conference in Industrial Talk (2016) 11. Fenton, N.E.: Software Metrics - A Rigorous Approach. Chapman & Hall, London (1991)

On Text Tiling for Documents: A Neural-Network Approach Siang Yun Yoong1 , Yao-Chung Fan1(B) , and Fang-Yie Leu2 1

National Chung Hsing University, Taichung, Taiwan [email protected] 2 TungHai University, Taichung, Taiwan

Abstract. Segmenting documents or conversation threads into semantically coherent segments have been one of the challenging tasks in natural language processing. In this work, we introduce three new text segmentation models that employ BERT for post-training. Extensive experiments are conducted based on benchmark datasets to demonstrate that our BERT-based models show significant improvements over the state-ofthe-art text segmentation algorithms.

1

Introduction

Text Segmentation [8] task that focuses on dividing a document or conversation thread into semantically coherent segments (topically similar units) is a fundamental text processing task and also a basic building block for text mining applications, such as text summarization [12], passage extraction [13], etc. In this paper, we explore the adaption of the BERT model [6] for the text segmentation problem. We propose to reformulate the text segmentation problem as a binary classification problem. Specifically, for a given document consisting of n sentences, our goal is to have a classification model to predict if there is a topic changing point between two adjacent sentences. Under such a reformulation, we propose three models based on BERT. Our first model called BERT-NSP is a naive BERT employment. We directly reuse the Next Sentence Prediction (NSP) task (the pre-trained setting) to have a baseline for the BERT employment. The next sentence prediction task is to predict if a given sentence is the next sentence in some context of another given sentence. Such a task is similar to our formulation. Therefore, we employ the architecture of NSP task as the first model for our text segmentation problem. However, we find that BERT-NSP considers only the information between adjacent sentences. We argue that the information between sentence pair may not be sufficient to determine a topic change in a document. In fact, the whole information in a document should be considered. Consequently, we propose the other two BERT models: BERT-SEP and BERT-SEGMENT. Both models are designed to consider the whole document information. In BERT-SEP, we mainly leverage sentence-level representation to detect topic change points. And, in c The Editor(s) (if applicable) and The Author(s), under exclusive license  to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 265–274, 2021. https://doi.org/10.1007/978-3-030-61108-8_26

266

S. Y. Yoong et al.

BERT-SEGMENT, we leverage token-level representation to detect topic change points. By validating on the benchmark datasets, we demonstrate our models outperform the existing models and advance the state-of-the-art performance on the benchmark datasets. The contributions of this paper are as follows. – To our best knowledge, this work is the first one to investigate the power of BERT on the text segmentation task. – We propose three novel BERT-based model for text segmentation. – Extensive experiments are conducted using benchmark datasets to show the effectiveness of the proposed models.

2

Related Work

The existing methods for segmenting documents or conversation thread into segments can be classified into supervised methods and unsupervised methods. For unsupervised method, [7] introduced a graph-based segmentation algorithm called Graphseg. The idea is to construct a semantic relatedness graph where the nodes are sentences and the edges indicate the semantic relatedness between nodes. After the graph construction, the Bron-Kerbosch [3] algorithm is employed to acquire the maximal cliques for extracting text segments. In Graphseg, edges between nodes are determined by a given threshold. However, the setting of the threshold is not a trivial task. Furthermore, authors [14] proposed a text segmentation algorithm called TopicTiling. The TopicTiling algorithm is a variation of the well-known TextTiling algorithm [8], which used Latent Dirichlet Allocation (LDA) [2] to obtain topic IDs for two adjacent blocks of sentences, and then use cosine similarity to compute a coherence score between the blocks. For supervised method, [10] indicates that the existing datasets for text segmentation are too small to enable supervised models. As a result, they generated a new dataset called Wiki-727K that consists of 727,726 English Wikipedia documents based on the hierarchical structure of Wikipedia. They construct a neural model composed of two sub-networks with each containing two layers of bidirectional LSTM [9] and use the Wiki-727K as training data. They first employ the lower-level sub-network to generate sentence representations and feed representations into the higher-level sub-network respectively to predict a topic-changing point probability for each of the sentence. [11] proposed a neural model called SEGBOT which employed a bidirectional recurrent neural network to encode sentence representation. The authors first use bi-directional GRU (BiGRU) [5] to capture the representation of sentence, and then pass the sentence representation into another GRU-based hidden layer with a pointer network to compute a distribution over all possible positions in the input sequence to decide whether there is a topic-changing boundary. Although existing works mainly employ recurrent neural networks to obtain the representation of the sentences or words, it requires significant training time

On Text Tiling for Documents: A Neural-Network Approach

267

as recurrent neural network needs to be processed sequentially and some information of the words may be missed when processing longer sequence. For solving the problems mentioned above, [15] proposes a neural network called Transformer, a novel the architectures of encoder-decoder with self-attention. Unlike the sequential nature of recurrent neural network, Transformer can be trained in a parallel manner by employing the self-attention modules. Thus, to leverage the power of Transformer, We employ BERT [6] which is a stack of multi-layer transformer, for addressing the text segmentation task.

3 3.1

Preliminaries Problem Definition

Text segmentation aims to divide a document or conversation thread into semantically coherent segments (topically similar units). Specifically, given a document d that consists of m sentences d = (s1 , s2 , ..., sm ), the objective is to find consecutive sentences with the same topic. In other word, the goal to find if there is a topic changing between two consecutive sentences si and si+1 , ∀i ∈ 1, ..., m − 1. 3.2

BERT Overview

The BERT model is a multi-layer bidirectional Transformer encoder [15]. For BERT’s model, the input is required to aligned as the BERT’s specific input sequence. First, an input sequence is formed by a token sequence that tokenizes the input by using WordPiece embeddings [16] with a vocabulary of 30,000 tokens. Then, a special token [CLS] is inserted as the first token of every sequence as it represents a token for the classification application in the original paper. Single text sentence or sentence pairs can both be represented as a sequence of tokens in BERT’s input representation. To distinguish the sentence pairs, a special token [SEP] is added between the sentences. In addition, a learned embedding is added to every token to denote whether it belongs to sentence A or sentence B. For example, a sentence pair (si , sj ) where si contains |si | tokens and sj contains |sj | tokens, the BERT input can be defined as following: x = ([CLS], ti,1 , ..., ti,|si | , [SEP], tj,1 ..., tj,|sj | ) The input representation of a given token is the sum of three embeddings: the token embeddings, the segmentation embeddings and the position embeddings. Then the input representation can feed forward into extra layers to perform a fine-tuning procedure.

4

Models

In this subsection, we introduce three models based on BERT: BERT-NSP, BERT-SEP, and BERT-SEGMENT. The details of our models are described in the following subsections.

268

4.1

S. Y. Yoong et al.

BERT-NSP

We experiment with the NSP task (the BERT’s pre-training setting) as a naive BERT employment. The goal of NSP task is to detect the contextual relatedness between two sentences. The idea is to divide the sentences in a document into sentence pairs and train a model to understand the relationship between two sentences. If two consecutive sentences are predicted as non-contextual relevant, we report there is a topic change. Thus, we employ the NSP task to fine-tune BERT model to acquire a sentence relatedness probability for each sentence pair. Given a document d contains n sentences d = (s1 , s2 , ..., sn ), we divide the adjacent sentences into n − 1 pairs of sentence. For each sentence pair (si , sj ), the BERT-NSP is applied as follows. First, an input sequence is formulated as x = ([CLS], si , [SEP], sj , [SEP]) Note that each sentence si is also a sequence of tokens ti,j . Specifically, si = [ti,1 , ..., ti,|si | ], where |si | denotes the length of si . Let BERT() be the BERT model. By BERT(x), we obtain

(V[CLS] , Vti,1 , ..., Vti,|si | , V[SEP] , Vtj,1 , ..., Vtj,|sj | V[SEP] )

Then, we take the hidden state vector of the [CLS] token V[CLS] ∈ Rh where h denotes as the hidden layer size. We add an affine layer W ∈ Rh×2 to the output of the [CLS] token. We compute the sentence relatedness probabilities P r(r|si , sj ) ∈ R2 by a softmax function as follows. P r(r|si , sj ) = sof tmax(V[CLS] · W + b) 4.2

BERT-SEP

In this model, we consider that the [SEP] token of the input sequence not only can be represented the token that separates different sentences, it can also be employed as the token that represent sentence relatedness between two adjacent sentences. The basic idea behind BERT-SEP is to leverage the representation of the [SEP] token in an input sequence of BERT; we think that the representation of the [SEP] token can be employed to compute relatedness between two adjacent sentences. When formulating the input sequence, we add [SEP] token to separate all sentences in a document. We aim to tag each token [SEP] in the input sequence with a label, where 1 indicates a topic change point, whereas 0 indicates not (Fig. 1). Our BERT-SEP works as follows. Given a document d contains n sentences where d = (s1 , s2 , ..., sn ), we apply BERT-SEP as below: First, an input sequence is formulated as x = ([CLS], s1 , [SEP], s2 , [SEP], ..., [SEP], sn )

On Text Tiling for Documents: A Neural-Network Approach

269

Fig. 1. The BERT-SEP architecture

Note that each sentence si is also a sequence of tokens ti,j . Specifically, si = [ti,1 , ..., ti,|si | ] Let BERT() be the BERT model. By BERT(x), we obtain (V[CLS] , Vti,1 , ..., Vti,|si | , V[SEP] , ..., V[SEP] , Vtn,1 , ..., Vtn,|si | V[SEP] ) Then, x is represented by the BERT embedding layers and then passed forward into BERT’s transformer blocks. We take the hidden vectors H[SEP] in R|x|×h where |x| is the number of [SEP] in input sequence. Then we compute the topic changing probabilities P r(r|sL ) for each sentence representation by a softmax function P r(r|sL ) = sof tmax(HL · WL + bL ). 4.3

BERT-SEGMENT

The BERT model can be also employed in token or word-level prediction tasks. Therefore, we propose to reformulate the text segmentation problem as a wordtagging task. Namely, we design to tag each of the token with a label. If the sentence is the last sentence of a segment, we label each token within the sentence as 1, where 1 denotes as the topic-changing point, while the other tokens of the sentence is labeled as 0. Figure 2 shows the architecture of BERT-SEGMENT. Our BERT-SEGMENT works as follows. Given a document d contains n sentences where d = (s1 , s2 , ..., sn ), we apply BERT-SEGMENT as below: First, an input sequence is formulated as x = ([CLS], s1 , [SEP], s2 , [SEP], ..., [SEP], sn ) Note that each sentence si is also a sequence of tokens ti,j . Specifically, si = [ti,1 , ..., ti,|si | ] Let BERT() be the BERT model. By BERT(x), we obtain (V[CLS] , Vti,1 , ..., Vti,|si | , V[SEP] , ..., V[SEP] , Vtn,1 , ..., Vtn,|si | V[SEP] )

270

S. Y. Yoong et al.

Fig. 2. The BERT-SEGMENT architecture

Then, x is represented by the BERT embedding layers and then passed forward into BERT’s transformer blocks. We take the hidden vectors Hb in R|x|×h where |x| is the length of the input sequence. After obtain the hidden vectors, we pass the hidden vectors through an additional layer to compute the topic-changing probabilities. Same as BERT-SEP, we experiment with three type of segmentation layers to fine-tuned with BERT. The three types of layers are described in the following subsections: Then we compute the topic changing probabilities P r(r|sL ) for each token representation by a softmax function P r(r|sL ) = sof tmax(HL · WL + bL ). Sentence Score Calculation. As the output is only the topic-changing probability of each of the tokens, we obtain the sentence score of a sentence by aggregating the topic-changing probability of every tokens in a sentence and averaging the score as the final sentence score. The sentence score Sc of a sentence consisting of n tokens is formulated as follows, while P r denotes the topic-changing probability Sc = (

5 5.1



|s |

i [P ri,j ]j=1 ). |si |

Experiments Datasets

We evaluate the proposed three BERT-based models on three benchmarking datasets, which are Wiki-727 datasets (small), Wiki-727 datasets (large) and Elements, where each of the datasets is based on the hierarchical structure of Wikipedia. – Wiki-727 datasets (large) [10] divide Wiki-727 dataset into 80% of training data, 10% of testing data and 10% of validation data. This dataset contains approximately 72775 documents from testing data.

On Text Tiling for Documents: A Neural-Network Approach

271

– Wiki-727 datasets (small) As some compared models take long time to run the evaluation on the Wiki-727 datasets (large), [10] also provided a small set of testing data from the testing data of Wiki-727 dataset, which contains only 50 documents. – Elements Elements is introduced by [4]. The dataset is derived from Wikipedia and contains 118 articles about chemical elements in the periodic table, with topics such as Biological Role, Occurrence and Isotopes. 5.2

Compared Models

We compare our models with previously proposed methods in the literature. The compared models in the experiment are: – Graphseg: An unsupervised algorithm proposed by [7] that works by building a graph, where the nodes denote the sentences of the documents and the edges is added if two sentences are semantically similar. The algorithm BronKerbosch [3] is employed for finding the maximal cliques of adjacent sentences and heuristically completing the segmentation. – TextSeg: [10] proposed a model that is constructed by a hierarchical of two sub-networks both based on bi-directional LSTM architecture [9]. First, the representation of the sentences are acquired through the lower-level subnetwork and then pass the representation of the sentences to the higher-level sub-network to predict whether the sentence is a topic-changing point or not. – BERT-NSP: The model introduced in Subsect. 4.1 which using a simple fullyconnected layer as a segmentation layer. The input of the model is a pair of sentences, while the output is the sentence relatedness probability. – BERT-SEP: The model introduced in Subsect. 4.2. We experiment with three variants of the different segmentation layer, BERT-SEP (simple linear), BERT-SEP-RNN (RNN), and BERT-SEP-Transformer (Transformer). – BERT-SEGMENT: The model introduced in Subsect. 4.3. We also experiment with three variants of the different segmentation layer, BERT-SEGMENT (simple linear), BERT-SEGMENT-RNN (RNN), and BERT-SEGMENTTransformer (Transformer). 5.3

Training Settings

Our BERT-based models: BERT-NSP, BERT-SEP and BERT-SEGMENT are trained by using the PyTorch version of BERT1 . For the use of the pre-training model, we employed the officially provided BERTbase with a vocabulary size of 30522. We complete the training of the models on two NVIDIA TITAN RTX GPUs on an Intel Core i7 machine with 128 GB memory. We use Adam as our optimization method, while the learning rate is initially set as 3e−5 and decayed by a linear warm scheme. We set the maximum sequence length of the input sequence up to 512 tokens. 1

https://github.com/huggingface/pytorch-pretrained-BERT.

272

S. Y. Yoong et al. Table 1. Comparison of Pk score on benchmarking datasets Method

Wiki727-small Wiki727-large Elements

Graphseg

63.56

–

49.12

BERT-NSP

44.24

42.68

36.17

Textseg

18.24

22.13

41.63

Bert-SEP

15.25

19.12

32.88

Bert-SEP-RNN

14.48

19.14

33.09

Bert-SEP-Transformer

13.92

19.01

31.11

Bert-SEGMENT

16.21

19.59

35.45

Bert-SEGMENT-RNN

16.40

19.73

33.70

Bert-SEGMENT-Transformer 15.91

19.72

34.52

Table 2. Comparison of Recall (R), Precision (P) and F1 score on Wiki727 datasets (Small) Method

Recall of class 0

Recall of class 1

Precision of class 0

Precision of class 1

F1 score of class 0

F1 score of class 1

Bert-SEP

0.98

0.61

0.95

0.83

0.96

0.70

Bert-SEP-RNN

0.98

0.66

0.95

0.83

0.97

0.73

Bert-SEP-Transformer

0.98

0.63

0.95

0.83

0.96

0.71

Bert-SEGMENT

0.98

0.62

0.95

0.79

0.96

0.69

Bert-SEGMENT-RNN

0.97

0.63

0.95

0.76

0.96

0.69

Bert-SEGMENT-Transformer 0.98

0.58

0.94

0.81

0.96

0.68

Table 3. Comparison of Recall (R), Precision (P) and F1 score on Wiki727 datasets (Large) Method

Recall of Recall of Precision of Precision of F1 score of F1 score of class 0 class 1 class 0 class 1 class 0 class 1

Bert-SEP

0.98

0.65

0.96

0.80

0.97

0.72

Bert-SEP-RNN

0.98

0.65

0.96

0.79

0.97

0.72

Bert-SEP-Transformer

0.98

0.66

0.96

0.78

0.97

0.72

Bert-SEGMENT

0.97

0.68

0.96

0.76

0.97

0.72

Bert-SEGMENT-RNN

0.98

0.65

0.96

0.78

0.97

0.71

Bert-SEGMENT-Transformer 0.99

0.60

0.95

0.85

0.97

0.70

Table 4. Comparison of Recall (R), Precision (P) and F1 score on dataset Elements Method

Recall of Recall of Precision of Precision of F1 score of F1 score of class 0 class 1 class 0 class 1 class 0 class 1

Bert-SEP

0.94

0.48

0.81

0.77

0.94

0.48

Bert-SEP-RNN

0.95

0.46

0.81

0.79

0.95

0.46

Bert-SEP-Transformer

0.92

0.53

0.82

0.73

0.92

0.53

Bert-SEGMENT

0.94

0.49

0.81

0.79

0.87

0.61

Bert-SEGMENT-RNN

0.95

0.45

0.80

0.79

0.87

0.57

Bert-SEGMENT-Transformer 0.94

0.47

0.80

0.77

0.87

0.58

On Text Tiling for Documents: A Neural-Network Approach

5.4

273

Evaluation Results

Performance Overview. We compared our models with two previously proposed methods: Graphseg and TextSe.g. We use Pk metric [1] for performance evaluation. The evaluation results are summarized in Table 1. The testing Wiki727K dataset-Large is too large for evaluation on Graphseg as it is computationally demanding, so the result is not shown in the table. From the result shown, our BERT-NSP outperforms Graphseg but underperforms TextSe.g. We speculate the reason is that that BERT-NSP considers only the information of sentence pair, while TextSeg consumes the information of a whole document to detect topic-changing points and therefore successfully segment the documents than the result by BERT-NSP. Alos, we can see that both BERT-SEP and BERT-SEGMENT outperform TextSeg, while the result of BERT-SEP slightly performs better than BERT-SEGMENT. Also, we can see that BERT-SEPTransformer shows best result among the models. Meanwhile, we find that the results of BERT with RNN layer and BERT with Transformer Layer does not affect too much comparing to BERT with Linear Layer. In addition, we calculate the recall score, precision score and F1 score for all the compared models on the datasets, which are summarized in Table 2, Table 3 and Table 4.

6

Conclusion

In this work, we explore the employment of the pre-trained BERT language model for text segmentation task. Three BERT-based models, BERT-NSP, BERT-SEP and BERT-SEGMENT, are introduced. We demonstrate that the word-tagging formulation and considering whole document information is key to the text segmentation performance. From the experiment evaluation, our BERTbased models outperform than state-of-the-art methods on the evaluation of the benchmarking datasets.

References 1. Beeferman, D., Berger, A., Lafferty, J.: Statistical models for text segmentation. Mach. Learn. 34(1–3), 177–210 (1999) 2. Blei, D.M., Ng, A.Y., Jordan, M.I.: Latent dirichlet allocation. J. Mach. Learn. Res. 3(Jan), 993–1022 (2003) 3. Bron, C., Kerbosch, J.: Algorithm 457: finding all cliques of an undirected graph. Commun. ACM 16(9), 575–577 (1973) 4. Chen, H., Branavan, S., Barzilay, R., Karger, D.R.: Global models of document structure using latent permutations. In: Proceedings of Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the Association for Computational Linguistics, pp. 371–379. Association for Computational Linguistics (2009) 5. Cho, K., Van Merri¨enboer, B., Gulcehre, C., Bahdanau, D., Bougares, F., Schwenk, H., Bengio, Y.: Learning phrase representations using RNN encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078 (2014)

274

S. Y. Yoong et al.

6. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: BERT: pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) 7. Glavaˇs, G., Nanni, F., Ponzetto, S.P.: Unsupervised text segmentation using semantic relatedness graphs. Association for Computational Linguistics (2016) 8. Hearst, M.A.: TextTiling: segmenting text into multi-paragraph subtopic passages. Comput. Linguist. 23(1), 33–64 (1997) 9. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Comput. 9(8), 1735–1780 (1997) 10. Koshorek, O., Cohen, A., Mor, N., Rotman, M., Berant, J.: Text segmentation as a supervised learning task. arXiv preprint arXiv:1803.09337 (2018) 11. Li, J., Sun, A., Joty, S.R.: SegBot: a generic neural text segmentation model with pointer network. In: IJCAI, pp. 4166–4172 (2018) 12. Mitrat, M., Singhal, A., Buckleytt, C.: Automatic text summarization by paragraph extraction. In: Intelligent Scalable Text Summarization (1997) 13. Oh, H.J., Myaeng, S.H., Jang, M.G.: Semantic passage segmentation based on sentence topics for question answering. Inf. Sci. 177(18), 3696–3717 (2007) 14. Riedl, M., Biemann, C.: Topictiling: a text segmentation algorithm based on LDA. In: Proceedings of ACL 2012 Student Research Workshop, pp. 37–42. Association for Computational Linguistics (2012) 15. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L  ., Polosukhin, I.: Attention is all you need. In: Advances in Neural Information Processing Systems, pp. 5998–6008 (2017) 16. Wu, Y., Schuster, M., Chen, Z., Le, Q.V., Norouzi, M., Macherey, W., Krikun, M., Cao, Y., Gao, Q., Macherey, K., et al.: Google’s neural machine translation system: bridging the gap between human and machine translation. arXiv preprint arXiv:1609.08144 (2016)

A High Sensing Accuracy Mechanism for Wireless Sensor Networks Li-Ling Hung1(B) and Fang-Yie Leu2 1 Aletheia University, New Taipei, Taiwan

[email protected] 2 Tunghai University, Taichung, Taiwan

[email protected]

Abstract. The wireless sensor networks have lots of popular applications for years. The efficiency of a wireless sensor network is subject to the monitoring accuracy and limited energy. Thus, event detection and energy efficiency are two important issues in Wireless Sensor Networks. In order to overcome the limit of yield rate for sensor hardware, some studies built other monitoring system to support monitoring for enhancing monitoring accuracy. According to limited energy, some studies designed energy efficient mechanisms to extend the lifetime of monitoring. In this paper, we propose a distributed cooperative mechanism that neighboring sensors mutual confirm the event occurrence for improving the monitoring accuracy and reducing the total numbers and energy consumption of transmissions. The result of simulations reveals that using the proposed mechanism, the monitoring lifetime is extended and the monitoring accuracy of proposed mechanism is much better than other existed mechanisms.

1 Introduction Wireless sensor networks (WSNs) have wide range of applications, including military, surveillance, environmental monitoring, and health care, proposed in the literatures [1– 3]. In addition, the technique of wireless sensor networks is the base of internet of things applications [4] which is applied much more in recent years. Hence, the issues of WSN are much more important than they were before. In WSNs, the objective of monitoring is to detect something abnormal which named as events. It is assumed that corresponding sinks should be notified about events. Regarding the efficiency of WSNs, event detection is a key and basic issue. In order to improve event detections, some researchers propose supporting another detecting system to verify or further reveal the event occurrence [5, 6]. In WSNs, the energy consumption for transmitting is much more than that of sensing and receiving. If each sensor sends out the abnormal situation and the abnormal is detected or transmitted by more than one sensor, many redundant packets of the same flow in the network forming the packet storm which wastes a great amount of energy for unnecessary redundant transmissions. Some researchers propose saving the energy by avoiding duplicate transmissions at event notification [8, 10–12]. © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 275–283, 2021. https://doi.org/10.1007/978-3-030-61108-8_27

276

L.-L. Hung and F.-Y. Leu

In this paper, we propose a High Sensing Accuracy Mechanism for applications of Wireless Sensor Network, which is named HSAM for short. The mechanism employs larger sensing range to overlap the monitoring region in the environment. When an event occurs, more than two sensors mutual confirm the occurrence of events and it will be sent by one of the sensors which detected it. Thus, the monitoring accuracy is enhanced and the transmission redundancy is reduced.

2 Related Studies Because the energy in WSN is often limited due to portable batteries, many studies pay more attention to deal with the energy efficient issue for prolong the monitoring lifetime of the WSNs. Rani et al. [9] proposed an energy efficient scheme for IoTs by using low power for intra-cluster communication and high power for inter-cluster communication. The communication mechanism reduces not only the energy consumption for communications but also the interference of communications among neighboring sensors. We propose that sensors employ different power for different targets. The sensors employ lower power for checking event with near neighbors, and they employ high power for transmitting checked events. Reddy et al. [8], Moosavi et al. [10], Marco et al. [11], and Cavalcante et al. [12] proposed saving the energy by avoiding duplicate transmissions for event notification. However, the duplicate transmissions sometime can be used to verify when some sensors do not detect events for some unpredictable mistakes. Faillettaz et al. [5, 6] and Meyer et al. [7] designed co-detection mechanisms to monitor the occurrence of important events by supporting another monitoring system. Our proposed mechanism eliminates both duplicate transmissions and undetected mistakes by means of checking events with neighbors before notifying them. In addition, only one sensor from those detected neighbor launches the event notification. Our study aims at announcing events to their sinks correctly and efficiently. To announce events correctly, the detected situation will be confirmed among neighbors before announcement. To announce confirmed events efficiently, we employ the longer transmission distance to reduce the times and consumption of transmissions.

3 The Proposed Mechanism This section introduces the high sensing accuracy mechanism for the sensor networks. Because the energy consumption for transmitting is much more than that for sensing and receiving of sensors, the sensors do not transmit all the sensing data to sinks, they transmit only the abnormal situations, or named events. The transmission policy can reduce a large number of transmissions and corresponding energy consumption. If an abnormal detected, the abnormal will be mutual checked among neighboring sensors. When more than two sensors have detected the same abnormal, the abnormal is assumed to be an event and should be transmitted to inform corresponding sinks. The following describes the details for network environment, error checking and event transmissions.

A High Sensing Accuracy Mechanism for Wireless Sensor Networks

277

3.1 Network Environment The monitored area is divided into n equal-size hexagons, named cells in this work. Each cell has six edges with length of r. In addition, a set of n static sensors S = {s1 , s2 ,…, sn } are deployed in the center of these cells. For ease of presentation, we define three axes which divide the area into 6 regions in the monitored environment, the center of the environment is set by coordinate (0, 0, 0). The horizontal line crosses the original point, (0, 0, 0), is named Z-axis; the Z-axis turns 60° in a counter-clockwise direction to be the Y-axis; and the Y-axis turns 60° in a counter-clockwise direction to be the X-axis. The location of each cell is represented by a coordinate (x, y, z). The points on the right of the original point have positive values with regard to the axis and the distance unit is 3r/2. By the definitions, the points on the X-axis, Y-axis, and Z-axis have the same characteristics as y + z = 0, x − z = 0, and x + y = 0, respectively. Figure 1 shows the axes example in an environment and partial of those coordinates.

Fig. 1. The coordinates for part of the environment.

While monitoring the environment, if any abnormal situation is detected by sensors, the sensors will notify corresponding sink of the abnormal situation for dealing with it. We assume that the sensors know the locations of corresponding sinks. The entire area of the environment must be monitored by some sensors at each time period. If the energy of some sensors is exhausted and causing some region unmonitored or abnormal situations without notified, the monitoring network does not work well, or said that the lifetime of sensor network ends. √ Each sensor is used for sensing √ and checking events in the distance of 3r. Because the sensing range of the sensors is 3r, any region in the environment is monitored by more than 2 sensors. Each cell is monitored by 7 sensors located around, as shown in Fig. 2, the regions colored in yellow ink are monitored by 3 neighboring sensors and the regions colored in green ink are monitored by 4 neighboring sensors. The monitoring

278

L.-L. Hung and F.-Y. Leu

Fig. 2. A cell is monitored by the sensors which are located in the center of neighboring cells.

accuracy is improved by means of mutual checking the abnormal among neighboring sensors in the√environment. On the other hand, the transmission distance for event transmission is 2 3r. 3.2 The Communications for Confirming Errors This period aims to confirm the occurrence of events. To reduce the energy consumption of transmissions, only confirmed events are transmitted to corresponding sinks in announcing periods. The sensor networks do not work well if there exits any event which is not notified to the sinks. In order to improve the accuracy of monitoring, the neighboring sensors communicate for√ confirming the occurrence of an event. As mentioned in Sect. 2, the sensing range is 3r and each cell has six edges with the length of r in the environment. The entire area in the environment must be monitored by three or four sensors. If one sensor does not detect abnormal situations for some unexpected reason, there still have two or more sensors will detect the anomaly. When √ sensors find an anomaly, they send an anomaly message to one-step neighbors, with 3r transmitting range. To avoid the collision of messages, the sensors send the anomaly message after a time period ts which derived by Eq. (1), where Di,j is the distance between sensor si and the nearest corresponding sink k j , Dmax is the longest distance in the environment; E i and E full are the residual energy of si and the full energy of each sensor, respectively; tp is the a unit waiting time period. Following (1), the sensors with shorter distance away from the corresponding sinks and having more energy will have smaller ts to wait and send the message earlier. After receiving one or more anomaly warnings including detected by themselves, the sensors confirm the occurrence of events. In addition, the event communicating sensor which sent the message first will transmit the event announcement in the following announcing period because it would be the most appropriate one. After communication, the other sensors detected the same anomaly are sure that the event would not be ignored or missed because the event will be announced by some neighbor. ts = (Di,j /Dmax )/(Ei /Efull ) × tp

(1)

A High Sensing Accuracy Mechanism for Wireless Sensor Networks

279

3.3 The Transmission for Notifying Sinks of Errors After confirming, the sensors transmit the confirmed events forward to corresponding sinks. In order to reduce the √ transmission steps and energy consumption, the transmission distance of sensors is 2 3r. In other words, the 2-step neighbors away from the transmitting sensor will receive the messages. For example, the sensors located at points (0, 2, 2), (1, −1, 2), (2, 2, 0) are 2-step neighbors of sensor located at point (0, 0, 0). When receiving an event packet, the sensor may relay the packet to corresponding sink through the neighboring sensors. To avoid interfering and reduce the energy consumption, the sensors relay the event packet after a time period which also derived by Eq. (1). In addition, when the distance between the receiving sensor and corresponding sink is greater than the distance between the original sending sensor, the receiving sensor ignore that packet because the relaying sensor should be located in the other side. Moreover, when the neighboring sensors, in the right side, hear the relaying message, they will abandon the relaying job because there exits the other neighboring sensor is more appropriate to relay that event. However, because the communication range is larger, the number of interfered sensors is larger. If the event occurrence rate is low, the transmissions will not collide frequently. When the event occurrence rate is higher, the interfering situation becomes serious.

4 Evaluations This section evaluates the performance of proposed mechanism against MLEACH [13] and EERH [14] in terms of the energy efficiency of sensors and WSN. In addition, the performance of proposed mechanism against EERH and CoDet [5] in terms of the monitoring accuracy is also evaluated. MLEACH is a cluster-based transmission mechanism for monitoring system. MLEACH elects cluster heads according to the residual energy of sensors. In addition, these head sensors transmit the statuses of sensors periodically. EERH shares the statuses of sensors by means of piggybacking instead of announcing periodically. Sensors using EERH do not negotiate before transmit packets, the time of event transmissions is derived according to the residual energy of sensors and their neighbors as well as the distance away from the sinks. In these simulations, the time duration of a round is set to 2 s. The sizes of a checking message and an event packet are 0.2 Kbits and 1 Kbits, respectively. For fair of comparison, the total amount of energy should be the same. Because of different deployment, the distance for sensing and transmission of EERH is 10 ms and 20 ms, respectively; the distance for sensing and transmission of MLEACH is 10 ms and 30 ms, respectively. In addition, the distances √ of √ HSAM for sensing √ abnormal, checking events and transmitting events are 10 3 ms, 10 3 ms and 20 3 ms, respectively. Moreover, the event detection performance of mechanisms is evaluated according to different fault ratios of sensors. Because the sensing range of neighboring sensors will overlap in some area. When an error occurs at some overlapping area, if one of monitoring sensors failures to detect it, the other sensor(s) monitoring the area may detect the error. Hence, we simulate the missing ratio for event detection in different error rates of sensors for mechanisms EERH, CoDet, and HSAM. CoDet is a mechanism which cooperates the sensor system with another supporting scheme [10]. The CoDet

280

L.-L. Hung and F.-Y. Leu

is designed to help confirm the detail of events. In order to avoid energy consumed for unimportant transmissions, CoDet redesign the sensors which are controlled by the supporting scheme. Only when supporting system senses abnormal, can the sensor system transmit event notification. The parameters in the simulations are listed in Table 1. Sensors know the geographical transmission distance of an event packet. The amount of amplified energy is determined by the distance between transmitter and receiver. For example, the amplifier of a sensor consumes energy 10 nJ × 1000 × 102 to transmit a packet of size 1 Kbits for a 10 ms distance transmission. Table 1. Simulation parameters. Parameter

Value

Size of network field

200 m × 200 m

Packet size

1 K bits/pkt

Initial total energy for sensors

72 J

Energy consumed for election

50 nJ/bit

Energy consumed for packet transmission/receiving 50 nJ/bit Energy consumed for sensing

0.045 mJ/s

Energy consumed by the amplifier

10 nJ/bit/m2

Event occurrence ratio

10%

When event occurrence rate is 10%, the simulation results on energy consumption according to the working time and the number of events are shown in Figs. 3 and 4. For the cluster-based schemes, such as MLEACH, the steps of packet transmission can be reduced. Because the transmission distance is longer, the cluster heads consume much more energy than that required by the sensor using other schemes for delivering packets. Moreover, the cluster-based schemes consume more energy for statuses exchange and heads election. Therefore, the average residual energy of MLEACH is much less than that of EERH and HSAM when the numbers of rounds are the same. Although sensors in HSAM confirm events before transmit them and they transmit the events in longer distance than those using MLEACH or EERH, the number of influenced sensors for each transmission using HSAM is much less than that using EERH. The reason is that energy consumption is not only for the sensors transmitting but also the sensors receiving. Following HSAM design, sensors adjust the directional antenna for transmitting and receiving. Hence, the receiving sensors only consume the energy for the events it should receive. Therefore, the energy consumption using HSAM is less than that using EERH. Generally, when the total residual energy is less than 2 J, there exist many sensors exhausting their energy. Thus some events cannot be sent out for notification because the routes to sinks are disconnected. Therefore, the monitoring lifetime of the WSN ends and the total residual energy of sensors does not decrease obviously. In addition, according to the number of notified events, simulation results for energy consumption are shown in Fig. 4. Compare with Fig. 3, as to each event notification,

A High Sensing Accuracy Mechanism for Wireless Sensor Networks

281

Fig. 3. The energy consumed by sensors according to the monitoring time.

the energy consumption for sensors using MLEACH and HSAM is less than that for sensors using EERH because the times of transmission for EERH is much more than that of MLEACH and HSAM. Although the times of transmission for MLEACH is less than that for HSAM, because the number of interfered sensors and the related energy consumption using MLEACH is much more, the total energy consumption for transmitting and receiving for MLEACH is more than that for HSAM. Therefore, from the simulation results as Fig. 4, energy consumption of HSAM is much less than that of MLEACH and EERH for the same number of event occurrence.

Fig. 4. The energy consumed by sensors according to the number of notified events.

Figure 5 shows the simulation result for missing rates of event detection of mechanisms EERH, HSAM and CoDet in two cases of error rate for supporting system, one

282

L.-L. Hung and F.-Y. Leu

is 0.02 (the line with green color, CoDet1) and the other is 0.04 (the line with cyan color, CoDet2). The result depicts that the more overlapping monitored region, the less the events missed. Therefore, HSAM has the least missing rate of event detections. Although CoDet has the higher overlapping area than EERH, the redesign scheme leads that the error notifications of sensors limited by the supporting system.

Fig. 5. The missing rates of event detection of mechanisms in different error rates of sensors.

5 Conclusions and Future Work This paper proposes a mechanism for improve monitoring efficiency and energy efficiency of WSNs. All the sensors are employed for communications and transmissions. By increasing the overlapping monitored region from sensors, the monitoring efficiency is enhanced because sensors confirm the event occurrence with neighbors. By using different power for communication between different neighbors, the energy efficiency is also improved because it reduces the interference of transmissions. Both the result of evaluations and simulations show that our missing rate of error detection is much less than other mechanisms. Because of longer transmission distance, the routing efficiency is enhanced. Following proposed mechanism, the transmissions are saving lots of energy consumptions by means of reducing communications and fault or duplicate announcements when the event occurrence rate is low. Our future work includes improving the transmission mechanism for overcome the low efficiency when the event occurrence rate is higher. Acknowledgments. This work was supported in part by the Ministry of Science and Technology, Taiwan, under Grant MOST 108-2221-E-156-002, Grant MOST 109-2221-E-156-002 and Grant MOST 109-2221-E-029-017-MY2.

A High Sensing Accuracy Mechanism for Wireless Sensor Networks

283

References 1. Kay, J.M., Frolik, J.: An expedient wireless sensor automaton with system scalability and efficiency benefits. IEEE Trans. Syst. Man Cybern. Part A Syst. Hum. 38(6), 1198–1209 (2008) 2. Erdelj, M., Razafindralambo, T., Simplot-Ryl, D.: Covering points of interest with mobile sensors. IEEE Trans. Parallel Distrib. Syst. 24(1), 32–43 (2013) 3. Wang, Z., Liao, J., Cao, Q., Qi, H., Wang, X.: Achieving k-barrier coverage in hybrid directional sensor networks. IEEE Trans. Mobile Comput. 13(7), 1443–1455 (2014) 4. Awasthi, R., Kaur, N.: Review of energy efficient techniques of IoT. Int. J. Recent Innov. Trends Comput. Commun. 7(5), 17–21 (2019) 5. Faillettaz, J., Or, D., Reiweger, I.: Codetection of acoustic emissions during failure of heterogeneous media: new perspectives for natural hazard early warning. Geophys. Res. Lett. 43, 1075–1083 (2016) 6. Faillettaz, J., Funk, M., Beutel, J., Vieli, A.: Towards early warning of gravitational slope failure with co-detection of microseismic activity: the case of an active rock glacier. Nat. Hazards Earth Sys. Sci. 19(7), 1399–1413 (2019) 7. Meyer, M., Farei-Campagna, T., Pasztor, A., Da Forno, R., Gsell, T., Faillettaz, J., Vieli, A., Weber, S., Beutel, J., Thiele, L.: Event triggered natural hazard monitoring with convolutional neural networks on the edge. In: ACM/IEEE International Conference on Information Processing in Sensor Networks (IPSN) (2019) 8. Reddy, P.K., Babu, R.: An evolutionary secure energy efficient routing protocol in Internet of Things. Int. J. Intell. Eng. Syst. 10(3), 337–346 (2017) 9. Rani, S., Talwar, R., Malhotra, J., Ahmed, S., Sarkar, M., Song, H.: A novel scheme for an energy-efficient Internet of things based on wireless sensor networks. Sensors 15(11), 28603–28626 (2015) 10. Moosavi, S., Gia, T., Nigussie, E.: End-to-end security scheme for mobility enabled healthcare Internet of Things. Future Gener. Comput. Syst. 64(1), 108–124 (2016) 11. Marco, P., Athanasiou, G., Mekikis, P., Fischione, C.: MAC-aware routing metrics for the Internet of Things. Comput. Commun. 74(15), 77–86 (2016) 12. Cavalcante, V., Pereira, J., Pitanga, M., Moura, R., Batista, T., Flavia, C., Paulo, F.: On the interplay of Internet of Things and cloud computing: a systematic mapping study. Comput. Commun. 89(1), 17–33 (2016) 13. Mohammed, I.Y., Elrahim, A.G.A.: Energy efficient routing protocol for heterogeneous wireless sensor networks. SUST J. Eng. Comput. Sci. 20(1), 1–10 (2019) 14. Hung, L.-L., Leu, F.-Y., Tsai, K.-L., Ko, C.-Y.: Energy-efficient cooperative routing scheme for heterogeneous wireless sensor networks. IEEE Access 8, 56321–56332 (2020)

A Novel Scheme of Schnorr Multi-signatures for Multiple Messages with Key Aggregation Rikuhiro Kojima(B) , Dai Yamamoto, Takeshi Shimoyama, Kouichi Yasaki, and Kazuaki Nimura FUJITSU Laboratories Ltd., 4-1-1, Kamikotanaka, Nakahara-ku, Kawasaki, Japan {kojima.rikuhiro,yamamoto.dai,shimo-shimo, yasaki.kouichi,kazuaki.nimura}@fujitsu.com Abstract. A digital signature is essential in verifying the reliability of people and data over networks, such as through web server certificates, authentication, and blockchain technologies. In blockchain, multisignature signature schemes have recently attracted attention for reducing the amount of data in transactions. While such schemes support only a single message, Interactive Aggregate Signatures (IAS), an extended Schnorr multi-signature scheme, supports some messages under the plain public key model. However, there are three problems with this scheme in certain use cases. We propose a key aggregatable IAS scheme called KAIAS. In contrast to the previous works, KAIAS solves these problems which means that KAIAS (1) includes a verification algorithm using only a single aggregated public key, (2) dynamically signature aggregation, and (3) requires signers to sign only their own messages. Recently, the Schnorr multi-signature scheme has been discussed mainly from its advantages of reducing the size of the signatures in the implementation of Bitcoin. Thus, we also propose a practical application of KAIAS that takes advantage of its feature to aggregate both signatures and public keys with low computational complexity of signing.

1

Introduction

Electronic signatures are used as “fingerprints” of data, mainly in identity verification and preventing spoofing from securing. Recent attention has been focused on improving their convenience and functionality. For example, in authentication protocols such as SSH (Secure SHell) [SSH] and FIDO (Fast Identity Online Universal Authentication Framework) [LT17], the client sends a valid signature to the server instead of a password hash. This modification is useful and effective because the client does not need to remember the password or keep any secret information in their secure storage. Multi-Signature. As one of the signature application scheme, Multi-signature schemes have been attracting attention for their convenience and functionality. c The Editor(s) (if applicable) and The Author(s), under exclusive license  to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 284–295, 2021. https://doi.org/10.1007/978-3-030-61108-8_28

A Novel Scheme of Schnorr Multi-signatures for Multiple Messages

285

Multi-signature scheme [Oka88] is a scheme in which n signers (who have their own private key ski and public key pki ) sign a signature σi for each i ∈ {1, · · · , n} in a common message m and output an aggregated single signature σ from σ1 , · · · , σn , which is a static size independent of n. The verifier is given m that is attached with σ with public key list {pk1 , · · · , pkn } (or aggregated single key apk) and confirms the correctness by running the verification algorithm, which outputs accept or reject. Multi-signature schemes are constructed by the extension of representative signature schemes with a security proof, e.g., a Schnorr signature scheme [BN06] or BLS (Boneh-Lynn-Shacham) signature scheme [Bol03]. Also, RSAbased [Oka88], pairing-based [BDN18], and lattice-based [EBS16] have been proposed. Even if we construct a secure model under any assumption, multi-signature schemes are vulnerable to “Rouge-key Attack”, which is caused by registration of the corrupted public key. To avoid this attack, Ristenpart and Yilek proposed a method of adding a proof of possession (PoP) to the public key [RY07], and Bellare and Neven proposed a method of including the individual public key in the hash value of the message [BN06]. As a different concept of signature aggregation in multi-signature schemes, Maxwell et al. proposed a (public) key aggregation method of which a verification algorithm does not require individual public keys but only uses an aggregated single public key [MPSW19]. Recent research on multi-signature schemes has been discussed particularly for Bitcoin application [MPSW19,BDN18]. Aggregate Signature [Bon11] is a scheme in which n parties sign their own messages and print an aggregated single σ (fixed size independent of n) in situations where each party has a single message. In general, by regarding set of messages as a single message, we can translate aggregate signatures into multisignatures. Thus, Aggregate Signature is defined as a generalization of a multisignature scheme; however, this scheme has restrictions, for example, the verifier must check which all messages are signed by each signers before verification, and all messages must differ from each other. As described above, multi-signature schemes only support a signature for only a single message, however, in some use cases, the signers may want to add a new message for signing dynamically and aggregate it. Maxwell et al. proposed to transform multi-signature schemes to secure Interactive Aggregate Signature (IAS), which includes signing the combined messages of m1 , · · · , mn ; however, it does not satisfy (1) the verification algorithm using only a single aggregated public key and (2) dynamic signature aggregation. Boneh et al. proposed the Aggregate Multi-Signatures Protocol (AMSP), which is an aggregation of multisignature schemes, however, (3) all signers need to sign all others’ signatures. We discussed the advantages derived from these specific requirements in more detail in Sect. 2.2. Our Proposal. We propose KAIAS, a novel key aggregatable IAS scheme that does not require additional assumptions in the public-key model and constructed three-round protocols. In this scheme, a signer outputs a single aggregated sig-

286

R. Kojima et al.

nature corresponding to multiple messages. This enables the verifier to verify only using a single aggregated public key. KAIAS satisfies requirements (1), (2), and (3) mentioned above. Compared with IAS and ASMP, KAIAS has a slightly larger aggregated signature (d +  → 2d + ). This is however negligible because it is a static value independent of the number of signers n, so by increasing n, we can gain the benefit of signature aggregation and key aggregation like the other multi-signature schemes. Table 1 shows a comparison of KAIAS, IAS, and ASMP. Table 1. Comparison of the Multi-Signature schemes supporting multiple messages with four evaluation points [1] size of signatures [2] key aggregation [3] dynamic aggregation [4] time of signing, when using a group G and hash functions H where d,  are the bit size of each elements of G and H Schemes

[1]

[2]

[3]

[4]

IAS [MPSW19]

d+

No No 1

AMSP [BDN18]

d+

Yes Yes n

KAIAS (proposal) 2d +  Yes Yes 1

Applications. Pieter Wuille et al. recently proposed a draft of Bitcoin Improvement Proposal (BIP) 340 (Schnorr signatures for secp256k1) [WNR20], which is the standard implementation to improve the Bitcoin protocols [Nak09] using the Schnorr signature scheme. BIP 340 uses the ECSDSA (Elliptic Curve Schnorr Digital Signature Algorithm) has the following three advantages, provable security, non-malleability and linearity, over the ECDSA (Elliptic Curve Digital Signature Algorithm), which is the standard signature protocol in the current Bitcoin. KAIAS can also be applied to Bitcoin applications, however, it increases the demand for the application of Trust Services following eIDAS (Electronic Identification, Authentication and trust Services) regulation [eID14] of the EU. Trust Services are components for determining how to trust others over a network in a situation such as a public service or business, and are composed of an electronic signature, timestamp, e-Seal, etc. [TS118]. The following advantages can be obtained by assigning the properties described above to the use case of Trust Services: • Even if there are many signers involved in approving a document (order of 100, 1000), the verifier can confirm the validity of an aggregate signature with an aggregated public key to verify in the same manner as in conventional signature verification. • It is possible to change the documents dynamically with addition, fix, and erase. • To clarify responsibility for documents, we can restrict the signers to sign only their texts.

A Novel Scheme of Schnorr Multi-signatures for Multiple Messages

287

Although IAS and AMSP only partially satisfy these benefits, KAIAS is the most appropriate. Overview. We first describe the notations used in this paper, discrete logarithm problem, multi-signature schemes, IAS, and AMSP in Sect. 2. We then present KAIAS and discuss its security proof in Sect. 3. Finally, we propose two applications that uses KAIAS in Sect. 4.

2 2.1

Preliminaries Definitions $

− S which means the Notation. Given a non-empty set S, we denote by s ← operation of sampling an element s from S uniformly at random. In all the following, Let G = g be a cyclic group of prime order p, where p is a κ-bit integer, with generator g, and we call (G, p, g) the group parameters. Discrete Logarithm Problem. The provable security of Schnorr Signature [Sch91] is based on the discrete logarithm problem (DLP). Definition 1 (Discrete Logarithm Problem (DLP)). Let (G, p, g) be the group parameter. We define a AdvDL G as $

$

P r[y = g x : y ← − G, x ← − A (y)] which is a probability is taken over the random choose of A and random elements y. An algorithm A is said to (τ, ) -solve DLP if it runs in time at most τ and AdvDL G > . Multi-Signature Scheme. We follows the definition of [BN06] and [BDN18], and we describe the scheme and security of multi-signatures. In general, a MultiSignature scheme MS is constructed by algorithms Pg, Kg, Sign, KAg, Vf defined by the following. Pg(1κ ) Given the security parameter κ, output the system parameters params. Kg(params) Given the params, output a key pair (sk, pk) where sk is a secret key and pk is a public key. Sign(params, Lpk , Lsk , m) Given the (params, Lpk , Lsk , m), where Lpk is the set of public key and Lsk is the set of secret key of n signers, compute the σi ← Signi (params, Lpk , ski , m) where ski ∈ Lsk for each i ∈ {1, · · · , n}, and output a signature σ aggregated by {σ1 · · · , σn }. KAg(params, Lpk ) Given the (params, Lpk ), output a aggregated public key apk. Vf(params, apk, m, σ) Given the (params, apk, m, σ), if σ is the valid signature for m, then output 1, otherwise output 0. This scheme should satisfy completeness and unforgeability.

288

2.2

R. Kojima et al.

Prior Works

MuSig. We follow the definition of Multi-Signature scheme MuSig = (Pg, Kg, Sign, KAg, Vf), which is denoted by Boneh, Drijvers and Neven [BDN18]. This scheme is executed by n signers, a verifier, and a “proxy”. This proxy is a semi-honest entity and has three functions: receive some parameters, computing (all proxy computations are verifiable by the signers), and broadcast some parameters to all signers. Parameter Generation. MuSig.Pg(1κ ) A trusted party generates the group parameters params ← Pg(1κ ), where params = (G, p, g), where κ is a security parameter, p is a κ-bit prime integer, G is cyclic group of order p, and g is a generator of G, and prepares three different hash function H0 , H1 , and H2 which returns non-zero value defined as H0 , H1 , H2 : {0, 1}∗ → Zp Key Generation. MuSig.Kg(params) $

− Zp and corresponding Each i-th signer generates a random parameter xi ← public key Xi = g xi , and let L = {X1 , · · · , Xn } be the multi-set of public keys. Key Aggregation. MuSig.KAg(L)  following steps. This algorithm outputs an aggregated public key X 1. Each i-th signer computes ai ← H1 (Xi , L) and sends ai to the proxy.   = n X ai and outputs X. 2. The proxy compute aggregated key X i=1 i This algorithm outputs a aggregated signature σ following an 3 round interactive protocol. 1. Each i-th signer generates a random integer ri ← Zp and computes Ri ← g ri , ti ← H2 (Ri ) and sends ti to the proxy, after that, the proxy broadcasts {tj }j∈{1,··· ,n} to each signer. 2. Each i-th signer send Ri to the proxy, after that, the proxy broadcasts {Rj }j∈{1,··· ,n} to each signer, and they check that tj = H2 (Rj ) for each j = i.  ← KAg(L) and let ai ← H1 (Xi , L) and broadcasts 3. The proxy compute X  = n X ai .  (X, {aj }j∈{1,··· ,n} ) to each signer, and they check that X i=1 i n  m) After 3-round protocols, the proxy compute R = i=1 Ri , c ← H0 (R, X, and broadcast c for all signers, and each i-th signer computes si = ri + cai xi mod p, where ai← H1 (Xi , L), and sends si to the proxy. Finally, the n proxy computes s ← i=1 si and outputs the final signature σ = (s, R).  Verification. MuSig.Vf(params, m, σ, X) This algorithm outputs 1 or 0 which means the verifier accepts or reject. Given a message m, a aggregated signature σ and a aggregated public key  the verifier computes c ← H0 (R, X,  m), and outputs 1 if and only if X, s c  g = RX , otherwise, outputs 0.

A Novel Scheme of Schnorr Multi-signatures for Multiple Messages

289

We can confirm the following equations to check the correctness. c g s = g s1 +···+sn = R1 X1a1 c · · · Rn Xnan c = RX Interactive Aggregate Signature. Compared with the multi-signature schemes, which include only a signing algorithm for a common message, IAS includes a signing algorithm for different messages. Bellare and Neven [BN06] suggested a generic approach to transform any multi-signature scheme into an IAS. Maxwell et al. [MPSW19] then proposed fixed secure IAS schemes. In IAS, instead of a public key list L = {X1 , · · · , Xn } and a single message m, signers and a verifier use an ordered set of public key/message pairs S = {(X1 , m1 ), · · · , (Xn , mn )}, so this transforms the multi-signature scheme MuSig defined above into IAS. Aggregate Multi-Signature Protocol. To aggregate signatures corresponding to different messages without using IAS, Boneh et al. introduced ASMP, which is a scheme that outputs a signature aggregated by multi-signature schemes [BDN18]. Namely, this scheme uses signature aggregation of aggregation. They also proposed a pairing-based approach to construct AMSP. Not only pairing, but this AMSP aggregation technique also allows to be applied to all multi-signatures whose aggregation process is homomorphic, i.e., the aggregated signature is verifiable using the normal verification algorithm. 2.3

Issues

Fundamentally, IAS is an advanced and extended multi-signature scheme; however, there are two points to be improved in terms of practical use. First, the verifier needs all of the individual public keys because it cannot verify with an aggregated public key. In MuSig.Vf, the verifier confirms the bind ing of a message and a signature by using only a single aggregated public key X, which is generated by all signer’s public keys; thus, the verifier does not require a set of public keys for verifying. With IAS, however, because the message hash  i) contains the message index i to prohibit cheating in which value ci ← H0 (S, X, any signer copies another signer’s signature, c1 , · · · , cn are all different and the verifier cannot use the aggregated public key for verifying. Second, the signers cannot generate a new aggregated signature using an already aggregated one. The ci , which is the message hash value in IAS, must be generated by {m1 , · · · , mn } ∈ S, which are static signer’s n messages before signing, unlike “original” aggregated signatures. Thus, we cannot aggregate dynamically from an existing σ and a new signature corresponding to mn+1 signed by any signer. If we want to do so, all the signers have to re-compute the hash value, which includes a new set S ∪ mn+1 , re-sign, and re-aggregate a signature. This is a non-negligible disadvantage considering the use cases such as documentation version control, in which any signer aggregates the signatures repeatedly. Alternatively, Schnorr-based AMSP satisfies the following condition that the verifier requires only aggregated a public key for verifying, and the signer can aggregate a new multi-signature to the old one if the new message is added; thus,

290

R. Kojima et al.

AMSP is considered an improved IAS. To construct multi-signatures, however, each i-th signer must generate a signature σi (mj ), which is a signature of mj signed by xi for each j ∈ {1, · · · , n}, which means that all signers must sign the message of other signers, which increases the cost of signing for every signer by n times and that of the total signature space to the square of n. From the above discussions, to the best of our knowledge, no multi-signature scheme has been proposed that satisfies the following conditions: (1) including the verification algorithm using only a single aggregated public key, (2) dynamically aggregating signatures, and (3) requiring signers to sign only their own messages.

3 3.1

Our Proposal Description

To consider applying more practical use cases, we propose a new IAS that satisfies all of (1), (2) and (3) under the same assumptions introduced by Maxwell et al. In prior IAS, the verifier needed the public key list L = {X1 , · · · , Xn } instead  to confirm the equation g s = R n X ci . of an aggregated public key X i=1 i As the improvement of it, note that we introduce the message hash aggregation.  as well as the aggregation of signac ← KAIAS.MHAg({m1 , · · · , mn }, R, X) tures and public keys. Also, we embed the additional parameter {di }i∈{1,··· ,n} in an aggregated signature σ , which is the “wormhole” multiplication of c1 , · · · , cn except ci . Using di , we can construct di · σi = (di si , di Ri ) as a pseudo signature of m1 , · · · , mn signed by a secret key xi , where σi = (di si , di Ri ) is a signature of mi signed by xi , after that, we can use usual signature aggregation technique.  to In consequence of this modification, the verifier can verify using only X s  c   confirm the equation g = RX . Now we introduce the new Key Aggregatable IAS KAIAS = (Pg, Kg, KAg, Sign, AVf). Hash Function. H0 Redefine H0 as H0 : {0, 1}∗ → Zp − {0}. Parameter gen. KAIAS.Pg(1κ ) Same as MuSig.Pg(1κ ). Key gen. KAIAS.Kg(params) Same as MuSig.Kg(params). Key Aggregation. KAIAS.KAg(L) Same as MuSig.KAg(L). Signing. KAIAS.Sign(params, {m1 , · · · mn }, {sk1 , · · · , skn }) This algorithm outputs a aggregated signature σ . After 3-round protocols in MuSig.Sign, each i-th signer computes R=

n 

 mi ), si = ri + ci ai xi mod p Ri , ci ← H0 (R, X,

i=1

where ai ← H1 (Xi , L), and sends (mi , si , ci ) to the proxy. Finally, the proxy  mi ) and computes checks that ci ← H0 (R, X, s =

n  i=1

= si di , R

n  i=1

Ridi

A Novel Scheme of Schnorr Multi-signatures for Multiple Messages

291

where {di }i∈{1,··· ,n} is defined as 

di =

cj = c1 · · · ci−1 ci+1 · · · cn

j∈{1,··· ,n}−{i}

 R). for each i, and outputs the signature σ  = ( s, R,  Message Aggregation. KAIAS.MHAg({m1 , · · · , mn }, R, X) This algorithm outputs a aggregated message hash value c.  mi ) for each i. 1. Compute ci ←H0 (R, X, n 2. Compute c = i=1 ci and outputs c.  , X) Aggregate Signature Verification. KAIAS.AVf(params, {m1 , · · · , mn }, σ This algorithm outputs 1 or 0 which means the verifier accepts or  = reject. Given a message {m1 , · · · , mn }, a aggregated signature σ   ( s, R, R) and a aggregated public key X, the verifier computes c ←  and outputs 1 if and only if g s = R X  c, KAIAS.MHAg({m1 , · · · , mn }, R, X), otherwise, outputs 0. Note that c = ci di for each i ∈ {1, · · · , n}, we can confirm the following equations to confirm the correctness. g s = g s1 d1 +···+sn dn = g r1 d1 +c1 d1 a1 x1 +···+rn dn +cn dn an xn X c = Rd1 · · · Rdn (X a1 · · · X an )c = R 1

N

1

n

This scheme should satisfy completeness, too. KAIAS.AVf(params, {m1 , · · · , mn }, KAIAS.Sign(params, {m1 , · · · mn }, {sk1 , · · · , skn }), KAIAS.KAg(L)) = 1. 3.2

Security Proof

Following the security proof of [MPSW19] and [BDN18], we use the double forking technique. Theorem 1. KAIAS is an unforgeable Multi-Signature scheme (as defined in Definition 2.1) in the random-oracle model if the DLP is hard. In other words, we can construct (τD , D )-solve DLP D by using KAIAS (τ, qS , qH , n, )-forger algorithm F of and τD = 4τ + 4nτexp + O(nqT ), D ≥

3 4 16nqT2 −  − 3 qT p 2

where qT = qS + qH + 1 and τexp is the time needed to compute the aggregated public key apk at most.

292

R. Kojima et al.

Sketch of Proof. We first construct an algorithm A which takes y as input and outputs a forgery multi-signature using forger algorithm F . we then construct an algorithm B takes y as input, and output an aggregated public key apk which includes y and its discrete logarithm ω, using generalized forking algorithm F ork A on the algorithm A which outputs two forgery multi-signatures which generated by same inputs. Finally, we construct a discreate-logarithm algorithm D takes y as input, and output its discrete logarithm ω, using the outputs of F ork B on the algorithm B, that means if we define the (τ, qS , qH , )-forger algorithm F , we can construct the DLP-solver D. 3.3

Implementation

We have experimented with toy-example implementation of KAIAS in Python 3.7.0, whose environment is Microsoft Windows 10 Pro 64bit, Intel Core i56500 3.20 GHz 3.19 GHz, RAM 8.00 GB in Table 2. Although our main work is not aimed at improving execution time, the result that there is no significant difference between the [2] and [3] is very important from the viewpoint of the practicality which delegates the computing cost required by the verifier to the signer. It is our future works that the experiments comparing with other implementation methods and open-source publishing. Table 2. Comparison of execution time in 100, 1000, 10000 signatures (all units are [ms]) with four evaluation points [1] time of signing [2] time of signature aggregation [3] time of verification [4] time of verification of a aggregated signature. Num of Sig [1] 10 100 1000

4

[2]

[3]

[4]

343

609

671 62

3,204

6,554

6,317 62

33,292 65,187 63,671 62

Proposed Application

In this section, we introduce our electronic signature application that uses KAIAS and can be applied in business. The eIDAS is an EU regulation, and Trust Services are digital signature components that are expected to be widely used for both corporate and private customers in the future. Trust Services are constructed using signature-application technologies such as time-stamps, e-Seal (the signature issued by a corporation), and e-Delivery. These services enable the users to verify not only the completeness of data but reliability, which contains the evidence of when signed and who signed. After publishing eIDAS in the EU in 2014 [eID14], the criteria defined by eIDs (Electronic IDentification) and Trust Services were recognized. Therefore, it is expected that Trust Services will spread not only throughout the EU but

A Novel Scheme of Schnorr Multi-signatures for Multiple Messages

293

also in Japan and the US to achieve the Digital Single Market envisioned by the EU. In the research on multi-signature schemes, it has been actively discussed how to reduce the transaction data size using the multi-signature application in Bitcoin and other transaction technologies. However, there has been less discussion on how to establish the trust of the signer’s data using the same schemes. We introduce the following electronic signature application for effectively applying KAIAS that is not only for Bitcoin. 4.1

Contract During the Companies

We introduce an electronic signature application that is effective in a situation involving a contract between corporations X and Y (X Corp. and Y Corp., respectively) using an electronic signature. X Corp. (the signer side) issues a contract document with an e-Seal, which is a signature that can be used to verify the correctness of the contract. Before signing with e-Seal, this document had been revised by approvers (the co-signers) who attached the signature as a sign of approval. In response, Y Corp. (the verifier side) can confirm the validity of the document by verifying its e-Seal by using a public key corresponding to it. In the above use case, the e-Seal signed by X Corp. allows Y Corp. to confirm only the validity of the final version of the document. However, the validity of the incomplete documents signed by the individual approver is outside the verifier’s scope. For example, the verifier of Y Corp. cannot detect internal rewriting of these documents by using only e-Seal. To avoid these abuses, we define the binding between a signature of the corporation and some signatures of the individual approvers to assign the KAIAS signature to the e-Seal. X Corp. should not disclose incomplete documents generated before the final version for Y Corp. for verification, which is shared only within X Corp. from the viewpoint of privacy and compliance. We introduce Message Hiding schemes which mean an extension of KAIAS(HCSign, HCVf, and DR) to meet the following requirements for the above use cases. • X Corp. can keep the incomplete documents confidential from Y Corp. • Y Corp. can verify that the final version has been issued by Corp X. according to the appropriate approval route without using the original incomplete documents. • Y Corp. does not usually require incomplete documents for the verification but can verify their validity using a disclosure request. Hash Chain Signing. KAIAS.HCSign(params, {m1 , · · · mn }, {sk1 , · · · , skn }) At first, the 1st signer broadcasts a random value seed corresponding to this document. Let mi be a message (document) and Ri = g ri be a part of Schnorr Signatures for each i-th signer described above. Each i-th signer generates (hi , σi = (si , Ri )) and send it to the proxy like following procedures.

294

R. Kojima et al.

hi ← H0 (mi ||Ri )  H0 (h1 ||seed)  hi ← hi−1 ) H0 (hi ||

(i = 1) (otherwise)

  ci ← H0 (R, X, hi ) si = ri + ci ai xi mod p Finally, the proxy outputs vals = (seed, h1 , · · · , hn−1 , mn , Rn ,  hn ) instead of  is a signature of the message list {m1 , · · · , mn } and an KAIAS signature σ { h1 , · · · ,  hn } signed by x1 , · · · , xn .  Hash Chain Verification. KAIAS.HCVerify(params, vals, σ , X) Before the normal verification, the verifier confirm the validity of Hash Chain to compute the recursively steps such as hn ← H0 (mn ||Rn )  H0 (h1 ||seed)   hi ← hi−1 ) H0 (hi ||

(i = 1) (otherwise)

If  hn =  hn then hash chain verification is successful which means {h1 , · · · , hn } and mi are certain non-spoofed hash value and generated the sequentially ascending order. Subsequent process is the same as  hn }, σ , X). KAIAS.AVf(params, { h1 , · · · ,  Disclosure Requirements. DR(h1 , · · · , hn , mi , Ri ) If the signer publishes the evidence mi , Ri for any i, then the verifier output 1 if there exist i s.t. hi = H0 (mi ||Ri ). Using Message Hiding, both X Corp. and Y Corp. can have the following advantages. • No matter how the structure of the internal approval path in X Corp. is constructed, the verifier can confirm that all signatures are valid by verifying only a single aggregated signature (Signatures Aggregation). • The verifier of Y Corp. does not require all X Corp. signer’s public keys and certificates (Key Aggregation). • The signer can extend the hash chain to add mn+1 , mn+2 , · · · that follows mn and aggregate its new signature and existing aggregated signatures (Dynamic Signature Aggregation). • The signers are only responsible for their signing, not other documents (Signing Only Own Message). • X Corp. keeps the incomplete documents confidential from Y Corp. using the hash chain until publishing it. Y Corp. can confirm the validity of whether the published document is really used in contract using the hash chain (Disclosure Requirements).

A Novel Scheme of Schnorr Multi-signatures for Multiple Messages

295

References [BDN18] Boneh, D., Drijvers, M., Neven, G.: Compact Multi-signatures for smaller blockchains. In: Peyrin, T., Galbraith, S., (eds.) Advances in Cryptology – ASIACRYPT 2018, pp. 435–464. Springer, Cham (2018) [BN06] Bellare, M., Neven, G.: Multi-signatures in the plain public-key model and a general forking lemma, pp. 390–399, January 2006 [Bol03] Boldyreva, A.: Threshold signatures, multisignatures and blind signatures based on the gap-diffie-hellman-group signature scheme, pp. 31–46, January 2003 [Bon11] Boneh, D.: Aggregate signatures, p. 27. Springer, Boston (2011). https:// doi.org/10.1007/978-1-4419-5906-5 139 [EBS16] El Bansarkhani, R., Sturm, J.: An efficient lattice-based multisignature scheme with applications to bitcoins. In: Foresti, S., Persiano, G., (eds.) Cryptology and Network Security, pp. 140–155. Springer, Cham (2016) [eID14] Regulation (EU) no 910/2014 of the European parliament and of the council, July 2014. https://eur-lex.europa.eu/legal-content/EN/TXT/? uri=uriserv:OJ.L .2014.257.01.0073.01.ENG [LT17] Lindemann, R., Tiffany, E.: FIDO UAF protocol specification, February 2017. https://fidoalliance.org/specs/fido-uaf-v1.1-ps-20170202/fidouaf-protocol-v1.1-ps-20170202.pdf [MPSW19] Maxwell, G., Poelstra, A., Seurin, Y., Wuille, P.: Simple schnorr multisignatures with applications to bitcoin. Designs Codes Cryptograh. 87, 02 (2019) [Nak09] Nakamoto, S.: Bitcoin: a peer-to-peer electronic cash system. Cryptography mailing list at, March 2009.https://metzdowd.com [Oka88] Okamoto, T.: A digital multisignature scheme using bijective public-key cryptosystems (1988) [RY07] Ristenpart, T., Yilek, S.: The power of proofs-of-possession: securing multiparty signatures against rogue-key attacks. Springer, Heidelberg (2007) [Sch91] Schnorr, C.: Efficient signature generation by smart cards. J. Cryptol. 4, 161–174 (1991) [SSH] SSH.COM. Public key authentication for SSH. https://www.ssh.com/ssh/ public-key-authentication [TS118] Trust services and electronic identification (EID), December 2018. https:// ec.europa.eu/digital-single-market/en/trust-services-and-eid [WNR20] Wuille, P., Nick, J., Ruffing, T.: Schnorr signatures for secp256k1, January 2020. https://github.com/bitcoin/bips/blob/master/bip-0340.mediawiki

A Fuzzy-Based Approach for Transmission Control of Sensory Data in Resilient Wireless Sensor Networks During Disaster Situation Daisuke Nishii1 , Makoto Ikeda2(B) , and Leonard Barolli2 1

Graduate School of Engineering, Fukuoka Institute of Technology, 3-30-1 Wajiro-higashi, Higashi-ku, Fukuoka 811-0295, Japan [email protected] 2 Department of Information and Communication Engineering, Fukuoka Institute of Technology, 3-30-1 Wajiro-higashi, Higashi-ku, Fukuoka 811-0295, Japan [email protected], [email protected]

Abstract. Wireless sensor networks can be used for long-term operation and they can collect data effectively from huge volumes of sensed data. There are many applications of wireless sensor networks. In this paper, we propose a fuzzy-based transmission control system of sensed data for resilient wireless sensor networks in disaster situations. From the evaluation results, we found that our proposed system can reduce the transmission interval and extend the lifetime of network for disaster situations. Keywords: Resilient wireless sensor network

1

· Fuzzy logic · Disaster

Introduction

Recently in Japan, weather-related disaster risk information such as rainfall, landslide disaster, typhoons and river levels can be found on websites provided by the Japan Meteorological Agency and municipalities. The alert function can be configured through the mobile app and is a standard function on mobile phones. There is a system that allows users to post and share the conditions in the area. In some cases, the posting function can be used in a high-risk area. With the spread of low-power wide-area networks, advanced data processing methods for dealing with weather-based big data should be developed. A variety of sensors are scattered around our daily life. In several fields of science, smart solutions based on sensors are providing us with the possibility to make decisions that are even better than human decisions [12,14,17]. The main function of a resilient wireless sensor network is the ability to collect information even if in normal and disaster situations [2,4,6,9,16,20,22–24]. c The Editor(s) (if applicable) and The Author(s), under exclusive license  to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 296–303, 2021. https://doi.org/10.1007/978-3-030-61108-8_29

A Fuzzy-Based Approach for Transmission Control of Sensory Data

297

For battery-operated devices, lifetime is also important. Resilient wireless sensor networks can be constructed on the emerging Internet of Things (IoT) toward achieving the Sustainable Development Goals (SDGs) [3,7,15,19]. In [13], the authors present a fuzzy-based approach to solve the distribution system restoration problems after disaster situation. In [21], we proposed a fuzzy-based disaster information gathering system for deciding the accident level. We considered fire strength, gas concentration and temperature as input parameters. The proposed system had good performance in indoor scenario. In this paper, we propose a fuzzy-based transmission control approach of sensed data in resilient wireless sensor networks during disaster situations. For evaluation, we use distance to receiver, data priority and data change amount as input parameters. The structure of the paper is as follows. In Sect. 2, we describe the resilient wireless sensor system design. In Sect. 3, we provide the evaluation results. Finally, conclusions and future work are given in Sect. 4.

2

Resilient Wireless Sensor System in Disaster Situation

In this section, we discuss in detail the design of the fuzzy-based resilient wireless sensor system in disaster situation. We start from the overview of wireless sensor network, development approach, then we present the fuzzy-based system parameters. 2.1

Overview of Wireless Sensor Networks

Wireless sensor networks are able to monitor physical phenomena, processing the sensed data, making decisions based on the sensed data, and completing the appropriate tasks when required. To reduce the transactions and resources within the database, sensor send not every bit of data within the limits of available resources [1]. When important data has to be sent, the sensors can send that data back to the sink, which controls the task from a distance, or they may send the data to a cluster head that can conduct the operation independently of the sink node. 2.2

Development Approach

In this work, we develop an integrated mobile device that serves as an edge device. The edge device is composed of a sensing element and a communication element. The sensing element is implemented in Raspberry Pi. We assume a sensor to measure parameters as distance to receiver, data priority, and data change amount. All sensing data are measured by network tool equipped with the wireless network interface. We implemented a fuzzy logic system written in Python to implement the resilient wireless sensor module in the sink. The sink or cluster head controls the transmission interval to sensors for both normal and disaster situations. The fuzzy-based system acts to different levels of responsibility to improve the transactions and resources.

298

D. Nishii et al. Far

1.0

Medium

Near

0.5

0.0

Low

1.0

Middle

High

0.5

1

0.8

0.6

0.4

0.2

Distance to Receiver(DR)

0

0.0

0

0.2

(a) Input: Distance to Receiver

Small

1.0

Normal

0.4

0.6

Data Priority(DP)

0.8

1

(b) Input: Data Priority

Big

VL

1.0

L

M

H

VH

0.5 0.5

0.0

0

0.2

0.4

0.6

0.8

Data Change Amount(DCA)

1

0.0

0

(c) Input: Data Change Amount

0.1

0.2

0.3

0.4

0.5

0.6

0.7

Transmission Interval(TI)

0.8

0.9

1

(d) Output: Transmission Interval

Fig. 1. Membership function.

2.3

Fuzzy-Based System Parameters

Fuzzy sets and fuzzy logic [25] have been developed to deal with vagueness and uncertainty in a inference process of a intelligent system such as knowledge-based system, expert system or logical control system [5,8,10,18]. The proposed system consists of a Fuzzy Logic Controller (FLC). The FLC basic elements are the fuzzifier, inference engine, fuzzy rule base and defuzzifier. We use triangular and trapezoidal membership functions for FLC, because they are suitable for real-time operation [11]. The membership functions are shown in Fig. 1. We use distance to receiver, data priority and data change amount input parameters for FLC. The output linguistic parameter is the Transmission Interval (TI). The fuzzy rule base is shown in Table 1. Distance to Receiver (DR): The edge device has a network interface controller, which connects to a wireless radio-based computer network. We use the Received Signal Strength Indicator (RSSI) to measure the DR. We consider three levels of DR for different distance level. Data Priority (DP): We consider the DP to support disaster situations. A sink or cluster head can ask for different priorities of data depending on the situation. For example, we can expect to increase the transmission interval in case of non-urgency situation.

A Fuzzy-Based Approach for Transmission Control of Sensory Data

299

Data Change Amount (DCA): The DCA is used to consider time sequence data. We consider three levels of DCA. For example, sometimes there is almost no change from the previous data at time of updating. The proposed system will help to increase the transmission interval at this time. Transmission Interval (TI): Our system can control the data transmission interval to selected sensors depending on the situation. The transmission interval has five different levels, which can be interpreted as: Very Low (VL), Low (L), Middle (M), High (H), Very High (VH).

3

Evaluation Results

We present the simulation and experimental results of a fuzzy-based resilient wireless sensor system in Fig. 2 and Fig. 3. As shown in Fig. 2, the TI increases with increase of DCA regardless the distance to receiver. For far condition, we observed that this case can reduce the transmission interval to the lowest level. For the near condition, we have the same simulation results with the cases when the DP is middle and high. This Table 1. Fuzzy rule base. Rule

Distance to receiver

Data priority

Data change amount

1

Far

Low

Small

Transmission interval VL

2

Far

Low

Normal

VL

3

Far

Low

Big

L

4

Far

Middle

Small

VL

5

Far

Middle

Normal

L

6

Far

Middle

Big

M

7

Far

High

Small

L

8

Far

High

Normal

M

9

Far

High

Big

H

10

Medium

Low

Small

VL

11

Medium

Low

Normal

L

12

Medium

Low

Big

M

13

Medium

Middle

Small

L

14

Medium

Middle

Normal

M

15

Medium

Middle

Big

H

16

Medium

High

Small

M

17

Medium

High

Normal

H

18

Medium

High

Big

VH

19

Near

Low

Small

L

20

Near

Low

Normal

M

21

Near

Low

Big

H

22

Near

Middle

Small

M

23

Near

Middle

Normal

H

24

Near

Middle

Big

VH

25

Near

High

Small

M

26

Near

High

Normal

H

27

Near

High

Big

VH

D. Nishii et al.

1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0

Distance to Receiver(DR) = Medium Transmission Interval(TI)

Transmission Interval(TI)

Distance to Receiver(DR) = Far Low-DP Middle-DP High-DP

1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Low-DP Middle-DP High-DP

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Data Change Amount(DCA)

Data Change Amount(DCA)

(a) DR: Far condition

(b) DR: Medium condition

Transmission Interval(TI)

Distance to Receiver(DR) = Near 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0

Low-DP Middle-DP High-DP

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Data Change Amount(DCA)

(c) DR: Near condition

Fig. 2. Simulation results for different cases.

Low-DP Middle-DP High-DP

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Distance to Receiver(DR) = Medium 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0

Low-DP Middle-DP High-DP

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Data Change Amount(DCA)

(a) DR: Far condition

Data Change Amount(DCA)

(b) DR: Medium condition Distance to Receiver(DR) = Near

Transmission Interval(TI)

1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0

Transmission Interval(TI)

Distance to Receiver(DR) = Far Transmission Interval(TI)

300

1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0

Low-DP Middle-DP High-DP

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Data Change Amount(DCA)

(c) DR: Near condition

Fig. 3. Experimental results for different cases.

A Fuzzy-Based Approach for Transmission Control of Sensory Data

301

is because the sensors are close to each other and can send messages with low energy for transmission. So, the system selects a high transmission interval for the cases with middle or higher priorities. The experimental results for each input and output parameters are shown in Fig. 3. The comparison of the experimental and simulation results shows that the difference is small. From these results, we conclude that the transmission interval of the sensor can be reduced and the lifetime of the network is increased when the DCA is small. The proposed system can be used for evaluation of the energy consumption of the transmission energy by considering the distance to the receiver.

4

Conclusions

In this paper, we propose a fuzzy-based transmission control approach of sensed data for resilient wireless sensor networks in disaster situations. We considered distance to receiver, data priority and data change amount parameters. From the evaluation results, we found that our proposed system can reduce the transmission interval and extend the lifetime of network for disaster situations. In the future work, we will consider the relationship between power consumption and transmission interval in different situations. Acknowledgments. This work has been partially funded by the research project from Comprehensive Research Organization at Fukuoka Institute of Technology (FIT), Japan.

References 1. Aazam, M., Huh, E.N.: Fog computing and smart gateway based communication for cloud of things. In: Proceedings of the International Conference on Future Internet of Things and Cloud (FiCloud-2014), pp. 464–470, August 2014 2. Akyildiz, I.F., Kasimoglu, I.H.: Wireless sensor and actor networks: research challenges. Ad Hoc Netw. J. (Elsevier) 2(4), 351–367 (2004) 3. Al-Fuqaha, A., Guizani, M., Mohammadi, M., Aledhari, M., Ayyash, M.: Internet of Things: a survey on enabling technologies, protocols, and applications. IEEE Commun. Surv. Tutorials 17(4), 2347–2376 (2015) ¨ 4. Akan, O.B., Akyildiz, I.F.: Event-to-sink reliable transport in wireless sensor networks. IEEE/ACM Trans. Netw. 13(5), 1003–1016 (2005) 5. Balan, K., Manuel, M.P., Faied, M., Krishnan, M., Santora, M.: A fuzzy based accessibility model for disaster environment. In: Proceedings of the IEEE International Conference on Robotics and Automation (ICRA-2019), pp. 2304–2310, May 2019 6. Forlizzi, J., DiSalvo, C.: Service robots in the domestic environment: a study of the roomba vacuum in the home. In: Proceedings of the 1st ACM SIGCHI/SIGART Conference on Human-Robot Interaction (ACM HRI-2006), Utah, US, pp. 258– 265, March 2006

302

D. Nishii et al.

7. Guo, Z., Li, G., Zhou, M., Feng, W.: Resilient configuration approach of integrated community energy system considering integrated demand response under uncertainty. IEEE Access 7, 87513–87533 (2019) 8. Gupta, I., Riordan, D., Sampalli, S.: Cluster-head election using fuzzy logic for wireless sensor networks. In: Proceedings of the 3rd Annual Communication Networks and Services Research Conference (CNSR-2005), pp. 255–260 (2005) 9. Jiang, X., Dawson-Haggerty, S., Dutta, P., Culler, D.: Design and implementation of a high-fidelity ac metering network. In: Proceedings of the International Conference on Information Processing in Sensor Networks 2009 (IPSN-2009), San Francisco, US, pp. 253–264, April 2009 10. Li, T.S., Chang, S.J., Tong, W.: Fuzzy target tracking control of autonomous mobile robots by using infrared sensors. IEEE Trans. Fuzzy Syst. 12(4), 491–501 (2004) 11. Mendel, J.M.: Fuzzy logic systems for engineering: a tutorial. Proc. IEEE 83(3), 345–377 (1995) 12. Petrakis, E.G.M., Sotiriadis, S., Soultanopoulos, T., Renta, P.T., Buyya, R., Bessis, N.: Internet of things as a service (iTaaS): challenges and solutions for management of sensor data on the cloud and the fog. Internet Things 3–4, 156–174 (2018) 13. Reddy, G.H., Chakrapani, P., Goswami, A.K., Choudhury, N.B.D.: Fuzzy based approach for restoration of distribution system during post natural disasters. IEEE Access 6, 3448–3458 (2018) 14. Ruan, J., Jiang, H., Li, X., Shi, Y., Chan, F.T.S., Rao, W.: A granular GA-SVM predictor for big data in agricultural cyber-physical systems. IEEE Trans. Ind. Inf. 15(12), 6510–6521 (2019) 15. Schmitt, S., Will, H., Aschenbrenner, B., Hillebrandt, T., Kyas, M.: A reference system for indoor localization testbeds. In: Proceedings of the International Conference on Indoor Positioning and Indoor Navigation (IPIN-2012), Sydney, Australia, pp. 1–8, November 2012 16. Sengupta, S., Das, S., Nasir, M., Vasilakos, A.V., Pedrycz, W.: An evolutionary multiobjective sleep-scheduling scheme for differentiated coverage in wireless sensor networks. IEEE Trans. Syst. Man Cybern. Part C (Appl. Rev.) 42(6), 1093–1102 (2012) 17. Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou, I., Huang, A., Guez, A., Hubert, T., Baker, L., Lai, M., Bolton, A., Chen, Y., Lillicrap, T., Hui, F., Sifre, L., van den Driessche, G., Graepel, T., Hassabis, D.: Mastering the game of Go without human knowledge. Nature 550, 354–359 (2017) 18. Su, X., Wu, L., Shi, P.: Sensor networks with random link failures: distributed filtering for T-S fuzzy systems. IEEE Trans. Ind. Inf. 9(3), 1739–1750 (2013) 19. Sung, J.Y., Guo, L., Grinter, R.E., Christensen, H.I.: My Roomba is Rambo: intimate home appliances. In: Proceedings of the 9th International Conference on Ubiquitous Computing (UbiComp-2007), Seoul, South Korea, pp. 145–162, September 2007 20. Tribelhorn, B., Dodds, Z.: Evaluating the roomba: a low-cost, ubiquitous platform for robotics research and education. In: Proceedings of the IEEE International Conference on Robotics and Automation (IEEE ICRA-2007), Roma, Italy, pp. 1393–1399, April 2007 21. Tsuchiya, G., Ikeda, M., Elmazi, D., Barolli, L., Kulla, E.: A disaster information gathering system design using fuzzy logic. In: Proceedings of The 12th International Conference on Broad-Band Wireless Computing, Communication and Applications (BWCCA-2017), pp. 854–861, Nov 2017

A Fuzzy-Based Approach for Transmission Control of Sensory Data

303

22. Xia, J., Yun, R., Yu, K., Yin, F., Wang, H., Bu, Z.: A coordinated mechanism for multimode user equipment accessing wireless sensor network. Int. J. Grid Util. Comput. 5(1), 1–10 (2014) 23. Yu, Y., Rittle, L.J., Bhandari, V., LeBrun, J.B.: Supporting concurrent applications in wireless sensor networks. In: Proceedings of the 4th ACM International Conference on Embedded Networked Sensor Systems (ACM SenSys-2006), Boulder, US, pp. 139–152, November 2006 24. Yuriyama, M., Kushida, T.: Integrated cloud computing environment with IT resources and sensor devices. Int. J. Space-Based Situated Comput. 1(2/3), 163– 173 (2011) 25. Zadeh, L.: Fuzzy logic, neural networks, and soft computing. ACM Commun. 77–84 (1994)

Parasitic Coil Effects on Communication Performance of Table Type 13.56 MHz RFID Reader: A Comparison Study for Different Coil Turns Yuki Yoshigai1 and Kiyotaka Fujisaki2(B) 1

2

Graduate School of Engineering, Fukuoka Institute of Technology, 3-30-1 Wajiro-higashi, Higashi-ku, Fukuoka 811-0295, Japan [email protected] Department of Information and Communication Engineering, Fukuoka Institute of Technology, 3-30-1 Wajiro-higashi, Higashi-ku, Fukuoka 811-0295, Japan [email protected] Abstract. RFID is very useful as an automatic recognition technology because it can contactless access to the information. However, the communication performance of RFID is easily changed by metal, water and other surrounding environments because this system use the electromagnetic fields for communication. In this paper, we evaluate the effect of the parasitic coil placed on the table type RFID reader on the communication performance between the reader and the tag by changing the number of turns of the parasitic coil.

1

Introduction

By using an induction field or radio wave as the medium for transmitting information between a communication terminal and tags, radio frequency identification (RFID) system realizes the contactless data exchange. RFID is an automatic recognition method that uses wireless communication technology and is used in various situations such as train tickets, driver’s licenses, passports and house keys. Because this technology is extremely useful for managing a large amount of objects and people, it is used for the management of goods in logistics and libraries and for the management of the flow lines of participants at event venues. In order to adapt this technology to new services, many studies are underway [1–6]. However, because the RFID system is easily affected by metal, water and other surrounding environments, the communication performance is significantly changed. In order to utilize the RFID system in familiar services, higher communication performance is important for improving the reliability of RFID systems. Therefore, experiments, evaluations and examination of improved methods are being conducted [7–16]. In [15], we evaluated the effect of parasitic coils placed at fixed positions on a table-type RFID reader and showed that this element may contribute to c The Editor(s) (if applicable) and The Author(s), under exclusive license  to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 304–312, 2021. https://doi.org/10.1007/978-3-030-61108-8_30

Effect of Coil Turns of Parasitic Element on RFID System Performance

305

improve the communication performance. Furthermore, in [16], by evaluating the magnetic field strength on the reader in the presence of the parasitic element, we clarified the effect of the parasitic element on the magnetic field distribution on the reader. In this paper, we show experimentally the effect of the difference in the number of turns of the parasitic coil placed on the table type RFID reader on the communication performance between the RFID reader and the tag. The paper structure is as follows. In Sect. 2, we introduce the RFID system. In Sect. 3, we present the effect of different coil turns on communication performance of a table-type RFID reader. Finally, in Sect. 4, we conclude the paper.

2

RFID System

RFID system is an important tool for achieving automatic identification. In this system, in order to get the ID from the tag without touching it, the wireless communication technique is used. An RFID system consists of two components as shown in Fig. 1. One is called RFID tag and is associated with the object that needs to be managed. Another is the reader/writer, which is used for exchanging information with tags and is called the interrogator.

Fig. 1. Construction of RFID system.

The shape of the tag and the interrogator are changed according to the purpose of use. In particular, tags come in a variety of shapes to support various services. For example, for the management of members is used the card style RFID tag, while for the management of goods is used seal style or label style RFID tag. In white coat rental services, button-shaped tags may be used to distinguish clothes of the same shape. Similarly, the RFID reader/writer is classified into the table type and portable type according to the purpose. In this research, we are studying how to improve communication performance by using a card type RFID tag and a table type reader/writer. Figure 2 shows a photograph of 13.56 MHz table type RFID reader/writer ST-RW01 produced by SOFEL, a RFID tag and a sample of the parasitic element used in our experiments. This RFID system is based on the ISO 15693

306

Y. Yoshigai and K. Fujisaki

13.56MHz table type RFID reader

Square open loop coil

RFID tag

Fig. 2. Photo 13.56 MHz table type RFID reader ST-RW01, RFID tag, and square open loop coil. y

Clear acrylic board Tag

x d

Parasitic coil

h Loop antenna of RFID reader

Fig. 3. Evaluation image of communication performance between reader and target tag when a parasitic coil is placed on the reader.

International Standard. The size of the housing of the table type RFID reader is 25 cm × 35 cm. This reader has an 20 cm square loop antenna in the housing and this system communicates using electromagnetic induction.

3

Effect of Different Coil Turns on Communication Performance

In [15], in order to improve the communication performance on the table type RFID reader, we made some coils with different number of turns as a parasitic element and evaluated the communication performance of RFID system when the parasitic element is placed parallel on the fixed height from the reader face as shown in Fig. 3. By using an open loop coil as a parasitic element, it was possible to increase the communication distance between the reader and tag by increasing the number of coil turns. Furthermore in [16], we evaluated the effect of parasitic elements to the electromagnetic field emitted from the table type RFID reader by observing the magnetic field 13.56 MHz using a magnetic field probe and showed that the parasitic coil helps to increase the magnetic field intensity, which contributes to the tag communication performance.

Effect of Coil Turns of Parasitic Element on RFID System Performance

307

Figure 4 shows the result of evaluating the electric power distribution on the plane at a height d = 0 cm from the reader face when changing the number of turns of the parasitic element placed at a height 6 cm from the reader. In this figure, the electric power distribution in the case of n = 0 is the result when there is no parasitic coil. The shades of colors in this figure indicate that the lighter the color (closer to white), the larger is the received power and the darker the color (closer to dark purple), the smaller is the received power. Considering the power observed by the measuring instrument, the value will be in the range of −35 dBm to −75 dBm. From Fig. 4, it can be seen that the parasitic coil affects the electromagnetic field on the reader and the magnitude of the effect of the coil depends on the number of turns. Figures 5, 6, 7, 8 shows the result of evaluating the distribution of points where reader is able to communicate with tag on a plane located at the height d [cm] from the reader face when the number of turns n of coil is changed and is installed at a height h [cm] from the reader face. In these figures, the red dot indicates the point where the tag and reader can communicate. The distribution points in the case of n = 0 is the result when there is no parasitic coil. From Fig. 5, in situations where the reader and tag can sufficiently communicate without a parasitic coil, it was shown that the reading performance becomes unstable when the number of turns of the parasitic coil is more than 3. Figure 6 shows the results of evaluating the influence of the parasitic coil on the reading performance when the distance between the tag and the reader is set to 12 cm. In this case, it can be seen that the communication performance without the parasitic coil is reduced to 1/4 compared with the communication performance of d = 6 cm. The results show that the reading performance is improved as the number of turns of the parasitic coil is increased, but the reading performance is significantly reduced when the number of turns is 4 or more. In Fig. 7 and Fig. 8 are shown the results when the distance d between the tag and the reader 15 cm and 18 cm, respectively. In the case of d = 15 cm, there are some points where the reader and tag can communicate when the number of turns n of the parasitic coil was 3, 3.5, or 4. When the number of turns of the parasitic coil was 3.5 or 4, the readable spots were circular and communication was not possible at the center of the reader. For d = 18 cm, the reader and tag could communicate when the number of turns n of the parasitic coil was 3.5 or 4. From these results, it is considered possible to extend the communication distance between the reader and tag by using the parasitic coil. These experiments showed that the difference in the number of turns of the parasitic coil affects the communication performance. As a result of the experiment, simply increasing the number of coil turns does not improve the performance. It is necessary to consider the optimum number of turns of the coil. Experiments have shown that the parasitic coil not only improves the communication performance between the reader and the tag, but also can reduce it. The reason why the parasitic coil placed on the reader deteriorates the reading performance of the tag is that the magnetic field emitted from the reader is interfering with the magnetic field re-emitted from the parasitic coil.

Y. Yoshigai and K. Fujisaki 25

-35

20

20

-50

10

-55 10

-65

5

-50

15

-60

5

10

15

20

0

25

5

10

15

20

20

25

'pd00h6t25.da' t

10

-55 10

-65

0 5

10

15

20

25

-70 0

-75 0

5

10

-60 -65

5

-75 20

25

'pd00h6t40.da' t

20

20

25

-35

'pd00h6t45.da' t

-40 20

-50

15

-55 10

-60 -65

5

-45 -50

15

-55 10

-60 -65

5

-70 0

-75 5

10

15 x

n = 3.5

25

-45

0

x

15

n=3 -35

-70 0

10 x

y

y

-55

25

-65

-40

-50

15

15

-60

5

n = 2.5 -35

-45

10

-55

25

-40

5

-50

x

n=2

0

-45

10

-75 0

x

20

-35

'pd00h6t30.da' t

-70

25

'pd00h6t35.da' t

25

15

-65

5

-75 20

20

-40

-60

-70 0

15

20

-50

15

-60

5

25

-45

y

y

-55

10

n = 1.5 -35

20

-50

15

25

5

-40

-45

15

-75 0

x

-40

10

-70 0

25

n=1 -35

'pd00h6t20.da' t

5

-65

x

n=0

0

-60

5

-75 0

x

25

-55

-70

-75 0

-50

10

-65

5

-45

15

-60

-70 0

-40 20

-45

y

y

-55

-35

'pd00h6t15.da' t

-40

-45

15

25

-35

'pd00h6t10.da' t

-40

y

'pd00hntn.da' t

y

25

y

308

n=4

20

25

-70 0

-75 0

5

10

15

20

25

x

n = 4.5

-35

'pd00h6t50.da' t

-40 20

-45 -50

15 y

-55 10

-60 -65

5

-70 0

-75 0

5

10

15

20

25

x

n=5

Fig. 4. Electric power distribution on a plane located at the height d = 0 [cm] from the reader face when the number of turns n of coil is changed and is installed at the height h = 6 cm from the reader face.

Figure 9 shows the results of evaluating the communication performance between the reader and the tag when the parasitic coils of 3.5 turns are placed at a height 6 cm 9 cm above the reader and the tag is 15 cm away from the reader. When the parasitic coil is attached to h = 6 cm, the central part of the reader cannot communicate with the tag. On the other hand, this situation did not

Effect of Coil Turns of Parasitic Element on RFID System Performance 24

24

24

18

18

18

y 12

y 12

y 12

6

6

6

0

6

12

x

18

24

0

6

n=0

12

x

18

24

0

n=1 24

24

18

18

18

y 12

y 12

y 12

6

6

6

6

12

x

18

24

0

6

n=2

12

x

18

24

0

24

18

18

18

y 12

y 12

y 12

6

6

6

12

x

18

24

n = 3.5

0

6

12

x

n=4

x

18

24

12

x

18

24

18

24

n=3

24

6

6

n = 2.5

24

0

12

n = 1.5

24

0

6

309

18

24

0

6

12

x

n = 4.5

24

18

y 12 6

0

6

12

x

18

24

n=5

Fig. 5. Distribution of points where reader is able to communicate with tag on a plane located at the height d = 6 [cm] from the reader face when the number of turns n of coil is changed and is installed at a height 6 cm from the reader face.

occur when the parasitic coil was attached at h = 9 cm. From this result, it can be seen that in order to improve the communication performance between the reader and the tag, it is also necessary to consider the location of the parasitic coil.

310

Y. Yoshigai and K. Fujisaki

24

24

24

18

18

18

y 12

y 12

y 12

6

6

6

0

6

12

x

18

24

0

6

n=0

12

x

18

24

0

n=1 24

24

18

18

18

y 12

y 12

y 12

6

6

6

6

12

x

18

24

0

6

n=2 24

18

18

y 12

y 12

6

6

6

12

x

x

18

24

0

6

n = 2.5

24

0

12

18

24

0

6

n = 3.5

12

x

12

x

18

24

18

24

n = 1.5

24

0

6

12

x

n=3

18

24

n=4

Fig. 6. Distribution of points where reader is able to communicate with tag on a plane located at the height d = 12 [cm] from the reader face when the number of turns n of coil is changed and is installed at a height 6 cm from the reader face.

24

24

24

18

18

18

y 12

y 12

y 12

6

6

6

0

6

12

x

n=3

18

24

0

6

12

x

n = 3.5

18

24

0

6

12

x

18

24

n=4

Fig. 7. Distribution of points where reader is able to communicate with tag on a plane located at the height d = 15 [cm] from the reader face when the number of turns n of coil is changed and is installed at a height 6 cm from the reader face.

Effect of Coil Turns of Parasitic Element on RFID System Performance 24

24

18

18

y 12

y 12

6

6

0

6

12

x

18

24

0

6

n = 3.5

12

x

18

311

24

n=4

Fig. 8. Distribution of points where reader is able to communicate with tag on a plane located at the height d = 18 [cm] from the reader face when the number of turns n of coil is changed and is installed at a height 6 cm from the reader face. 24

24

18

18

y 12

y 12

6

6

0

6

12

x

18

(a) h = 6 cm

24

0

6

12

x

18

24

(b) h = 9 cm

Fig. 9. Comparison of distribution of points where reader is able to communicate with tag on a plane located at the height d = 15 [cm] from the reader face when the number of turns of coil is 3.5 times. Figure 9(a) is installed the parasitic coil at a height 6 cm from the reader face and Fig. 9(b) is installed the parasitic coil at a height 9 cm from the reader face.

These experiments clarified the effect of the parasitic coil on the communication performance of the RFID system. From these results, in order to improve the performance of the RFID system, it is necessary to consider optimization of the number of turns of the parasitic coil and its installation location.

4

Conclusion

In this paper, we evaluated the effect of the difference in the number of parasitic coil turns on the communication performance between a table-type RFID reader and a tag. By experimental results was shown that the number of turns of the coil and the placement of the parasitic coil must be considered in order to improve the communication performance of the RFID tag system. With the goal of improving the performance of RFID systems, we will continue research to solve these problems.

312

Y. Yoshigai and K. Fujisaki

References 1. Li, N., Gerber, B.B.: Performance-based evaluation of RFID-based indoor location sensing solutions for the built environment. Adv. Eng. Inf. 25(3), 535–546 (2011) 2. Minami, T.: RFID tag based library marketing for improving patron services. In: Advances in Knowledge Acquisition and Management, vol. 2303 (2006) 3. Prasad, N.R.K., Rajesh, A.: RFID-based hospital real time patient management system. Int. J. Comput. Trends Technol. 3(3), 509–517 (2012) 4. Symonds, J., Seet, B.C., Xiong, J.: Activity inference for RFID-based assisted living applications. J. Mob. Multimed. 6(1), 15–25 (2010) 5. Tajima, M.: Strategic value of RFID in supply chain management. J. Purchasing Supply Manag. 13(4), 261–273 (2007) 6. Zhonga, R.Y., Dai, Q.Y., Qu, T., et al.: RFID-enabled real-time manufacturing execution system for mass-customization production. Robot. Comput.-Integr. Manuf. 29(2), 283–292 (2013) 7. Basat, S.S., Kyutae, L., Laskar, J., Tentzeris, M.M.: Design and modeling of embedded 13.56 MHz RFID antennas. In: Proceedings of 2005 IEEE International Symposium Antennas and Propagation, pp.64–67 (2005) 8. Cantatore, E., Geuns, T.C.T., Gelinck, G.H., et al.: A 13.56-MHz RFID system based on organic transponders. IEEE J. Solid-State Circ. 42(1), 84–92 (2007) 9. Fujisaki, K.: Implementation of a RFID-based system for library management. Int. J. Distrib. Syst. Technol. 6(3), 1–10 (2015) 10. Fujisaki, K.: Evaluation and measurements of main features of a table type RFID reader. J. Mob. Multimed. 11(1–2), 21–33 (2015) 11. Fujisaki, K.: Evaluation of table type reader for 13.56 MHz RFID system considering distance between reader and tag. In: Proceedings of The 21st International Conference on Network-Based Information Systems, vol. 22, pp. 56–64 (2019) 12. Fujisaki, K.: Evaluation of 13.56 MHz RFID system performance considering communication distance between reader and tag. J. High Speed Netw. 25(1), 61–71 (2019) 13. Potyrailo, R.A., Morris, W.G., Sivavec, T., et al.: RFID sensors based on ubiquitous passive 13.56-MHz RFID tags and complex impedance detection. Wirel. Commun. Mob. Comput. 9(10), 1318–1330 (2009) 14. Yoshigai, Y., Fujisaki, K.: Evaluation of 13.56 MHz RFID system considering tag magnetic field intensity. In: Proceedings of the 21st International Conference on Network-Based Information Systems, vol. 1036, pp. 620–629 (2020) 15. Fujisaki, K., Yoshigai, Y.: Effect of parasitic element on communication performance of 13.56 MHz RFID system. In: Proceedings of The 14th International Conference on Broad-Band Wireless Computing, Communication and Applications, pp. 646–654 (2020) 16. Fujisaki, K., Yoshigai, Y.: Effect of parasitic coil on communication performance on table type 13.56 MHz RFID reader. In: Proceedings of The 23th International Conference on Network-Based Information Systems, pp. 479–487 (2020)

Tuning of Output Optical Signal Wavelength Through Resonant Filter for WDM System Hiroshi Maeda(B) Department of Information and Communication Engineering, Faculty of Information Engineering, Fukuoka Institute of Technology, Fukuoka, Japan [email protected]

Abstract. In two-dimensional photonic crystal waveguide with passive switching function depending on the input light wavelength, the output characteristic is simulated by using frequency dependent FDTD method. For the wide range of input optical wavelength, distribution ratio to output ports is calculated from the electric field profile. For add/drop circuit of wavelength division multiplexing (WDM) optical communication systems, it was found that tuning of add/drop signal wavelength is possible by designing duplexer at branching point with its radius of pillars. Finally, a configuration of drop circuit with simulated structure is proposed.

1 Introduction In recent years, the internet traffic has increased by the development of optical fiber communication and wireless communication. Along with that, information processing of electronic integrated circuits is increasing, and there are problems such as heat generation and increase in power consumption. There is a possibility that these problems can be solved by using optical technology. Therefore, optical integrated circuits have been investigated various research institutions [1–3]. Photonic crystal has attracted attention for the realization of optical integrated circuits. The photonic crystal is artificial structure that blocks light of a specific range of wavelength depending on the material, the periodic lattice and the thickness of the pillar etc. The blocked wavelength band is called the photonic band gap (PBG) [4]. In this research, linear dispersion and nonlinear dielectric material is installed as a resonator in branch point of the two-dimensional photonic crystal waveguide. In previous study, optical switching was confirmed by using the structure [5]. In this paper, distribution characteristic according to the change of the amplitude and wavelength of the input signal is examined by frequency dependent FDTD (Finite Difference Time Domain) method [6]. The method is used to calculate the electromagnetic field in the waveguide, and it can be applied to linear dispersion and nonlinear dielectric materials [7]. Finally, it was found that tuning of add/drop signal wavelength is possible by designing duplexer at branching point with its radius of pillars.

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 313–320, 2021. https://doi.org/10.1007/978-3-030-61108-8_31

314

H. Maeda

2 Frequency Dependent FDTD Method   ∂ According to our previous work [8, 9], we employed two dimensional i.e. ∂y =0 , discretized Maxwell’s equation for transverse electric (TE) mode, which propagates to x and z axis. As is mentioned in next section, we supposed dispersive dielectric material which shows frequency dependency on the dielectric constants and nonlinear Kerrtype dielectric medium which shows nonlinear refractive index change in proportion to electric field intensity |E|2 . The Maxwell equations are formulated to involve those dispersive and nonlinear characteristics by frequency dependent finite difference time domain (FDTD) method [7]. The analysis space is surrounded by Berenger’s perfectly matched layer [10] for absorption of radiated wave. The detail of formulation for this problem are to be referred to Refs. [8, 9]. In this article, linear dispersive and nonlinear dispersive effect of the material plays a key role for optical add/drop function. Optical Kerr effect is undesirable due to generation of radiated field. Therefore, input optical amplitude is fixed to quite low level. Those two dispersive effects are available in the following simulation.

3 Settings for Simulation Optical signal propagating in two-dimensional photonic crystal waveguide with a duplexer with three ports is analyzed. Furthermore, in order to catch the optical signal leaked into the photonic crystal, the output port is added in the photonic crystal. Figure 1 shows illustration of two-dimensional photonic crystal waveguide used in simulation. In the figure, background of the waveguide is free space and the refractive index n0 = 1.0. Black circles are linear dielectric material supposed to be pure silicon (Si). The refractive index n1 = 3.6 as is in Ref. [11]. The green circles are dispersive linear and nonlinear dielectric material to be silica (SiO2 ). The refractive index n2 = 1.5, relative permittivity of dispersive linear and nonlinear dielectric material ε∞ = 2.25, εs = 5.25, linear relaxation frequency fL = 63.7 [THz], linear decaying factor δL = 2.5×10−4 , nonlinear relaxation frequency fNL = 14.8 [THz], nonlinear decaying factor δNL = 3.36 × 10−1 , (3) nonlinear susceptibility χ0 = 0.07, and weight factor for Kerr effect α = 0.7 as is in Ref. [7]. The lattice period L = 551.8 [nm], and the radius of dielectric material R = 0.2L are used. The FDTD discretization are x = z = 9.85 [nm] and t = 0.0209 [fs], respectively. The analysis area is W = 9.69 [μm] in the propagation axis and H = 9.69 [μm] in the transverse axis. The wavelength range of λ = 1.350 − 1.600 [μm]. Berenger’s perfectly matched layer [10] is installed as absorbing boundary condition. The electric field input to port 1 is Gaussian beam and given by       x 2 2π ct (1) sin Ey (x, t) = E0 exp − w0 λ where the fixed input signal amplitude E0 = 0.1 [V/m], the beam spot w0 = 28 [nm], optical speed in vacuum c = 2.998 × 108 [m/s].

Tuning of Output Optical Signal Wavelength Through Resonant Filter

315

Standard part of the cross-correlaon coefficient PORT1 PORT2 INPUT PORT

PORT3 Fig. 1. Top of view 2-D photonic crystal waveguide (duplexer is composed of linear dispersive and nonlinear material at the center of the waveguide)

Optical signal propagating in two-dimensional photonic crystal waveguide with a duplexer with three ports is analyzed. Furthermore, in order to catch the optical signal leaked into the photonic crystal, the output port is added in the photonic crystal. In Fig. 1, two-dimensional photonic crystal waveguide used in simulation is illustrated. The light intensity Pi (i = 1, 2, 3) is calculated from the integral of the square of the output electric field of each port. Definition of the power distribution ratio Si is described as follows;  PORTi Pi dw    (2) Si = (i = 1, 2, 3) P dw + PORT 1 1 PORT 2 P2 dw + PORT 3 P3 dw where w is the axis which is transverse to wave propagation (x or z axis). Cross-correlation coefficient Ci is calculated from the normalized electric field mode output from each port. Ci is described as follows;



 1 n ¯ ¯ j=1 Ey|CC (j) − Ey|CC Ey|porti (j) − Ey|porti 2 Ci =

2 1 n

2 (i = 2, 3) (3) n 1 ¯ ¯ E E − E − E (j) (j) y|CC y|CC y|porti y|porti j=1 j=1 n

n

where, E y|CC (j) is the standard part, E y|porti (j) is the output mode at each port, E¯ y|CC and E¯ y|porti are the time average value of each output mode, and n is the maximum cell number of the waveguide. Also, the mode changes between plus and minus with increase of time step number. In this calculation, the positive maximum mode in the steady state of the electric field is used.

316

H. Maeda

4 Simulation Results

100

1.0

80

0.8

60

0.6

40

0.4

20

0.2

0 0.0 1.35.E-6 1.40.E-6 1.45.E-6 1.50.E-6 1.55.E-6 1.60.E-6 Wavelength [m] S₁

S₂

S₃

C₂

Cross-correla on coefficient

Distribu on ra o %

Distribution ratio and cross-correlation coefficient were calculated by the converged electric field while the input amplitude is changed. Input amplitude E0 was changed in 0.1, 0.5, 1.0, 1.5 [V/m]. At first, Fig. 2 shows distribution ratio of port 1, port 2 and port 3 and Cross-correlation coefficient of port 2 and port 3 for E0 = 0.1 [V/m]. It was confirmed that the distribution ratio to the port 2 and port 3 is switched along with change of the input signal wavelength. The optical signal does not reach port 1 at any wavelength because it is blocked by array of periodic linear dielectric pillars.

C₃

Fig. 2. Distribution ratio of port 1, port 2 and port 3 and cross-correlation coefficient of port 2 and port 3 (E0 = 0.1 [V/m])

In addition, the cross-correlation coefficient exceeds 0.9 in most of the wavelength band, and high correlation is confirmed between the reference part and the mode of each port. However, at the wavelength λ = 1.388 [μm], the cross-correlation coefficient of port 3 is very small. In other words, it is shown that there is no correlation between the mode of the reference part and the mode of port 3. At the wavelength λ = 1.388 [μm] and 1.516 [μm], the electric field distribution is shown in Fig. 3. It is shown that the optical signal does not reach to port 3 at the wavelength λ = 1.388 [μm] from Fig. 3. Finally, variation of radius of pillars of silica (totally eight green pillars in Fig. 1) was examined for tuning output frequency to port 2. In Table 1, relation between magnified radius of green pillars R and power distribution ratio to port 2 is listed. It is also plotted as function of R in Fig. 4. From the table and the figure, we found this structure has tunable transmission characteristics. This is applicable for wavelength division multiplexed (WDM) signal systems. To consider signal add/drop circuit for wavelength multiplexed signal, for example, suppose cascaded circuits with different magnified ‘R’s in Table 1. In first stage of the cascaded circuit with magnification 1.0, only the signal with l = 1.389 [mm] is transmitted to port 2 to be dropped or demultiplexed output signal. The other frequency components are mostly transmitted to port 3 and meet the second stage of the cascaded

Tuning of Output Optical Signal Wavelength Through Resonant Filter

317

Fig. 3. Electric field distribution for E0 = 0.1 [V/m] (left side: λ = 1.388 [μm] and right side: λ = 1.516 [μm])

circuit. In the second stage, when magnification of radius of duplexer pillar to be 2.0, the signal with l = 1.468 [mm] is transmitted to port 2 of the second stage to be dropped output signal. By increasing cascaded stages in an integrated optical circuit, many channels can be multiplexed/demultiplexed in a monolithic structure. This is great advantage for realization of optical signal exchange with high functionality, high stability, with low cost by mass production like semiconductor ICs. To consider signal add/drop circuit for wavelength multiplexed signal, for example, suppose cascaded circuits with different magnified ‘R’s in Table 1. In first stage of the cascaded circuit with magnification 1.0, only the signal with l = 1.389 [mm] is transmitted to port 2 to be dropped or demultiplexed output signal. The other frequency components are mostly transmitted to port 3 and meet the second stage of the cascaded circuit. In the second stage, when magnification of radius of duplexer pillar to be 2.0, the signal with l = 1.468 [mm] is transmitted to port 2 of the second stage to be dropped output signal. By increasing cascaded stages in an integrated optical circuit, many channels can be multiplexed/demultiplexed in a monolithic structure. This is great advantage for realization of optical signal exchange with high functionality, high stability, with low cost by mass production like semiconductor ICs. In Fig. 5, a configuration of drop circuit using simulated structure above is illustrated. The remarkable point is that three resonators with different diameters are situated between input (left side) and output (right side) ports to classify the input wavelength components and to lead toward each drop (output) port. In the figure, input WDM signal which is multiplexed with four kinds of optical signals with different wavelength λ1 to λ2 is given at the left side port. The WDM signal encounters first cavity which resonates for λ1, then the signal of λ1 is transmitted downward to the first drop (output) port. The other wavelength components from λ2 to λ4 propagate to the right and the component of λ2 and λ3 are dropped to each output port, respectively. The remaining component of λ4 is obtained from the most righthand-side port.

318

H. Maeda

Table 1. Relation between magnified radius R of eight pillars which composes cavity duplexer at branch point in Fig. 1 and the power distribution ratio to port 2 and wavelength.

Wa vel ength that S 2 is maximum value [μm]

Tuning of Output Optical Signal Wavelength Through Resonant Filter

319

1.499 1.489 1.479 1.469 1.459 1.449 1.439 1.429 1.419 1.409 1.399 1.389 1.0 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2.0 2.1 2.2 2.3 2.4 2.5 Ma gni fication of ra dius R of silica pillar

Fig. 4. Output wavelength to port 2 as a function of magnified radius of R of six pillars at branch point.

Fig. 5. Demultiplexing (DROP) circuit based on proposed structure for WDM input signal with four different wavelengths.

320

H. Maeda

5 Conclusion In two-dimensional photonic crystal waveguide with a linear dispersive and nonlinear medium duplexer, it was found that the distribution ratio changes depending on the input signal wavelength. From our previous research, it is found that E0 = 0.5, 1.0, 1.5 [V/m] is not suitable for optical signal switching, and the most suitable input amplitude for optical signal switching is E0 = 0.1 [V/m]. Therefore, the input signal amplitude level is fixed in this article. We demonstrated that the output signal amplitude changes due to variation of diameter of resonator pillars. Then, a configuration of drop circuit utilizing the simulated structure is proposed. This is a great advantage for fabrication of optical integrated circuits based on photonic crystal structure, because we can make use of lithography technique for fabrication of semiconductor integrated circuits. As future challenge, we would like to demonstrate WDM signal propagation in the proposed drop circuit by a large scale simulation.

References 1. The Telecommunications Association: NTT Tech. J. 22(5) (2010) 2. Takahashi, Y., Inui, Y., Chihara, M., et al.: A micrometer-scale Raman silicon laser with a microwatt threshold. Nature 498, 470–474 (2013) 3. Al Islam, P., Sultan, N., Nayeem, S., et al.: Optimization of photonic crystal waveguide based optical filter. In: The IEEE Region 10 Symposium, pp. 162–167 (2014) 4. Noda, S., Baba, T. (eds.): Roadmap on Photonic Crystals. Kluwer Academic Publishers, The Netherlands (2003) 5. Higashinaka, N., Maeda, H.: Routing of optical baseband signal depending on wavelength in periodic structure. In: Advance on Broad-Band Wireless Computing, Communication and Applications, pp. 621–629, November 2019 6. Yee, K.S.: Numerical solution of initial boundary value problems involving Maxwell’s equation. IEEE Trans. Antennas Propag. 14(3), 302–307 (1996). IEEE Trans. EMC 32(3), 222–227 (1990) 7. Sullivan, D.M.: Nonlinear FDTD formulations using Z transforms. IEEE Trans. Microwave Theor. Tech. 43(3), 676–682 (1995) 8. Higashinaka, N., Maeda, H.: Input amplitude dependency of duplexer with dispersive and nonlinear dielectric in 2D photonic crystal waveguide. In: Proceedings of EIDWT 2020, LNDECT, vol. 47, pp. 226–236, February 2020. https://doi.org/10.1007/978-3-030-397463_25 9. Maeda, H., Higashinaka, N.: Wavelength tuning of output optical signal through resonant filter for WDM system by periodic structure composed of silica glass. In: Proceedings of NBiS-2020, to be presented, August 2020 10. Berenger, J.: A perfectly matched layer for the absorption of electromagnetic waves. J. Comput. Phys. 114(2), 185–200 (1994) 11. Li, X., Maeda, H.: Numerical analysis of photonic crystal directional coupler with kerr-type nonlinearity. In: Proceedings of 5th Asia-Pacific Engineering Research Forum on Microwaves and Electromagnetic Theory, pp. 165–168 (2004)

Design and Implementation of a DQN Based AAV Nobuki Saito1(B) , Tetsuya Oda1 , Aoto Hirata2 , Yuto Hirota1 , Masaharu Hirota3 , and Kengo Katayama1 1

2

Department of Information and Computer Engineering, Okayama University of Science (OUS), 1-1 Ridaicho, Kita-ku, Okayama 700-0005, Japan {t18j057ny,t17j033sn}@ous.jp, {oda,katayama}@ice.ous.ac.jp Engineering Project Course, Okayama University of Science (OUS), 1-1 Ridaicho, Kita-ku, Okayama 700-0005, Japan [email protected] 3 Department of Information Science, Okayama University of Science (OUS), 1-1 Ridaicho, Kita-ku, Okayama 700-0005, Japan [email protected]

Abstract. The Deep Q-Network (DQN) is a method of deep reinforcement learning algorithm. DQN is a deep neural network structure used for the estimation of Q value of the Q-learning technique. The authors have previously developed a simulation system on DQN-based behavioral control methods for actuator nodes in Wireless Sensor Actor Networks (WSANs). In this paper, an Autonomous Aerial Vehicle (AAV) testbed is designed and implemented for DQN-based mobility control. We evaluate the performance of the AAV testbed for a indoor single-path environment. For simulation results show that the DQN can control the AAV.

1

Introduction

In recent years, Japan has been devastated by torrential rains. For example, after 2017, torrential rains occurred in northern Kyushu in July 2017 (total damage: 224 billion yen), torrential rains in July 2018 (total damage: 1,405 billion yen), and torrential rains in July 2020. Meteorological monitoring methods for heavy rainfall disasters include meteorological radar, Doppler radar and polarization radar. However, all of these methods are difficult to predict localized torrential rains with small spatial scales and to detect the victims. The authors study Wireless Sensor Actor Networks (WSANs) that are capable of autonomous behavior in light of monitoring in heavy rainfall. The WSANs consist of wireless network nodes, with the ability to sense events (sensors) and to perform actuations (actors) based on the sensing data collected by all sensors. There is many applications of WSANs such as Autonomous Aerial Vehicles (AAV), Autonomous Underwater Vehicles (AUV), Heating, Ventilation, and Air Conditioning (HVAC), Internet of Things (IoT), Ambient Intelligence and Ubiquitous Robot. The WSAN nodes in these applications are integrated sensor and c The Editor(s) (if applicable) and The Author(s), under exclusive license  to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 321–329, 2021. https://doi.org/10.1007/978-3-030-61108-8_32

322

N. Saito et al.

actuator nodes that are high processing ability, high communication ability, large battery capacity and may include another functions such as mobility. The authors have previously developed a simulation system for Deep QNetwork (DQN) based behavioral control methods for actuator nodes in WSANs [1–4]. DQN is a type of deep reinforcement learning and is the deep learning algorithm estimates of the Q-value in Q-learning algorithm. Deep reinforcement learning is a function approximation method using deep neural network for value function and policy function in reinforcement learning. Deep Q-Network (DQN) using a Convolution Neural Network (CNN) as a function approximation of Q-leaning is a Deep Reinforcement Learning method proposed by Mnih et al. [5,6]. In paper [5], the authors present the implementation and performance evaluation of DQN for different Atari 2600 games. For experimental results, DQN can use game screen input without feature value design, and better than conventional reinforcement learning method with feature value design using linear function [7]. DQN combines the methods of neural fitted Q iteration [8], experience replay [9], sharing the hidden layer of action value function in each behavior pattern, and learning can be stabilized even with a nonlinear function such as CNN [10,11]. In this paper, we implement an AAV testbed and it’s control method based on DQN. The performance ecaluation of AAV testbed is evaluated for a indoor single-path environment.

2

Implementation of an AAV Testbed

In this section, we describes the design of AAV testbed. 2.1

Design of a Quadrotor

In Fig. 3 are shown the snapshot and structure of our AAV. The design of AAV makes use of a quadcopter, which is a type of multicopter. Multicopters are high maneuverable and can go into places that are difficult for humans to enter, such as disaster areas and danger zones. It also has the advantage of not requiring space for takeoffs and landings and being able to stop at mid-air during the flight, therefore enabling activities at fixed points. The Quadrotor is a type of rotary-wing aircraft that uses 4 rotors for takeoff and propulsion. They are also less expensive to manufacture and operate with less power than hexacopters and octocopters. Figure 1(a) shows a snapshot of the actual quadrotor that we designed and created. The quadrotor frame is mainly composed of a PVC pipe and acrylic plate. The parts for connecting the battery, motor, sensor, etc. to the frame were created using an optical three-dimetional printer. Figure 1(b) shows the structure of AAV configuration. The Flight Controller (FC) is a component that calculates the optimum motor rotation speed for flight based on the information sent from the built-in acceleration sensor and gyro

Design and Implementation of a DQN Based AAV

323

Fig. 1. AAV testbed implementation.

sensor. The Electronic Speed Controller (ESC) is a part that controls the rotation speed of the motor in response to commands from FC. The Raspberry Pi uses telemetry communication to send commands such as up, down, front, back, left and right to the flight controller. In addition, multiple Time-of-Flight (ToF) range sensors using Inter-Integrated Circuit (I2 C) communication and General-Purpose InputOutput (GPIO) are used to acquire and save flight data. Output and save values such as up, down, left, right, front and back for the process of movement when simulationg a single indoor path environment using DQN. The movement process during the DQN simulation is stored values

324

N. Saito et al. Table 1. Components of C quadrotor. Component

Model

Manufacture

Propeller

15 × 5.8

Hobby king

Morter

MN3508 700kv

T-motor

Electric Speed Controller

F45A 32bitV2

T-motor

Flight Controller

Pixhawk 2.4.8

Hobby ant

Power Distribution Board MES-PDB-KIT

Lynxmotion

Li-Po Battery

22.2v 12000mAh XT90 YoWoo

Mobile Battery

Pilot Pro 2 23000mAh Poweradd

ToF Ranging Sensor

VL53L0X

Kookye

Camera

IMX219

SainSmart

Raspberry Pi 3

Model B Plus

ABOX

Table 2. Size of quadrotor. Size (Including propeller) Values Length [cm]

87

Width [cm]

87

Height [cm]

30

Diagonal [cm] weight [g]

107 4259

and reproduces the simulation results. Table 1 shows the components used in the quadrotor. Table 2 shows the size and weight of the quadrotor including the propeller. 2.2

DQN for Moving Control of AAV

We present the design and implementation of proposed simulation system based on DQN for actor node mobility control in WSANs. The actor node can choose its networking, moving direction, actuation and sensing policies to maximize the number of sensing events and the number of connected integrated sensor and actor nodes. The DQN for moving control of AAV structure is shown in Fig. 2. The proposed simulating system is implemented by Rust programming language [12,13]. In this work, we use the Deep Belief Network (DBN), where computational complexity is smaller than CNN for DNN part in DQN. The environment is set as vi . At each step, the agent selects an action at from the action sets of the mobile actor nodes and observes a communication and sensing coverage vt from the current state. The change of the mobile actor node score rt was regarded as the reward for the action. For a reinforcement learning, we can complete all of

Design and Implementation of a DQN Based AAV

325

Fig. 2. DQN for moving control of AAV.

these mobile actor nodes sequences mt as Markov decision process directly, where sequences of observations and actions mt = v1 , a1 , v2 , . . . , at−1 , vt . Likewise, it uses a method known as experience replay in which it store experiences of the agent at each timestep, et = (mt , at , rt , mt+1 ) in a dataset D = e1 , . . . , eN , cached over many episodes into a Experience Memory. Defining the discounted reward for the future T by a factor γ, the sum of the future reward until the end would be Rt = t =t γ t −t rt . T means the termination time-step of the mobile actor nodes. After running experience replay, the agent selects and executes an action according to an -greedy strategy. Since using histories of arbitrary length as inputs to a neural network can be difficult, Q function instead works on fixed length format of histories produced by a function φ. The target was to maximize the action value function Q∗ (m, a) = maxπ E[Rt |mt = m, at = a, π], where π is the strategy for selecting of best action. From the Bellman equation, it is equal to maximize the expected value of r+γQ∗ (m , a ), if the optimal value Q∗ (m , a ) of the sequence at the next time step is known. Q∗ (m , a ) = Em ∼ξ [r + γmax Q∗ (m , a )|m, a]  a

(1)

Not using iterative updating method to optimize the equation, it is common to estimate the equation by using a function approximator. Q-network in DQN was such a neural network function approximator with weights θ and Q(s, a; θ) ≈ Q∗ (m, a). The loss function to train the Q-network is: Li (θi ) = Es,a∼ρ(· ) [(yi − Q(s, a; θi ))2 ].

(2)

326

N. Saito et al. Table 3. Simulation parameters of DQN. Parameters

Values

Number of episode

10000

Number of iteration

400

Number of hidden layers

3

Number of hidden units

15

Initial weight value

Normal Initialization

Activation function

ReLU

Action selection probability () 0.999 − (t/Number of episode) (t = 0, 1, 2, . . ., Number of episode) Learning rate (α)

0.04

Discount rate (γ)

0.9

Experience memory size

300 × 100

Batch size

32

Number of actor nodes

1

The yi is the target, which is calculated by the previous iteration result θi−1 . ρ(m, a) is the probability distribution of sequences m and a. The gradient of the loss function is shown in Eq. (3): ∇θi Li (θi ) = Em,a∼ρ(· );s ∼ξ [(yi − Q(m, a; θi ))∇θi Q(m, a; θi )].

(3)

We consider tasks in which an agent interacts with an environment. In this case, the actor node moves step by step in a sequence of observations, actions and rewards. We took in consideration the connectivity, sensing and mobility of actor nodes. For an actor node are considered 7 mobile patterns (up, down, back, forward, right, left, stop). The actor nodes have networking, sensing, mobility and actuation mechanisms. In order to decide the reward function, we considered Distance between AAV and Obstacle (DAO) parameter. The reward function r is defined as follows: ⎧ −5 (if Cur. X ≤ M in. X ∧ Cur. X ≥ M ax. X ⎪ ⎪ ⎪ ⎪ ∧Cur. Y ≤ M in. Y ∧ Cur. Y ≥ M ax. Y ⎪ ⎪ ⎨ ∧Cur. Z ≤ M in. Z ∧ Cur. Z ≥ M ax. Z) r= (M ax. X/2 − Cur. X)+ ⎪ ⎪ ⎪ ⎪ (M ax. Y − Cur. Y )+ ⎪ ⎪ ⎩ (M ax. Z/2 − Cur. Z) (else). In equation, “X”, “Y ” and “Z” means X-axis, Y -axis and Z-axis, respectively. The initial weights values are assigned as Normal Initialization [14]. The input layer is using actor nodes and the position of events, total reward values in Experience Memory and mobile actor node patterns. The hidden layer is connected with 256 rectifier units in Rectified Linear Units (ReLU) [15]. The output Q values are actor node movement patterns.

Design and Implementation of a DQN Based AAV

327

Fig. 3. Problem area.

3

Simulation Results

Table 3 shows the parameters used in the simulation. In Fig. 3, we show the problem area. Figure 3(a) is an snapshot of the indoor path environment and Fig. 3(b) shows the problem area of single indoor path environment used for simulation by DQN. The simulation is executed for one actor node assuming a single AAV. The initial placement of the actor node is the X-axis is the mobile range on the X-axis divided by 2, and the Y -axis and Z-axis are 0. There are 400 repetitions in one episode.

328

N. Saito et al. 300

Best Median Worst

250

Reward

200 150 100 50 0 -50

0

50

100

150

200

250

300

350

400

Number of Iteration

Fig. 4. Simulation results.

Figure 4 shows a simulation result of reward values for movement in a singlepath environment. The best reward value means that AAV move without hitting the wall. Also, a higher reward value means that it is moving closer to the center of the width and height toward the back of the single route.

4

Conclusion

In this paper, we designed and implemented AAV testbed for DQN-based mobility control. We also evaluated the indoor single-path environment. The simulation results showed that DQN could control AAV. In the future, we would like to conduct more extensive experiments assuming various environments and improve the autonomous control system using DQN. Acknowledgement. This work was supported by Grant for Promotion of Okayama University of Science (OUS) Research Project (OUS-RP-20-3).

References 1. Oda, T., Obukata, R., Ikeda, M., Barolli, L., Takizawa, M.: Design and implementation of a simulation system based on deep q-network for mobile actor node control in wireless sensor and actor networks. In: Proceedings of The 31-th IEEE International Conference on Advanced Information Networking and Applications Workshops (IEEE WAINA-2017) (2017) 2. Oda, T., Kulla, E., Cuka, M., Elmazi, D., Ikeda, M., Barolli, L.: Performance evaluation of a deep q-network based simulation system for actor node mobility control in wireless sensor and actor networks considering different distributions of events. In: Proceedings of The 11-th International Conference on Innovative Mobile and Internet Services in Ubiquitous Computing (IMIS-2017), pp. 36–49 (2017)

Design and Implementation of a DQN Based AAV

329

3. Oda, T., Elmazi, D., Cuka, M., Kulla, E., Ikeda, M., Barolli, L.: Performance evaluation of a deep q-network based simulation system for actor node mobility control in wireless sensor and actor networks considering three-dimensional environment. In: Proceedings of The 9-th International Conference on Intelligent Networking and Collaborative Systems (INCoS-2017), pp. 41–52 (2017) 4. Oda, T., Kulla, E., Katayama, K., Ikeda, M., Barolli, L.: A deep q-network based simulation system for actor node mobility control in WSANs considering threedimensional environment: a comparison study for normal and uniform distributions. In: Proceedings of The 12-th International Conference on Complex, Intelligent, and Software Intensive Systems (CISIS-2018), pp. 842–852 (2018) 5. Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A.A., Veness, J., Bellemare, M.G., Graves, A., Riedmiller, M., Fidjeland, A.K., Ostrovski, G., Petersen, S., Beattie, C., Sadik, A., Antonoglou, I., King, H., Kumaran, D., Wierstra, D., Legg, S., Hassabis, D.: Human-level control through deep reinforcement learning. Nature 518, 529–533 (2015) 6. Mnih, V., Kavukcuoglu, K., Silver, D., Graves, A., Antonoglou, I., Wierstra, D., Riedmiller, M.: Playing Atari with Deep Reinforcement Learning, pp. 1–9 (2013). arXiv:1312.5602v1 7. Lei, T., Ming, L.: A robot exploration strategy based on q-learning network. In: IEEE International Conference on Real-time Computing and Robotics (RCAR2016), pp. 57–62 (2016) 8. Riedmiller, M.: Neural fitted Q iteration - first experiences with a data efficient neural reinforcement learning method. In: The 16-th European Conference on Machine Learning (ECML-2005), vol. 3720 of the series Lecture Notes in Computer Science, pp. 317–328 (2005) 9. Lin, L.J.: Reinforcement learning for robots using neural networks. Technical report, DTIC Document (1993) 10. Lange, S., Riedmiller, M.: Deep auto-encoder neural networks in reinforcement learning. In: The 2010 International Joint Conference on Neural Networks (IJCNN2010), pp. 1–8 (2010) 11. Kaelbling, L.P., Littman, M.L., Cassandra, A.R.: Planning and acting in partially observable stochastic domains. Artif. Intell. 101(1–2), 99–134 (1998) 12. The Rust Programming Language. https://www.rust-lang.org/. Accessed 14 Oct 2019 13. Takano, K., Oda, T., Kohata, M.: Design of a DSL for converting rust programming language into RTL. In: Proceedings of The 8-th International Conference on Emerging Internet, Data & Web Technologies (EIDWT-2020), pp. 342–350 (2020) 14. Glorot, X., Bengio, Y.: Understanding the difficulty of training deep feedforward neural networks. In: The 13-th International Conference on Artificial Intelligence and Statistics (AISTATS-2010), pp. 249–256 (2010) 15. Glorot, X., Bordes, A., Bengio, Y.: Deep sparse rectifier neural networks. In: The 14-th International Conference on Artificial Intelligence and Statistics (AISTATS2011), pp. 315–323 (2011)

A Dynamic Tree-Based Fog Computing (DTBFC) Model for the Energy-Efficient IoT Keigo Mukae1(B) , Takumi Saito1 , Shigenari Nakamura1 , Tomoya Enokido2 , and Makoto Takizawa1 1

Hosei University, Tokyo, Japan [email protected], [email protected], [email protected], [email protected] 2 Rissho University, Tokyo, Japan [email protected]

Abstract. In order to decrease the energy consumption of the IoT (Internet of Things), the TBFC (Tree-based Fog Computing) model is proposed in our previous studies. Here, fog nodes are hierarchically structured where a root node shows a cluster of servers and a leaf node indicates an edge node which communicates with sensors and actuators. Each fog node supports a subsequence of subprocesses of an application process. As a volume of sensor data increases and decreases, some fog node gets overloaded and underloaded, respectively. In the DTBFC (Dynamic TBFC) model, the tree structure is dynamically changed by creating and dropping fog nodes so that the energy consumption of fog nodes can be reduced. A fog node is splitted, i.e. some of the subprocesses move to another fog node and is replicated, i.e. the subprocesses are supported by multiple fog nodes. We newly propose a DFC (Dynamic FC) algorithm to dynamically change the tree structure of the fog nodes to reduce the energy consumption and execution time of the fog nodes in a DTBFC model. Keywords: IoT · TBFC (Tree-Based Fog Computing) model · DTBFC (Dynamic TBFC) model · DFC (Dynamic FC) algorithm

1

Introduction

In the IoT (Internet of Things) [8], not only computers like servers and clients but also millions of devices supporting sensors and actuators are interconnected in networks. The FC (Fog Computing) model [16] is proposed to reduce the communication and processing traffic to handle sensor data in the IoT. Here, subprocesses of an application process to handle sensor data are distributed to not only servers but also fog nodes. c The Editor(s) (if applicable) and The Author(s), under exclusive license  to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 330–340, 2021. https://doi.org/10.1007/978-3-030-61108-8_33

A DTBFC Model for Energy-Efficient IoT

331

In the TBFC (Tree-Based FC) models [6,7,9–14], fog nodes are hierarchically structured in a tree, where a root node is a server and a leaf node is an edge node which communicates with sensors and actuators. Each fog node receives input data from child nodes, calculates output data on the input data, and sends the output data to a parent fog node. Thus, a child-parent relation of fog nodes shows the output-input relation of data. Here, subprocesses supported by each fog node and the tree structure of the fog nodes are static, i.e. not changed even if the amount of sensor data is so changed that the some fog nodes are overloaded or underloaded. In the DTBFC (Dynamic TBFC) model [15], fog nodes are dynamically created and dropped in a tree as the traffic of each node increases and decreases, respectively, so that the energy consumption and execution time of fog nodes can be reduced. In this paper, an application process is assumed to be realized as a sequence of subprocesses. An edge subprocess receives input data from sensors and sends output data calculated on the input data to a succeeding subprocess. Thus, each subprocess receives input data from a preceding subprocess and sends output data to a succeeding subprocess. Initially, every subprocess of an application process is supported by one fog node f like the CC (Cloud Computing) model [2]. Here, the fog node f receives data from every sensor. As the amount of sensor data increases, the fog node f is more heavily loaded and consumes more energy. The fog node f is splitted or replicated to a pair of nodes f and f1 to reduce energy consumption and execution time. In the splitting way, some postfix subsequence of subprocesses supported by a node f is supported by a new node f1 and the node f supports only the prefix of subprocesses. In the replication way, the new fog node f1 supports the same subsequence of the subprocesses as the node f . In this paper, we newly propose a DF C (DynamicF C) algorithm to dynamically change the tree structure by splitting and replicating fog nodes. In the evaluation, we show the energy consumption and execution time of fog nodes can be reduced in the DFC model. In Sect. 2, we propose the DTBFC model. In Sect. 3, we discuss the energy consumption and execution time of a fog node. In Sect. 4, we propose the DFC algorithm to split and replicate fog nodes.

2

Dynamic Tree-Based Fog Computing (DTBFC) Model

An application process P to handle data from sensors and to issue actions to actuators is assumed to be a sequence of subprocesses p1 , ..., pm (m ≥ 1) for simplicity. The subprocesses p1 and pm are a root and edge ones, respectively. Here, the edge subprocess pm receives input data from sensors. The subprocess pm obtains output data dm by calculating on the input data from the sensors. Then, the subprocess pm sends the output data dm to the preceding subprocess pm−1 . Thus, each subprocess pi receives input data di+1 from a succeeding subprocess pi+1 and sends output data di to a preceding subprocess pi−1 . Let |d| show the size of data d. The output ratio ri of the subprocess pi is |di |/|di+1 |.

332

K. Mukae et al.

A sequence P  = pi , pi+1 , ..., pj  (1 ≤ i ≤ j ≤ m) of subprocesses is a subsequence of a sequence P = p1 , ..., pi , ..., pj , ..., pm  (P   P ). Subsequences p1 , ..., pi  and pj , ..., pm  (1 ≤ i, j ≤ m) are a pref ix and postf ix of a sequence P = p1 , ..., pm , respectively. Initially, all the subprocesses p1 , ..., pm of an application process P are supported by one fog node f which is a root node, e.g. servers in a cloud. A sensor sends sensor data to the root node f , where the subprocesses p1 , ..., pm calculate output data on a sensor data. As the amount of sensor data increases, the execution time and energy consumption of the node f increase and network traffic to the fog node f also increases. In order to increase the performance and reduce the energy consumption, a prefix p1 , ..., pl (l ≤ m) and postfix pl+1 , ..., pm of the subprocesses supported by the node f are distributed to the nodes f and f1 , respectively, in a splitting way as shown in Fig. 1 (1). The fog node f1 receives all the input data from the child nodes, i.e. sensors and calculates output data d1 on the input data by the subprocesses pl+1 , ..., pm . The fog node f1 then sends the output data d1 to the node f . Here, the nodes f is a parent node of the node f1 . The processing load of the node f thus decreases since the subprocesses pl+1 , ..., pm  are not performed. If the amount of input data of the nodes f and f1 decreases, the node f1 is merged into the node f , i.e. the node f1 is dropped and the node f supports all the subprocesses p1 , ..., pl , pl+1 , ..., pm . Next, if the fog node f1 gets too heavily loaded, the node f1 is replicated to a fog node f2 , i.e. the subsequence pl+1 , ..., pm  of the subprocesses are supported by both the fog nodes f1 and f2 as shown in Fig. 1 (2). Each of the child nodes sends the output data to one of the nodes f1 and f2 . Thus, the nodes f1 and f2 receive smaller amount of sensor data than before replication. The nodes f1 and f2 send the output data to the parent node f . The nodes f1 and f2 are child nodes of the node f . If the amount of sensor data to be sent to the nodes f1 and f2 decreases so that one node can calculate on all the sensor data, one node, say f2 is dropped and every sensor data is sent to the node f1 . In the DTBFC model [15], a root node f has child nodes f1 , ..., fc (c ≥ 1). Each node fi has also child nodes fi1 , ..., fi,ci (ci ≥ 1). Thus, a fog node fR has child nodes fR1 , ..., fR,cR (cR ≥ 1) where the label R is a sequence of numbers. The label R of a node fR shows a path from the root node f to the node fR . Here, fR is a parent node of each node fRi . C(fR ) is a set of child nodes and pt(fR ) shows a parent node of a node fR . Each node fR supports a subsequence SP (fR ) ( P ) of subprocesses of an application process P . The output ratio orR of a node fR is Πpi ∈SP (fR ) ri . For each edge node fR where R = i1 ... ih (1 ≤ h ≤ m), a sequence SP (fi1 ), SP (fi2 ), ..., SP (fi1 ...ih ) of subprocesses is a sequence p1 , ..., pm  of the application process P .

A DTBFC Model for Energy-Efficient IoT

333

Fig. 1. Splitting and replication of subprocesses.

3 3.1

Computation and Energy Consumption Models of a Fog Node Computation Model

An application process P is a sequence p1 , ..., pm  (m ≥ 1) of m subprocesses. That is, subprocesses are sequentially performed. A fog node fR receives input data DR = {dR1 , ..., dR,cR } of size iR (= |DR |) from the child nodes fR1 , .., fR,cR . The fog node fR calculates the output data dR of size oR (= |dR |) on the input data DR by the subprocesses SP (fR ). Here, let SP (fR ) be a subsequence psR , psR +1 , ..., peR  ( P ) (sR = 1, eR ≤ m ) of subprocesses supported by a fog node fR . Then, the fog node fR sends the output data dR to a parent node pt(fR ), where oR = orR · iR for the output ratio orR . Let CRR show the computation rate of a fog node fR to the root node. We assume the computation rate CR of a root node f is 1 and CRR ≤ CR for every fog node fR . This means, each fog node fR supports CRR (≤ 1) times slower computation rate than the root node f . T Pi (x) and T PiR (x) show the execution time [sec] of a subprocess pi to calculate on input data of size x by a root node f and fog node fR , respectively. The execution time T Pi (x) of each subprocess pi on a root node f to calculate on input data of size x is cti · Ci (x) where cti is a constant and Ci (x) shows the computation complexity, which is assumed to be x or x2 in this paper. The size of the output data which the subprocesses pi calculates on the input data of size x is ri · x where ri is the output ratio of the subprocess pi . If a subprocess pi is performed on a fog node fR , the execution time T PiR (x) of the subprocess pi is T Pi (x)/CRR . T CR (SP (fR ), x) shows the execution time [sec] of a node fR to calculate on input data of size x by the subprocesses SP (fR ). The execution time T CR (SP (fR ), x) of a subsequence SP (fR ) = fsR , fsR +1 , ... , feR  of subprocesses on a node fR to calculate on input data DR of size x is given as follows: eR eR T CR (SP (fR ), x) = Σi=s T PiR (rfi · x) = Σi=s T Pi (rfi · x)/CRR . R R

(1)

Here, rfeR = 1 and rfi = reR · reR −1 · ... ·ri−1 for sR ≤ i < eR . The output ratio orR of the node fR with the subprocesses SP (fR ) = fsR , fsR +1 , ... , feR  is

334

K. Mukae et al.

rfsR ·rsR = reR ·reR −1 · ... ·rsR . The size of the output data dsR of the subprocess psR , i.e. the size of the output data dR of the node fR is orR · x. A pair of the execution time T IRi (x) and T ORi (x) of a node fR to receive and send data of size x from a child fog node fRi are proportional to the data size x, i.e. T IRi (x) = rcRi + rtRi · x and T ORi (x) = scRi + stRi · x. Here, rcRi , rtRi , scRi , and stRi are constants. A node fR receives input data dR1 , ..., dR,cR from child nodes fR1 , ..., fR,cR , respectively, where xi = |dRi | for i = 1, ..., cR and x = |DR | = x1 +· · ·+xcR . We assume T IRi (x) = T IR (x) and T ORi (x) = T OR (x) for every fog node fRi . It takes T T IR (x) = T IR (x1 ) + · · · + T IR (xcR ) [sec] to receive the input data dR1 , ..., dR,cR [15]. It takes totally T FR (x) (= T T IR (x) + T CR (SP (fR ), x) + δR · T OR (orR · x)) [sec] to calculate on input data DR = {dR1 , ..., dR,cR } of size x in each node fR . Here, if fR is a root, orR = 0, else orR = 1. The total execution time T TR (psR , ..., peR , x) [sec] of a fog node fR with subprocesses SP (fR ) = psR , ..., peR  for input data DR of size x is given as follows: T TR (psR , ..., peR , x) = T T IR (x) + T CR (psR , ..., peR , x) + δR · T OR (orR · x) (2) eR = T T IR (x) + Σi=s T Pi (rfi · x)/CRR + δR · T OR (orR · x). R

3.2

Energy Consumption Model

Next, we consider electric energy to be consumed by of each fog node fR . Let EIR (x), ECR (SP (fR ), x), and EOR (x) show electric energy [J] consumed by a fog node fR supporting subprocesses SP (fR ) to receive, calculate on, and send data of size x, respectively. In this paper, we assume each fog node fR follows the SPC (Simple Power Consumption) model [3–5] for simplicity. Here, the power consumption of a node fR to calculate on data is maxER [W]. A fog node fR consumes the electric power maxER [W] for time T CR (SP (fR ), x) [sec] to calculate on data of size x. Hence, the energy ECR (SP (fR ), x) [J] consumed by a node fR to perform subprocesses SP (fR ) on input data of size x (> 0) is given as follows: ECR (SP (fR ), x) = maxER · T CR (SP (fR ), x).

(3)

A pair of the power consumption P IR and P OR [W] of a fog node fR to receive and send data are reR · maxER and seR · maxER , respectively, where reR (≤ 1) and seR (≤ 1) are constants. A fog node fR receives input data from child nodes fR1 , ... , fR,cR and send the output data to a parent node. Here, the energy consumption EIR (x) and EOR (x) [J] of a fog node fR to receive and send data of size x (> 0), are given as follows: EIR (x) = P IR · T T IR (x) = reR · maxER · (cR · rcR + rtR · x). EOR (x) = P OR · T OR (x) = seR · maxER · (scR + stR · x).

(4) (5)

A fog node fR consumes energy EFR (SP (fR ), x) to receive and calculate on input data dR1 , ..., dR,cR of size x1 , ..., xcR , respectively, where x = x1 +· · ·+xcR

A DTBFC Model for Energy-Efficient IoT

335

by the subprocesses SP (fR ) = psR , ..., peR  and to send the output data dR of size orR · x to a parent node pt(fR ): EFR (SP (fR ), x) = EIR (x) + ECR (SP (fR ), x) + δR · EOR (orR · x) = {reR · T T IR (x) + T CR (SP (fR ), x) + δR · seR · T OR (ρR · x)} · maxER e

R = {(reR · (cR · rcR + rtR · x)) + Σi=s cti · Ci (rfi · x)/CRR R

+ δR · seR · (scR + stR · orR · x)} · maxER .

(6)

Here, δR = 0 if fR is a root node, else δR = 1.

4 4.1

Algorithms to Split and Replicate Fog Nodes Ways to Split and Replicate Fog Nodes

First, suppose a fog node fR has initially child nodes cfR1 , ..., cfR,qR (qR ≥ 0) and a parent node fR ( = pt(= fR )) as shown in Fig. 2. An application process P is a sequence p1 , ..., pm  (m ≥ 1) of subprocesses. Suppose the node fR supports a subsequence SP (fR ) = psR , psR +1 , ..., peR  ( P ) (1 ≤ sR ≤ eR ≤ m) of the subprocesses of the application process P . Let CF (fR ) be a set {cfR1 , ..., cfR,qR } (qR ≥ 0) of child fog nodes of a fog node fR . A node fR receives input data dR1 , ..., dR,qR from child nodes cfR1 , ..., cfR.qR , respectively, and calculates output data d|R on the input data. Let oi be the size of the input data dRi . Here, x is the total size of input data DR , x = o1 + ... + oqR . This means, the fog node fR supports a subsequence psR , ..., pl ( P ) of subprocesses and each child fog node fRi supports a subsequence pl+1 , ..., peR ( P ) of subprocesses (sR ≤ l < eR ). First, suppose that the subprocesses SP (fR ) of the fog node fR is splitted to fR and a new fog node fR1 . Here, the fog node fR1 is a child node of the fog node fR . Then, the fog node fR1 is replicated to the fog nodes fR1 , ..., fR,cR

Fig. 2. Fog nodes.

336

K. Mukae et al.

Fig. 3. Splitting and replication of fog node.

(cR ≥ 1) as shown in Fig. 3. Each fog node fRi (i = 2, ..., cR ) supports the same subprocesses pl+1 , .., peR  as the fog node fR1 . Each child node cfRi of the node fR is connected to some fog node fRj . Here, qR ≥ cR . Let C(fRi ) (⊆ CF (fR )) be a set of child nodes which are connected to the fog node fRi . Here, C(fRi ) = φ and C(fRi ) ∩ C(fRj ) = φ for every pair of different fog nodes fRi and fRj , and C(fR1 ) ∪ ... ∪ C(fR,cR ) = CF (fR ). As shown in Fig. 3, each fog node fRi receives the input data DRi of size xi from child nodes in C(fRi ). Here, x = x1 + ... +xcR . A fog node fRi receives the input data DRi of size xi from the nodes C(fRi ) and calculates the output data dRi on the input data DRi by the subprocesses peR , peR −1 , ..., pl+1 . The output ratio orRi of each child node fRi is the same corR (= reR ·...·rl+1 ) since each node fRi supports the same subprocess sequence SP (fRi ) = pl+1 , ..., peR . Then, the fog node fRi sends the output data dRi of size corR · xi to the parent node fR as shown in Fig. 1. The node fR receives the input data DR = {dR1 , ..., dR,cR } from all the child nodes fR1 , ..., fR,cR . The size of the output data dR which is calculated on the input data DR of the node fR is corR · x (= x1 + ... +xcR ). The node fR sends the output data dR of size orR · corR · x to the parent node fR (= pt(fR )). 4.2

Energy Consumption and Execution Time

In the DTBFC model, each fog node fR calculates output data on input data from child nodes fR1 , ..., fR,cR and the output data is sent to the parent node to do the further calculation. First, we consider simple case each edge node receives one sensor data. Each child fog node fRi receives input data IDRi from the child nodes and calculates the output data odRi on the data IDRi by a subsequence SP (fRi ) = peR , ..., pl+1  of subprocesses. The parent node fR supports a subsequence SP (fR ) = fl , ..., fsR  of subprocesses. Since every child node fRi supports a same sequence of subprocesses, CSP (fR ) shows subprocesses SP (fRi ) of any child node fRi . We consider how much electric energy a

A DTBFC Model for Energy-Efficient IoT

337

Fig. 4. Energy consumption of subprocesses.

fog node fR and the child fog nodes fRi , ..., fR,cR consume after receiving input data from the child nodes C(fRi ), ..., C(fR,cR ) until sending output data. First, we assume every child fog node fRi receives the input data idRi from the child nodes at the same time τ . The fog node fR starts calculating output data odR on a collection of the input data idRi , ..., idR,cR on receipt of all the input data. We also assume the delay time between every pair of a parent node and a child node is zero. This means, once a child node fRi starts sending the output data odRi to a parent node fR , the parent node fR starts receiving the data idRi . We consider each child node fRi which receives the input data IDRi of size xi from the child nodes. The execution time T TRi (CSP (fR ), xi ) of the child node fRi is T IRi (xi ) + T CRi (CSP (fR ), xi ) + T ORi (corR · xi ) [sec]. The child node fRi consumes energy EFRi (CSP (fR ), xi ) = EIRi (xi ) + ECRi (CSP (fR ), xi ) + T ORi (corR ·xi ) [J]. Each time a fog node fRi sends the output data odRi (= idRi ) of size corR · xi to the parent node fR , the parent node fR receives the input data idRi . In Fig. 4, reci and recj show that a fog node fR receives input data from child nodes fRi and fRj , respectively. Here, the parent node fR spends time T IRi (corR · xi ) [sec] and consumes energy EIRi (cprR · xi ) [J]. Let maxT TR be a maximum one of T TRi (CSP (fR ), xi ), ..., T TRi,cR (CSP (fR ), xR ). This means, the parent node fR starts calculating output data odR on the input data idRi , ..., idR,cR at the time τ + maxT TR . Here, x = xi + ... + xcR . Then, it takes P ETR (SP (fR , corR · x) [sec] and the node fR consumes energy EFR (SP (fR ), corR · x) to do the calculation on the input data of size corR · x and transmit the output data of size orR · corR · x:

338

K. Mukae et al.

P ETR (SP (fR ), corR · x) = T CR (SP (fR ), corR · x) + T OR (orR · corR · x).(7) cR EFR (SP (fR ), corR · x) = Σi=1 EIR (corR · x) + ECR (SP (fR ), corR · x) + EOR (orR · corR · x). (8) Each child node fRi consumes not only energy EFRi (CSP (fR ), ·xi ) to handle the input data idRi of size xi but also energy minERi · (maxT TR − T TRi (CSP (fR ), xi ) where the child node fRi does nothing. The time maxT TR − T TRi (CSP (fR ), xi ) is idle time of a node fRi . The parent node fR also spends cR T IR (corR · xi )[sec]. idle timeIER = maxT TR − Σc=1 The idle timeIERi of a child node fRi and IER of a parent node fR are given as follows: IERi = minERi · (maxT TR − T TRi (CSP (fR) ), xi ) + P ETR (SP (fR ), corR · x)). (9) cR IER = minERi · (maxT TR − Σc=1 T IR (corR · xi )).

(10)

The parent node fR supporting subprocesses and the child nodes fRi , ..., fR,cR supporting subprocesses SP (fR ) = fl+1 , ..., feR  and CSP (fR ) = feR , ..., fl , l,cR totally consume energy SER [J] for time units STRl,cR [sec]: STRl,cR = maxT TR + T CR (SP (fR ), corR · x) + T OR (orR · corR · x). l,cR SER

=

cR Σi=1 EFRi (xi )

(11)

+ EFR (SP (fR ), corR · x) − T TR (corR · x) + IERi + IER .

(12) On receipt of input data idRi from every child node fRi , if a fog node fR still does the calculation, every input data idRi is enqueued in a receipt queue. Let DT be a DTBFC tree of a set F of fog nodes. First, we select a fog node fR where the long a of the receipt gueue is longer. Next, we find l (sR ≤ l ≤ eR where 1,l cR ,l SER is minimum. Then, we find the number cR where SER is minimum. The prefix fl+1 , ..., fsR  is supported by a fog node fR and the postfix feR , ..., fl  is supported by child fog nodes fR1 , ..., fR,cR .

5

Concluding Remarks

In this paper, we discussed the DTBFC (Dynamic Tree Based Fog Computing) model. A fog node is splitted and replicated to multiple fog nodes to reduce the total energy consumption of a fog a fog node, as the traffic of the fog node increases in the tree. We made clear the execution time and energy consumption of nodes. By using the model, we propose an algorithm to split and replicate nodes so that the energy consumption of nodes can be reduced. We are now evaluating a DTBFC model where each fog node is implemented in a Raspberry pi 3 [1].

A DTBFC Model for Energy-Efficient IoT

339

References 1. Raspberry pi 3 model b. URL https://www.raspberrypi.org/products/raspberrypi-3-model-b/ 2. Creeger, M.: Cloud computing: an overview. Queue 7(5), 3–4 (2009) 3. Enokido, T., Ailixier, A., Takizawa, M.: A model for reducing power consumption in peer-to-peer systems. IEEE Syst. J. 4(2), 221–229 (2010) 4. Enokido, T., Ailixier, A., Takizawa, M.: Process allocation algorithms for saving power consumption in peer-to-peer systems. IEEE Trans. Industr. Electron. 58(6), 2097–2105 (2011) 5. Enokido, T., Ailixier, A., Takizawa, M.: An extended simple power consumption model for selecting a server to perform computation type processes in digital ecosystems. IEEE Trans. Industr. Inf. 10(2), 1627–1636 (2014) 6. Guo, Y., Oma, R., Nakamura, S., Duolikun, D., Enokido, T., Takizawa, M.: Data and subprocess transmission on the edge node of TWTBFC model. In: Proceeding of the 11-th International Conference on Intelligent Networking and Collaborative Systems (INCoS-2019), pp. 80–90 (2019) 7. Guo, Y., Oma, R., Nakamura, S., Duolikun, D., Enokido, T., Takizawa, M.: Evaluation of a two-way tree-based fog computing (TWTBFC) model. In: Proceeding of the 13th International Conference on Innovative Mobile and Internet Services in Ubiquitous Computing (IMIS-2019), pp. 72–81 (2019) 8. Hanes, D., Salgueiro, G., Grossetete, P., Barton, R., Henry, J.: IoT Fundamentals: Networking Technologies, Protocols, and Use Cases for the Internet of Things. Cisco Press, United States (2018) 9. Oma, R., Nakamura, S., Duolikun, D., Enokido, T., Takizawa, M.: An energyefficient model for fog computing in the internet of things (IoT). Internet of Things 1–2, 14–26 (2018) 10. Oma, R., Nakamura, S., Duolikun, D., Enokido, T., Takizawa, M.: Energy-efficient recovery algorithm in the fault-tolerant tree-based fog computing (FTBFC) model. In: Proceeding of the 33rd International Conference on Advanced Information Networking and Applications (AINA-2019), pp. 132–143 (2019) 11. Oma, R., Nakamura, S., Duolikun, D., Enokido, T., Takizawa, M.: Evaluation of data and subprocess transmission strategies in the tree-based fog computing (TBFC) model. In: Proceeding of the 22nd International Conference on NetworkBased Information Systems (NBiS-2019), pp. 15–26 (2019) 12. Oma, R., Nakamura, S., Duolikun, D., Enokido, T., Takizawa, M.: A fault-tolerant tree-based fog computing model. Int. J. Web Grid Serv. (IJWGS) 15(3), 219–239 (2019) 13. Oma, R., Nakamura, S., Duolikun, D., Enokido, T., Takizawa, M.: Subprocess transmission strategies for recovering from faults in the tree-based fog computing (TBFC) model. In: Proceeding of the 13th International Conference on Complex, Intelligent, and Software Intensive Systems (CISIS-2019), pp. 50–61 (2019) 14. Oma, R., Nakamura, S., Enokido, T., Takizawa, M.: A nodes selection algorithm for fault recovery in the GTBFC model. In: Proceeding of the 14th International Conference on Broad-Band Wireless Computing, Communication and Applications (BWCCA-2019), pp. 81–92 (2019)

340

K. Mukae et al.

15. Oma, R., Nakamura, S., Enokido, T., Takizawa, M.: A dynamic tree-based fog computing (DTBFC) model for the energy-efficient IoT. In: Proceeding of the 8th International Conference on Emerging Internet, Data and Web Technologies (EIDWT-2020), pp. 24–34 (2020) 16. Rahmani, A., P. Liljeberg, J.S.P., Jantsch, A.: Fog Computing in the Internet of Things. Springer (2018)

An Energy-Efficient Algorithm for Virtual Machines to Migrate Considering Migration Time Naomichi Noaki1(B) , Takumi Saito1 , Dilawaer Duolikun1 , Tomoya Enokido2 , and Makoto Takizawa1 1 Hosei University, Tokyo, Japan {naomichi.noaki.2k,takumi.saito.3j}@stu.hosei.ac.jp, [email protected], [email protected] 2 Rissho University, Tokyo, Japan [email protected]

Abstract. It is critical to reduce electric energy consumed by servers in clusters to reduce carbon dioxide emission. Here, application processes issued by clients are performed on servers in clusters. In this paper, we take the migration approach that application processes on a server migrate to a guest server by using the live migration technologies of virtual machines. By making a virtual machine where application processes are performed migrate from a host server to a guest server, the energy consumption of the servers to perform the application processes may be reduced as discussed in our previous studies. On the other hand, it takes time for a virtual machine to migrate from a host server to a guest server, since it may take a longer time to perform application processes on the virtual machine. In this paper, we first measure migration time of a virtual machine. Then, we propose an MDMG (Maximum energy consumption Difference by virtual machine MiGration) algorithm to select a virtual machine on a host server and a guest server to which the virtual machine to migrate in order to reduce the total energy consumption of servers in a cluster. In the evaluation, we show the energy consumption of servers and the average execution time of application processes can be reduced in the MDMG algorithm. Keywords: Live migration of virtual machine · Energy-efficient virtual machine migration · Migration time · Power consumption model · MDMG algorithm

1

Introduction

It is critical to reduce electric energy consumed by information systems to reduce carbon dioxide emission on the earth. Information systems are composed of clusters of servers and clients which are interconnected in networks. An application c The Editor(s) (if applicable) and The Author(s), under exclusive license  to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 341–354, 2021. https://doi.org/10.1007/978-3-030-61108-8_34

342

N. Noaki et al.

on a client issues a request to a cluster of servers. Here, one sever is selected where an application process to handle the request is performed. Since servers consume more energy than clients, it is critical to reduce the energy consumption of servers to perform application processes issued by clients. The SPC (Simple Power Consumption) [6,7] and MLPC (Multi-Level Power Consumption) [9,10] models are proposed to show the electric power [W] to be consumed by a server to perform application processes. These are macro-level models since the power consumption of a whole server to perform application processes is considered while the power consumption of each hardware component is not considered. By using the models, the energy consumption of a server and the execution time of application processes can be estimated. In order to reduce the energy consumption of servers in a cluster, a more energy-efficient server, which is expected to consume less energy, is selected to perform an application process [8–10]. In another approach, an application process issued by a client is performed on a virtual machine of a server in a cluster. In addition, each virtual machine on a host server migrates to another guest server in a live manner [3] if the guest server is expected to consume smaller energy than the host server [4,5]. By making a virtual machine migrate from a host server to a guest server, the total energy consumed by the host and guest servers can be reduced as discussed in papers [4,5]. On the other hand, it takes time to make a virtual machine vmk migrate, i.e. transmit the image of vmk to a guest server in networks and restart vmk on the guest server. Thus, it may take a longer time to perform application processes on a virtual machine while the host and guest servers may consume less energy if the virtual machine migrates from the host server to the guest servers. We have to make a virtual machine migrate from a host server to a guest server so that not only the energy consumption of servers but also the execution time of application processes can be reduced by taking into consideration the migration time. In this paper, we first measure the live migration time of a virtual machine from a host server to a guest server. By taking the migration time into consideration, we propose an MDMG (Maximum energy consumption Difference by Virtual machine MiGration) algorithm. Each server initially supports one active virtual machine and some number of idle virtual machines. Application processes are issued to active virtual machines and each active virtual machine migrates from a host server to a guest server so as to reduce the energy consumption of the servers. If every active virtual machine on a server st migrates to another server, no process can be performed on the server st . Hence, if no active virtual machine is supported by a server, an idle server transits to active so that each server always supports at least one active virtual machine. A virtual machine migrates from a host server to a guest server so that the energy consumption of the host and guest servers can be mostly reduced. In the evaluation, the total energy consumption of servers and the average execution time of application processes can be reduced in the MDMG algorithm.

An Energy-Efficient Algorithm for Virtual Machines

343

In Sect. 2, we present the system model. In Sect. 3, we measure the migration time of a virtual machine. In Sect. 4, we propose the MDMG algorithm. In Sect. 5, we evaluate the MDMG algorithm.

2 2.1

System Model Virtual Machines

A cluster S is composed of servers s1 , . . . , sm (m ≥ 1). Each server st supports applications with virtual machines. V is a set of all the virtual machines in the cluster S. Let Vt be a set of virtual machines supported by a server st . V = ∪t=1,...,m Vt . A server st supporting a virtual machine vmti is referred to as host server of vmti . Let h (vmk ) stand for a host server which hosts a virtual machine vmk . A client issues a request to a cluster S. One virtual machine vmti on a server st is selected and the request is sent to the virtual machine vmk . An application process to handle the request is created and performed on the virtual machine vmk . On termination of the application process, the virtual machine vmk sends the reply to the client (Fig. 1). In this paper, a term process means an application process to be performed on a virtual machine.

Fig. 1. Virtual machines in a cluster.

A virtual machine vmk can migrate from a host server st to another guest server su in the live manner [3]. That is, processes on the virtual machine vmk move from the host server st to the guest server su without suspension. Let P (vmk ) indicate a set of application processes performed on a virtual machine vmk . Let nvk show the number of processes on a virtual machines vmk , i.e. nvk = |P (nvk )|. A virtual machine vmk is larger than a virtual machine vmh (vmk > vmh ) iff nvk > nvh .

344

2.2

N. Noaki et al.

Power Consumption and Computation Models

A server st is composed of npt (≥ 1) homogeneous CPUs. Each CPU [1] is composed of nct (≥ 1) homogeneous cores. Each core supports the same number ctt of threads. A server st supports processes with totally ntt (= npt · nct · ctt ) threads. Each process is at a time performed on one thread. A thread is active if at least one active process is performed, otherwise idle. CPt (τ ) is a set of active processes performed on a server st at time τ . Here, the electric power N Et (n) [W] of a server st to concurrently perform n(≥ 0) processes is given as follows [8,10]: [Power consumption for n processes] ⎧ ⎪ ⎪minEt if n = 0. ⎪ ⎪ ⎪ minEt + n · (bEt + cEt + tEt ) if 1 ≤ n ≤ npt . ⎪ ⎪ ⎪ ⎪ ⎪ ⎨minEt + npt · bEt + n · (cEt + tEt ) N Et (n) = if npt < n ≤ nct · npt . ⎪ ⎪ ⎪ minEt + npt · (bEt + nct · cEt ) + ntt · tEt ⎪ ⎪ ⎪ ⎪ ⎪ if nct · npt < n < ntt . ⎪ ⎪ ⎩ maxEt if n ≥ ntt .

(1)

The electric power consumption Et (τ ) [W] of a server st at time τ is assumed to be N Et (| CPt (τ ) |) in this paper. Let minTti show the minimum execution tim of a process pi , i.e. only the process pi is performed on the server st without any other process. Let minTi be a minimum one of minT1i , . . . , minTmi , i.e. minTi = minTf i on the fastest thread which is on a server sf . A server sf with the fastest thread is f astest in a cluster S. We assume one virtual computation step [virtual steps (vs)] is performed on a thread of the fastest server sf for one time unit [sec]. This means, the computation rate T CRf of a thread in a fastest server sf is one [vs/sec]. The thread computation rate T CRt of a server st is (minTi / minTit ) ·T CRf = minTi / minTit [vs/sec]. The maximum computation rate XSCRt (≤ ntt ) of a server st is ntt · T CRt . The total amount V Ci = minTi [sec] ·T CRf [vs/sec] = minTi [vs] of computation is performed by a process pi . The maximum computation rate maxP CRti of a process pi on a server st is V Ci / minTti = minTi / minTti (≤ 1) where only the process pi is performed. Hence, for every pair of processes pi and pj on a server st , maxP CRti = maxP CRtj = T CRt (≤ XSCRt ). The process computation rate N P Rti (n) (≤ T CRt ) [vs/sec] of a process pi on a server st where n processes are concurrently performed at time τ is defined as follows [4,5,9]: [MLCM (Multi-Level Computation with Multiple CPUs) model]  ntt · T CRt / n if n > ntt . N P Rit (n) = (2) T CRt if n ≤ ntt .

An Energy-Efficient Algorithm for Virtual Machines

345

For every pair of processes pi and pj where n (≥ 1) processes are concurrently performed on a server st , N P Rti (n) = N P Rtj (n). Hence, N P Rt (n) means N P Rit (n) for some process pi on a server st . The computation rate P Rt (τ ) of a process a process pi at time τ is assumed to be N P Rt (| CPt (τ ) |). Suppose et pi on a server st starts at time st and ends at time et. Here, τ =st P Rti (τ ) = et τ =st N P Rt (| CPt (τ ) |) = V Ci [vs] = minTi . Thus, minTi shows the amount of computation of a process pi . Each process pi is performed on a server st as follows: [Computation model of a process pi ] 1. At time τ a process pi starts, the computation residue Ri of a process pi is V Ci , i.e. Ri = V Ci (= minTi ); 2. At each time τ , Ri = Ri − N P Rt (|CPt (τ )|); 3. If Ri (≤ 0), the process pi terminates at time τ . The server computation rate N SRt (n) of a server st to perform n processes is n · N P Rt (n), i.e. ntt · T CRt (= XSCRt ) for n > ntt and n · T CRt for n ≤ ntt .

3

Migration Time

As discussed in papers [4,5], if a virtual machine on a host server sh migrates to a guest server sg , the total energy consumed by the servers sh and sg may be reduced. On the other hand, it takes time to make a virtual machine migrate from a host server to a guest server. This means, the execution time of each process on a virtual machine may increase as the virtual machine of the process migrates. We measure time for a virtual machine to migrate from a host server to a guest server. We consider a pair of homogeneous Linux (CentOS 7.5.1804) servers sh and sg , i.e. the same architecture and operating system, as shown in Table 1. The servers sh and sg are interconnected in a network of one [Gbps] [4]. A virtual machine vmk is realized on the servers in KVM [3]. The virtual machine vmk supports one [GiB] virtual memory and ten [GiB] virtual storage. The number n(≥ 1) of processes p1 , . . . , pn are performed on the virtual machine vmk . The processes are created by forking a process p to n (≥ 1) child processes p1 , . . . , pn [9]. No other process is performed on the servers sh and sg . The process p finds prime factors of numbers from 1 to 1,000,000 using the computation loop. The minimum execution time minTi of each process pi on the servers sh and sg is 472 [sec] to find prime numbers of the number, i.e. minTi = 472 [sec] on each of the servers sh and sg . First, the virtual machine vmk is hosted by the host server sh . Before creating the child processes p1 , . . . , pn , the parent process p obtains current time τ by the gettimeof day system call. Then, each process pi waits for (st−τ ) [nsec] until the total specified time st after pi is created. Thus, every process pi starts on the virtual machine vmk at time st. The parent process p waits for termination

346

N. Noaki et al. Table 1. Servers. Server

sh

CPU

Intel Core i5–8400 Intel Core i5–8400

sg

Memory [GB] 8

8

HDD [TB]

1.0

1.0

OS

CentOS 7.5.1804

CentOS 7.5.1804

Fig. 2. Measurement of migration time.

of every child process. The parent process p obtains time et every child process terminates. We measure the execution time of the processes in a pair of cases. In the first case, the virtual machine vmk does not migrate as shown in Fig. 2(1). The processes p1 , . . . , pn on the virtual machine vmk start at time st and terminate at time et2 on the host server sh . In the second case, the virtual machine vmk migrates from the host server sh to the guset server sg in the live migration manner and then migrates back to the host server sh before all the processes p1 , . . . , pn terminate as shown in Fig. 2(2). The virtual machine vmk migrates among the servers sh and sg some number mc of times so that the virtual machine vmk finally backs to the host server sh before every process terminates. Then, every process pi on the virtual machine vmk terminate on the server sh at time et2 . We measure the time st, et1 , and et2 by using the physical clock of the host server sh . Let xt1 be et1 – st and xt2 be et2 – st. The difference xt = xt2 – xt1 gives the migration time mtkhg of the virtual machine vmk among the servers sh and sg is xt/mc. Table 2 shows the execution time xt1 , xt2 , xt = xt2 – xt1 , and mthg . For example, mtkhg = 0.796/2 = 0.398 [sec] for mc = 2. The average migration time mtkhg is 0.405 [sec]. This means, if a virtual machine vmk migrates

An Energy-Efficient Algorithm for Virtual Machines

347

from a host server sh to a guest server sg , it takes additionally about 0.41 [sec] to perform each process on the virtual machine vmk . Table 2. Migration time. mc xt1 [sec] xt2 [sec] xt2 − xt1 [sec] mtkhg [sec]

4 4.1

0

471.944 471.944 0

0

2

471.944 472.740 0.796

0.398

4

471.944 473.513 1.569

0.392

8

471.944 475.340 3.396

0.4245

MDMG (Maximum Energy Consumption Difference by Virtual Machine Migration) Algorithm Selection of a Virtual Machine

Let P be a set of processes p1 , . . . , pn (n ≥ 1) to be issued by clients and S be a set of servers s1 , . . . , sm (m ≥ 1). Let V be a set of virtual machines vm1 , . . . , vml (l ≥ 1) on the servers in the set S. Each server st supports one active virtual machine and some number of idle virtual machines. An active virtual machine can perform processes while an idle virtual machine cannot perform any process. Active virtual machines can migrate among servers in the live manner [3]. First, a client issues a process pi to a load balancer L. The load balancer L selects a host server st in some selection algorithm and then selects an active virtual machine vmk supported by the server st . The process pi is then performed on the virtual machine vmk . Ri shows the computation residue of each active process pi as presented in the computation model. RSt and RV t is the total computation residue of active processes on a server st , i.e. RSt = pi ∈CPt (τ ) Ri  and a virtual machine vmk , i.e. RVk = pi ∈P (vmk ) Ri , respectively. Let nt and nvk be the numbers of active processes on a server st and a virtual machine vmk , respectively. The execution time ETt (RSt , pi ) [time unit] of a server st to perform both all the active processes whose total computation residue is RSt and a new process pi is RSt /N SRt (nt + 1). The energy EEt (RSt , pi ) to be consumed by a server st is ETt (RSt , pi ) [time unit] ·N Et (nt + 1) [W]. The execution time ET Nt (RS, n) [time unit] is RS/N SRt (n) where RS is a computation residue of a server st and n is the number of active processes on the server st . The execution time ETt (RSt , pi ) is ET Nt (RSt , nt +1), since a new process pi is performed in addition to the nt current active processes. For a new process pi issued by a client, the load balancer L selects a host server st where the energy consumption EEt (RSt , pi ) is smallest on the cluster S. Then, an smallest active virtual machine vmh is selected on the server st in the VM selection (VMS) algorithm [Algorithm 1].

348

N. Noaki et al.

Algorithm 1: VMS algorithm 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18

4.2

input : pi = new process issued by a client; S = set of servers ; output : st = host server ; vmk = virtual machine; vmk = ⊥ ; HS = S ; while (HS = φ and vmk = ⊥) do select st in HS where EEt (RSt , pi ) is smallest; Vt = set of active virtual machines on st ; IDt = set of idle virtual machines on st ; if Vt = φ; then select vmk in Vt where nvk is the smallest; else if IDt = φ; then select vmk in IDt ; /* vmk is activated */ Vt = Vt ∪ {vmk }; else HS = HS - {st };

Migration of a Virtual Machine

First, we consider case that no virtual machine migrates from a host server st to a guest server su . Here, a pair of execution time N ETt and N ETu of the servers st nad su are ETt (nt , 0) and ETu (nu , 0), respectively. A pair of the servers st and su consume energy, EEt (RSt , 0) and EEu (RSt , 0), respectively. Suppose N ETt < N ETu , i.e. processes on the server st terminate before the server su . Here, the server st just consumes minEt [W] from time N ETt to time N ETu . Hence, the servers st and su totally consume the energy T EEtu : T EEtu = EEt (RSt , 0) + EEu (RSu , 0) + N T EEtu .

N T EEtu

 (N ETt − N ETu ) · minEt = (N ETu − N ETt ) · minEu

if N ETt ≥ N ETu . otherwise.

(3)

(4)

Next, we consider case a virtual machine vmk migrates from a host server st to a guest server su . Suppose the virtual machine vmk starts migrating from the host server st to the guest server su at time τ . The execution time N ETt of the host server st decreases to M ETt = EN Tt (RSt − RVk , nt − nvk ) = (RSt − RVk )/N SRt (nt − nvk ) since the virtual machine vmk leaves the server st . The virtual machine vmk restarts on the server su at time τ + mtktu . Here, mtktu is the migration time of the virtual machine vmk from the host server st to the guest server su . As presented in the preceding section, mtktu is about 0.41 [sec]. In this paper, we assume mtktu = mt for every pair of servers st and su and

An Energy-Efficient Algorithm for Virtual Machines

349

Fig. 3. Migration of a virtual machine mvk (mt ≤ M ETu ) and M ETt < M ETu ).

Fig. 4. Migration of a virtual machine mvk (mt > N ETu ).

every virtual machine vmk . The computation residue RSu of the guest server su is reduced to αu · RSu at time τ + mt when the virtual machine vmk restarts on the guest server su if N ETu ≥ mt, where αu is (N ETu − mt)/N ETu (≤ 1) [Fig. 3].

350

N. Noaki et al.

The hatched area of Fig. 3 shows the total energy consumption of the servers st and sn . On the other hand, if N ETu < mt, every active process on the guest server su terminates at τ + N ETu before the virtual machine vmk restarts at time τ + mt on the guest server su [Fig. 4]. Hence, the execution time M ETu of the guest server su is given as follows : ⎧ ⎪ ⎨ET Nu ((1 − αu ) · RSu , nu ) + M ETu = (5) ET Nu (αu · RSu + RVk , nu + nvk ) if N ETu ≥ mt. ⎪ ⎩ mt + ET Nu (RVk , nvk ) otherwise. k consumed by the servers st and su is given as The total energy M EEtu follows : ⎧ ⎪ mt · EEu (RSu · (1 − αu ), 0) + ⎪ ⎪ ⎪ ⎪ ⎪ ⎨ ET Nu (αu · RSu + RVk , nu + nvk ) · N Eu (nu + nvk ) + k M EEtu = (6) M T EEtu if N ETu ≥ mt. ⎪ ⎪ ⎪EEu (RSu , 0) + (mt − M ETu ) · minEu + ⎪ ⎪ ⎪ ⎩ ET N (RV , nv ) + M T EE u k k tu otherwise.

M T EEtu

 (M ETt − M ETu ) · minEt = (M ETu − M ETt ) · minEu

if M ETt ≥ M ETu . otherwise.

(7)

In the EVMS algorithm [11], virtual machines on servers migrate to some servers which support more computation rates. In this paper, we propose an M DM G (Minimum energy consumption Difference by virtual machine MiGration) algorithm. Here, a tuple vmk , st , su  of a virtual machine vmk on a host server st and a guest server su is first selected where the difference (T EEtu k ) of the expected energy consumption is the largest in the cluster S. This M EEtu means, the energy consumption of the servers st and su can be mostly reduced by making the virtual machine vmk migrate from the host server st to the guest server su . The virtual machine vmk migrates from the host server st to the guest server su . Then, a next tuple vmk , sh , sg  are selected where a guest server sg is not the guest server su . Thus, a server su to which a virtual machine already migrates is not selected as a guest server of another virtual machine. The MDMS algorithm is shown in Algorithm 2.

An Energy-Efficient Algorithm for Virtual Machines

351

Algorithm 2: MDMG algorithm 1 2 3 4 5 6 7 8 9 10 11

input : S = set of servers; output : {vmk , st , sg |st = host server, su = guest server, vmk = virtual machine on st }; GS = S: while GS = φ do for each server st in S do D = 0; for each active virtual machine vmk in Vt do k select a guest server su (= st ) in GS where (T EEtu - M EEtu ) k > D and (T EEtu - M EEtu ) is largest; if su exists; then k D = T EEtu - M EEtu ; for end;

14

if su exists; then migrate vmk from st to su ; GS = GS - {su }

15

for end;

12 13

16

5

while end;

Evaluation

We evaluate the MDMG algorithm to select a host virtual machine to perform a process issued by a client in terms of the total energy consumption of servers and the average execution time of processes. We consider a cluster composed of four servers s1 , . . . , s4 (m = 4) as shown in Table 3. The fastest thread computation rate T CR1 of the servers1 is one, i.e. T CR1 = 1. For the other servers s2 , s3 , and s4 , T CR2 = 0.8, T CR3 = 0.6, and T CR4 = 0.4. The performance and energy parameters of the servers are shown in Table 2. For example, the server s1 supports the maximam server computation rate XSCR1 = 16 by sixteen threads where the maximum power consumption maxE1 is 230 [W] and the minimum power consumption minE1 is 150 [W]. The server s4 supports XSCR4 = 3.2 by eight threads where maxE4 = 77 and minE4 = 40. The servers s2 and s3 support the same number, twelve threads while maxE2 > maxE1 and minE2 > minE1 . The server s3 is more energy-efficient than the server s2 . Each server st supports two virtual machines in the cluster S. One time unit means one step of the simulation. There are totally eight virtual machines. The migration time mT is one [time unit] which shows 0.41 [sec] based on the measurement discussed in this paper. Each server st is checked every four time units, i.e. md = 4 if an active virtual machine on the server st is to migrate to another server. Let P be a set of processes p1 , . . . , pn (n ≥ 1) to be issued. For each process pi in the set P , the starting time stimei , 0 < stimei ≤ xtime and 10 ≤ minTi ≤ 50.

352

N. Noaki et al.

Here, xtime = 100 [time unit]. At each time τ , if there is a process pi whose {stimei } is τ , one server st is selected by a selection algorithm. The process pi is added to the set Pt of the selected server st , i.e. Pt = Pt ∪ {pi }. For each server st , active processes in the set Pt are performed. The energy variable Et is incremented by the power consumption N Et (| Pt |). If | Pt | = φ, Et is just incremented by minEt . If | Pt |> 0, the variable Tt is incremented by one [time unit] since the server st is active. The variable Tt shows how long the server st is active, i.e. some process is performed. For each process pi in the set Pt , the computation residue Ri of the process pi is decremented by the process computation rate N P Rt (nt ). If Ri ≤ 0, the process pi terminates, i.e. Pt = Pt − {pi } and P = P − {pi }. Until the set P gets empty, the steps are iterated. The variables Et and Tt give the total energy consumption [W · time unit] and execution time [time unit] of each server respectively. In the evaluation, we consider a non-migration (NMG) algorithm [12] and the MDMG algorithm proposed in this paper. In the NMG algorithm, each server st supports only one active virtual machine which just stays on the server st , i.e. does not migrate. For each process pi , a server st where energy consumption EEt (RSt , pi ) is minimum. No virtual machine migrates. The simulator is implemented in SQL on a Sybase [2] database. Information on servers, virtual machines, and processes are stored in tables of the database and the tables are manipulated in SQL. Table 3. Parameters of servers. npt nct ntt T CR XSCR minE maxE pE

cE tE

s1 1

8

16

1.0

16.0

150.0

270.0

40.0 8.0 1.0

s2 1

6

12

0.8

9.6

128.0

200.0

30.0 5.0 1.0

s3 1

6

12

0.6

7.2

80.0

130.0

20.0 3.0 1.0

s4 1

4

8

0.4

3.2

40.0

67.0

15.0 2.0 0.5

Figure 3 shows the total energy consumption of the servers s1 , . . . , s4 for number n of processes. The total energy consumption of the servers of the MDMG algorithm is smaller than the NMG algorithm. The total energy consumption of the servers can be reduced in the MDMG algorithm. Figure 4 shows the average execution time of n processes. The average execution time is shorter in the MDMG algorithm than the NMG algorithm.

An Energy-Efficient Algorithm for Virtual Machines

353

Fig. 5. Total energy consumption.

Fig. 6. Average execution time.

6

Concluding Remarks

It is critical to reduce energy consumption of servers to realize eco society. In this paper, we discussed the virtual machine migration approach that virtual machines on host servers migrate to more energy-efficient guest servers. In this paper, we first measured the migration time mt of a virtual machine among servers Fig. 5. Here, mt is about 0.41 [sec]. By considering the migration time mt, we proposed the MDMG algorithm to make a virtual machine migrate from a host server to a guest server. Here, a virtual machine vmk on each server st is selected to migrate to a guest server su so that the energy consumption of the host server st and the guest server su can be mostly reduced. In the evaluation, we showed the energy consumption of servers and the average execution time of processes can be reduced in the MDMG algorithm compared with the nonmigration NMG algorithm Fig. 6.

354

N. Noaki et al.

References 1. Intel xeon processor 5600 series: The next generation of intelligent server processors, white paper [online] (2010). URL http://www.intel.com/content/www/us/ en/processors/xeon/xeon-5600-brief.html 2. Sybase (2014). URL http://www.cultofmac.com/167829/sybasesap-afaria-offersios-and-pcmanagement-options-mobile-management-month/ 3. A virtualization infrastructure for the linux kernel (kernel-based virtual machine). URL https://en.wikipedia.org/wiki/Kernel-basedVirtualMachine 4. Duolikun, D., Aikebaier, A., Enokido, T., Takizawa, M.: Energy-aware passive replication of processes. Int. J. Mobile Multimedia 9(1,2), 53–65 (2013) 5. Duolikun, D., Kataoka, H., Enokido, T., Takizawa, M.: Simple algorithms for selecting an energy-efficient server in a cluster of servers. Int. J. Commun. Netw. Distrib. Syst. 21(1), 1–25 (2018) 6. Enokido, T., Duolikun, D., Takizawa, M.: Accepted for publication in IEEE transactions on industrial electronics. Execution of Processes. Int. J. Commun. Netw. Distrib. Syst. (IJCNDS), 15(4), 366–385 (2015) 7. Enokido, T., Takizawa, M.: An integrated power consumption model for distributed systems. Accepted for Publication IEEE Trans. Industr. Electron. 15(4), 366–385 (2012) 8. Kataoka, H., Duolikun, D., Enokido, T., Takizawa, M.: Energy-efficient virtualisation of threads in a server cluster. In: Proceeding of the 10th International Conference on Broadband and Wireless Computing, Communication and Applications (BWCCA-2015), pp. 288–295 (2015) 9. Kataoka, H., Nakamura, S., Duolikun, D., Enokido, T., Takizawa, M.: Multi-level power consumption model and energy-aware server selection algorithm. Int. J. Grid Util. Comput. (IJGUC) 8(3), 201–210 (2017) 10. Kataoka, H., Sawada, A., Duolikun, D., Enokido, T., Takizawa, M.: Energy-aware server selection algorithm in a scalable cluster. In: Proceeding of IEEE the 30th International Conference on Advanced Information Networking and Applications (AINA-2016), pp. 565–572 (2016) 11. Noaki, N., Saito, T., Duolikun, D., Enokido, T., Takizawa, M.: Energy-efficient migration of virtual machine. In: The 23nd International Conference on NetworkBased Information Systems (NBiS-2020) accepted (2020) 12. Noguchi, K., Saito, T., Duolikun, D., Enokido, T., Takizawa, M.: An algorithm to select a server to minimize the total energy consumption of a cluster. In: 15-th International Conference on P2P, Parallel, Grid, Cloud and Internet Computing (3PGCIC-2020) accepted (2020)

A Coverage Construction Method Based Hill Climbing Approach for Mesh Router Placement Optimization Aoto Hirata1(B) , Tetsuya Oda2 , Nobuki Saito2 , Masaharu Hirota3 , and Kengo Katayama2 1

2

3

Engineering Project Course, Okayama University of Science (OUS), 1-1 Ridaicho, Kita-ku, Okayama 700–0005, Japan [email protected] Department of Information and Computer Engineering, Okayama University of Science (OUS), 1-1 Ridaicho, Kita-ku, Okayama 700–0005, Japan {oda,katayama}@ice.ous.ac.jp, [email protected] Department of Information Science, Okayama University of Science (OUS), 1-1 Ridaicho, Kita-ku, Okayama 700–0005, Japan [email protected]

Abstract. Wireless mesh networks (WMNs) are one of the wireless network technologies that have received much attention in recent years, and as the name implies, routers can provide a stable network over a wide area by configuring the network like a mesh. In order to provide a lower cost and more stable network, various methods for optimizing the placement of mesh routers are being studied. In a previous work, we proposed a Coverage Construction Method (CCM) for this mesh router placement problem. In this paper, we propose a CCM based Hill Climbing (HC) for mesh router placement optimization problem.

1

Introduction

The Wireless Mesh Networks (WMNs) [1–3] are one of the wireless network technologies that have received much attention in recent years, and as the name implies, routers can provide a stable network over a wide area by configuring the network like a mesh. In order to provide a lower cost and more stable network, various methods for optimizing the placement of mesh routers are being studied. In our previous work [4–8], we proposed and evalueted the different meta-heuristics such as Genetic Algorithms (GA) [9], Hill Climbing (HC) [10], Simulated Annealing (SA) [11], Tabu Search (TS) [12] and Particle Swarm Optimization (PSO) [13] for mesh router placement optimization. Also, we proposed a Coverage Construction Method (CCM) for mesh router placement problem [14]. The CCM is able to rapidly create a group of mesh routers with the radio communication range of all the mesh routers linked to each other. In addition, CCM c The Editor(s) (if applicable) and The Author(s), under exclusive license  to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 355–364, 2021. https://doi.org/10.1007/978-3-030-61108-8_35

356

A. Hirata et al.

Fig. 1. Flowchart of the CCM.

could cover a sufficient number of mesh clients with this solution deployment. Therefore, CCM can increase the number of client coverage by using the results of the solution construction method as the initial solution and then applying the meta-heuristics algorithm. In this paper, we propose a CCM based HC for mesh router placement optimization problem. As evaluation metrics, we considered Size of Giant Component (SGC) for connection between mesh routers and Number of Covered Mesh Clients (NCMC) for mesh clients within radio communication range of SGC of mesh routers. The structure of the paper is as follows. In Sect. 2, we defines the mesh router placement problem. In Sect. 3, we describes proposed system of the CCM and

A CCM Based HC Approach for Mesh Router Placement Optimization

357

CCM based HC. In Sect. 4, we presents the simulation results and compares them with other studies. Finally, conclusions and future work are given in Sect. 5.

2

Mesh Router Placement Problem

In this problem, we are given a two-dimensional continuous area where to deploy a number of mesh routers and a number of mesh clients of fixed positions. The objective of the problem is to optimize a location assignment for the mesh routers to the two-dimensional continuous area that maximizes the network connectivity and mesh clients coverage. Network connectivity is measured by the SGC of the each mesh routers link, while the NCMC is the number of mesh clients that within the radio communication range of at least one mesh router. An instance of the problem consists as follows. • An area W idth × Height where to problem area in mesh router placement. Positions of mesh routers are not pre-determined, and are to be computed. • The mesh routers, each having its radio communication range, defining thus a vector of routers. • The mesh clients located in arbitrary points of the considered area, defining a matrix of clients.

3

Proposed Method

In this section, we present the proposed system. In Fig. 1 and Fig. 2 are shown the flowchart of proposed system. 3.1

CCM for Mesh Router Placement Optimization

In this section, we describe a CCM [14] of our previous proposed algorithm for mesh router placement problem. The proposed method is summarized in Fig. 1. First, randomly generate mesh clients in the problem area. Next, randomly determine a single point coordinate and let it be mesh router 1. Once again, randomly determine a single point coordinate and let it be mesh router 2. Each mesh router has a radio communication range of a radius 2. Consider the mesh routers as circles and perform collision detection for two mesh routers. If the radio communication ranges of the two routers do not overlap, delete router 2 and once again randomly determine a single point coordinate and make it as mesh router 2. This process is repeated until the radio communication range of the two mesh routers overlaps. If the radio communication ranges of the two mesh routers overlap, generate next mesh routers. Determine the collision between the next mesh routers generated and the one generated so far. If there is no overlap in radio communication range with any mesh routers, the next mesh routers generated is removed and generate randomly again. If any of the other mesh routers have overlapping radio communication ranges, generate next mesh routers. Continue this process until the setting number of mesh routers.

358

A. Hirata et al.

Fig. 2. Flowchart of the CCM based HC method for mesh router placement optimization.

A CCM Based HC Approach for Mesh Router Placement Optimization

359

This allows for the creation of a group of routers with all radio communication ranges linked together without the derivation of connected component using Depth First Search (DFS) [15,16]. However, this method only creates a population of mesh routers at set area and does not take into account the location of the mesh clients. So, repeat this process a setting number of loop. When it is repeating, determine how many mesh clients are included in the radio communication range group of the mesh router, and the one with the highest number of coverage in the repetition is the solution. 3.2

CCM Based HC for Mesh Router Placement Optimization

The aforementioned CCM allowed us to generate a group of routers with all radio communication ranges linked together, but it would be difficult to cover all mesh clients with this method. So we put the best results obtained by the CCM as the initial solution and thought that we could get better results by adapting the approximate algorithm. In this paper, we adapt the HC, known as the simplest search algorithm, to the CCM. The implementation of the HC in the mesh router placement problem is shown in Fig. 2. First, we randomly select one of the routers in the group of mesh routers in the initial solution obtained by the CCM and change the coordinates of that chosen mesh router randomly. Then, decision the NCMC by the entire mesh router. If the NCMC is greater than that of the mesh router placement obtained so far, then the changed mesh router placement is the neighbor solution. If the NCMC is less than that of mesh router placement’s NCMC obtained so far, the changed mesh router coordinates are restored. Repeat this process a setting number of loop. This sequence of process is the HC. However, this process alone is inadequate for the mesh router placement problem. This is because, depending on the placement of the changed mesh routers, the radio communication range of all mesh routers will not be linked. Therefore, it is necessary to create an adjacency list for a mesh router each time the mesh router placement is changed and use depth-first search to find out if the radio communication range of all the mesh routers are linked. And, NCMC is decision only when the radio communication range of all mesh routers is linked, and only when NCMC is greater than the neighbor solution, the placement of the mesh routers is the neighbor solution. In this algorithm, in order to increase the probability that all the radio communication ranges are linked, loops the randomly change of coordinates until it overlaps with the radio communication range of one of the mesh routers. We also tightened the conditions for collision detection to cover as many clients as possible. Specifically, the collision is recognized only when the sum of the square of the difference between the x-axis of the two routers and the square of the difference between the y-axis is bigger than 12 and less than 16.

360

A. Hirata et al. Table 1. Parameters and Value for Simulation

Width of problem area

32

Hight of problem area

32

Number of mesh routers

16

Radius of radio communication range of mesh 2 routers Number of mesh clients

48

Distributions of mesh clients

Normal distribution and uniform distribution

Number of loop for CCM

3000

Number of loop for HC method

1000000

Table 2. Comparison of best SGC considering another method. Distributions of mesh clients

Initial solution CCM

CCM based Xhafa et al.’s HC method [17]

Normal distribution

16

16

16

16

Uniform distribution 16

16

16

16

Table 3. Comparison of avg. SGC considering another methods. Distributions of mesh clients

Initial solution CCM

CCM based Xhafa et al.’s HC method [17]

Normal distribution

16

16

16

16

Uniform distribution 16

16

16

16

Table 4. Comparison of best NCMC considering another method. Distributions of mesh clients

Initial solution CCM

CCM based Xhafa et al.’s HC method [17]

Normal distribution

7

46

48

27

Uniform distribution 7

18

26

9

Table 5. Comparison of avg. NCMC considering another methods. Distributions of mesh clients

Initial solution CCM

CCM based Xhafa et al.’s HC method [17]

Normal distribution

7

43

47

26

Uniform distribution 7

14

20

9

32 31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0

Mesh Clients Mesh Router Communication Range

Y-AXIS [unit]

Y-AXIS [unit]

A CCM Based HC Approach for Mesh Router Placement Optimization

0

1

2

3

4

5

6

7

8

9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32

32 31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0

Mesh Clients Mesh Router Communication Range

0

1

2

3

4

5

6

7

8

9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32

X-AXIS [unit]

X-AXIS [unit]

(a) Result of initial solution.

Y-AXIS [unit]

361

(b) Result of CCM.

32 31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0

Mesh Clients Mesh Router Communication Range

0

1

2

3

4

5

6

7

8

9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32

X-AXIS [unit]

(c) Result of CCM based HC.

Fig. 3. Visualization results in Normal distribution.

4

Simulation Results

In this section, we evaluate the proposed method. The parameters used in the simulation are shown in the Table 1. In this simulation, we ran two types of clients, one using a Normal distribution and the other using a Uniform distribution, 15 times each. The simulation results are summarized in Table 2, Table 3, Table 4 and Table 5. In Table 2 and Table 3, we show the simulation results of best SGC and avg. SGC. For each simulation results, all mesh router nodes are connected. In Table 4 and Table 5, we show the simulation results of best NCMC and avg. NCMC. When the Normal distribution of mesh clients is used to generate mesh clients, we are able to cover all mesh clients 13 times out of 15 runs. We can find that all mesh routers are connected and all mesh clients covered by mesh routers in best results. Also, more than 90 [%] of the clients are covered on average when the solution construction method is used to generate the initial solution. For Uniform distribution of mesh clients, we can confirm that about 30

A. Hirata et al. 32 31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0

Mesh Clients Mesh Router Communication Range

Y-AXIS [unit]

Y-AXIS [unit]

362

0

1

2

3

4

5

6

7

8

9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32

32 31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0

Mesh Clients Mesh Router Communication Range

0

1

2

3

4

5

6

7

8

9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32

X-AXIS [unit]

X-AXIS [unit]

Y-AXIS [unit]

(a) Result of initial solution.

(b) Result of CCM.

32 31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0

Mesh Clients Mesh Router Communication Range

0

1

2

3

4

5

6

7

8

9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32

X-AXIS [unit]

(c) Result of CCM based HC.

Fig. 4. Visualization results in Uniform distribution.

[%] of the mesh clients are covered on average when the CCM is used to generate the initial solution. The CCM based HC is can cover about 40 [%] of the mesh clients on average. In both distributions, we compair the Xhafa et. al.’s method [17]. For the best results in Normal distribution of mesh clients, the simulation results are shown 178 [%] better than Xhafa et al.’s method. For the best results in Uniform distribution of mesh clients, the simulation results are shown 289 [%] better than Xhafa et al.’s method. The visualization results is shown in Fig. 3 and Fig. 4. In Fig. 3 and Fig. 4, we can see that the mesh routers group generated by the CCM are able to expand its coverage and cover more mesh clients by applying the HC.

A CCM Based HC Approach for Mesh Router Placement Optimization

5

363

Conclusion

In this paper we propose and implement a CCM based HC for mesh router placement optimization. For simulation results, we can find that all mesh routers are connected and all mesh clients covered by mesh routers in Normal distribution. In the future, we would like to apply it to the Simulated Annealing, which is an advanced version of the CCM based HC. In addition, we would like to adapt our method to the genetic algorithm which is often mentioned as a solution to the mesh router placement problem. We would also like to challenge what kind of change is caused by changing the parameters used in the experiments. Acknowledgement. This work was supported by JSPS KAKENHI Grant Number 20K19793.

References 1. Akyildiz, I.F., et al.: Wireless mesh networks: a survey. Comput. Netw. 47(4), 445–487 (2005) 2. Jun, J., et al.: The nominal capacity of wireless mesh networks. IEEE Wireless Commun. 10(5), 8–15 (2003) 3. Oyman, O., et al.: Multihop relaying for broadband wireless mesh networks: from theory to practice. IEEE Commun. Mag. 45(11), 116–122 (2007) 4. Oda, T., et al.: WMN-GA: a simulation system for WMNs and its evaluation considering selection operators. J. Ambient Intell. Humanized Comput. 4(3), 323– 330 (2013) 5. Ikeda, M., et al.: Analysis of WMN-GA simulation results: WMN performance considering stationary and mobile scenarios. In: Proc. of The 28-th IEEE International Conference on Advanced Information Networking and Applications (IEEE AINA-2014), pp. 337–342. IEEE (2014) 6. Oda, T., et al.: Analysis of mesh router placement in wireless mesh networks using Friedman test considering different meta-heuristics. Int. J. Commun. Netw. Distrib. Syst. 15(1), 84–106 (2015) 7. Oda, T., et al.: A genetic algorithm-based system for wireless mesh networks: analysis of system data considering different routing protocols and architectures. Soft Comput. 20(7), 2627–2640 (2016) 8. Sakamoto, S., et al.: Performance evaluation of intelligent hybrid systems for node placement in wireless mesh networks: a comparison study of WMN-PSOHC and WMN-PSOSA. In: Proc. of The 11-th International Conference on Innovative Mobile and Internet Services in Ubiquitous Computing (IMIS-2017), pp. 16–26 (2017) 9. Holland, J.H.: Genetic algorithms. Sci. Am. 267(1), 66–73 (1992) 10. Skalak, D.B.: Prototype and feature selection by sampling and random mutation hill climbing algorithms. In: Proc. of The 11-th International Conference on Machine Learning (ICML-1994), pp. 293–301 (1994) 11. Kirkpatrick, S., et al.: Optimization by simulated annealing. Sci. 220(4598), 671– 680 (1983) 12. Glover, F.: Tabu search: a tutorial. Interfaces 20(4), 74–94 (1990)

364

A. Hirata et al.

13. Kennedy, J., Eberhart, R.: Particle Swarm Optimization. In: Proc. of The IEEE International Conference on Neural Networks (ICNN-1995), pp. 1942–1948 IEEE (1995) 14. Hirata, A., et al.: Approach of a solution construction method for mesh router placement optimization problem. In: Proc. of The IEEE 9-th Global Conference on Consumer Electronics (IEEE GCCE-2020), accepted to appear, pp. 1–2. IEEE (2020) 15. Tarjan, R.: Depth-first search and linear graph algorithms. SIAM J. Comput. 1(2), 146–160 (1972) 16. Lu, K., et al.: The depth-first optimal strategy path generation algorithm for passengers in a metro network. Sustain. 12(13), 1–16 (2020) 17. Xhafa, F., et al.: Solving mesh router nodes placement problem in Wireless Mesh Networks by Tabu Search algorithm. J. Comput. Syst. Sci. 81(8), 1417–1428 (2015)

Review of Intelligent Data Analysis and Data Visualization Kang Xie1(B) , Linshan Han2 , Maohua Jing3 , Jingmin Luan3 , Tao Yang1 , and Rourong Fan1 1 Key Lab of Information Network Security of Ministry of Public Security,

The Third Research Institute of the Ministry of Public Security, Shanghai, China [email protected] 2 Software College, Northeastern University, Shenyang, China [email protected] 3 School of Computer and Communication Engineering, Northeastern University at Qinhuangdao, Qinhuangdao, China [email protected]

Abstract. In recent years, machine learning, deep learning, neural network and other artificial intelligence technologies have made significant development and breakthroughs. Thus more mature and reliable technologies start to be applied in the field of data analysis. Data analysis is the process of studying and summarizing data in detail in order to extract useful information, that is, collecting, sorting, processing and analyzing data. Moreover, in order to make the analysis results easier to understand, data visualization technology is widely employed. This paper tries to review the state-of-the-art data analysis methods based on artificial intelligence. Besides, the mainstream data visualization technology and cases are elaborated. Finally, some open problems and challenges are also put forward for the future research. Keywords: Graph · Data analysis · Data modeling · Data representation · Data extraction

1 Introduction With the breakthrough of genetic engineering, virtual reality, artificial intelligence, Internet of things, quantum information technology, clean energy and biotechnology, that is, the era of using information technology to promote industrial change is the era of intelligence. In response, China launched the “Made in China 2025” plan, the main line of which is the digital networked intelligent manufacturing that reflects the deep integration This work was supported by the National Key Research and Development Program of China (Project No. 2018YFC0806903,2018YFC0820104) and Key Lab of Information Network Security of Ministry of Public Security C19606 (The Third Research Institute of Ministry of Public Security). © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 365–375, 2021. https://doi.org/10.1007/978-3-030-61108-8_36

366

K. Xie et al.

of information and manufacturing technology [1]. It can be expected that the realization of intelligent data analysis in terms of artificial intelligence and other technologies will further reflect its value. With the rapid development of the Internet of things, cloud computing, neural networks and other technologies, information technology and political, economic, military, scientific research, life and other fields continue to cross-integrate, which has generated huge amounts of data beyond any previous era, making the world into the data era relying on the Internet [2]. Various websites, applications, mobile devices, etc. in different fields around the world are generating huge data traffic at all times, which has spurred and promoted the development of related industries relying on data, and has also presented severe challenges for data analysis and processing. In this regard, the intelligent analysis of data has become a research hotspot. Intelligent data analysis can be divided into two directions: the use of artificial intelligence technology for data analysis and mining, and the adoption of visualization technology to cooperate with data analysis and human-computer Interaction. This paper describes the research direction of intelligent data analysis based on machine learning and deep learning together with some classic algorithms. Moreover, the design methods of data visualization and some reliable data visualization tools are also elaborated. Finally, the open problems and challenges are presented for the future research.

2 Data Analysis Based on Artificial Intelligence Artificial intelligence technology is an important method for big data analysis. Machine learning and computational intelligence are important analysis methods in artificial intelligence technology. Machine learning technology is an important artificial intelligence technology that can perform data mining from data clustering, data association analysis, data classification, data prediction, and other aspects to achieve effective data analysis and processing. Deep learning is currently the most important machine learning method. It can be combined with the Hadoop, MapReduce, and Spark frameworks for distributed computing analysis. 2.1 Big Data Analysis Based on Machine Learning Machine learning is the most important branch of artificial intelligence and the most important method in big data analysis. According to the main tasks of big data mining, big data analysis based on machine learning is divided into four aspects: big data clustering, big data association analysis, big data classification, big data prediction. 1) Big data clustering: By using parallel implementation or improvement of existing clustering algorithms to solve the problems of large data volume and high complexity of traditional clustering algorithms in the era of big data, important research has been made based on the MapReduce framework and K-means algorithm. Parallel operations can be realized based on the MapReduce framework, and analysis speed

Review of Intelligent Data Analysis and Data Visualization

367

and data processing efficiency can be improved through data deduplication technology and cluster analysis technology [3]. Combined with the feature weighting method, a weighting k-means clustering approach by integrating Intra—Cluster and Inter-Cluster distances (KICI) can be proposed to solve the problem of clustering high-dimensional data and obtain better clustering results [4]. 2) Big data association analysis. Association analysis: also known as association mining, is to find frequent patterns, associations, correlations, or causal structures between items or objects in various data. Currently, the most commonly used association analysis algorithms include Apriori association rule mining (breadth first) and frequent pattern growth (FP-Growth) algorithm association rule mining (depth first). The Apriori algorithm performs filtering based on the set support threshold, and then finds all frequent itemsets and obtains the corresponding association rules. The Apriori algorithm has the disadvantages of large time overhead and a large number of candidate frequent sets, so the adaptability in breadth and depth when processing large data is not good. Parallelizing Apriori algorithm based on MapReduce or Spark is an important strategy to improve computing efficiency [5, 6]. The FP-Growth algorithm uses a divide-and-conquer strategy. By constructing a frequent pattern tree (FP-tree) and mining frequent patterns on the FP-tree, the algorithm only needs to scan the data set twice to obtain the association rules contained in the data set. The FP-Growth algorithm avoids the large number of candidate sets generated by the Apriori algorithm. The calculation and grouping for the FP-Growth algorithm can further improve the distributed and parallel processing process, and shorten the calculation time [7, 8]. 3) Big data classification: Big data classification is an important method in big data mining in applications and industries. The classic methods include: random forest based on decision tree model, cost-sensitive learning strategies, extreme learning machines (ELM), K-nearest neighbor classifier, discriminant analysis, Bayesian, neural networks, etc. [9–13]. 4) Big data prediction: Big data prediction based on machine learning can be applied to many industries, such as breast cancer survival prediction model in medical treatment [14], E-GDP prediction based on big data of electric power [15], wear predication of tunnel boring machine cutters during mining [16]. The difficulty of big data prediction is to further improve the efficiency and performance of the algorithm, ensure the accuracy of data prediction, and control the cost of prediction. 2.2 Big Data Analysis Based on Deep Learning Deep learning is one of the most important machine learning methods, and it is widely used in the fields of image, speech, and natural language processing [17]. It takes a large number of iterative calculations to determine various types of parameters during the computation-intensive deep learning model training. Because the amount of data for deep learning training is very large and the entire training process is time-consuming, supporting distributed computing platforms such as MapReduce and Spark has become an important research direction to reduce the training time cost of deep learning. MapReduce’s distributed computing framework is conducive to processing largescale data, and has achieved certain results in the basic model of deep learning: Deep

368

K. Xie et al.

Belief Nets (DBNs) based on the MapReduce framework have been implemented and training time has been shortened [18], a distributed algorithm based on the MapReduce framework to implement an effective mapping mechanism back propagation (BP) for the Map phase [19], optimizes the initial weights and thresholds of the BP neural network and implements the parallel processing of the BP neural network [20], distributed classification model and mining operator of streaming big data based on MapReduce framework, etc. [21]. Compared with MapReduce, the Spark distributed computing platform based on memory is more conducive to iterative computing, and has achieved many research results: Accelerating deep learning training process based on Real-valued Conditional Restricted Boltzmann Machine (R_CRBM) model based on Spark distributed computing platform [22], the convolutional neural network was implemented through the Spark platform which improved the classification performance and training time of the ImageNet dataset [23], distributed computing of neural network training based on Spark distributed framework and deep learning tool Deeplearning4j, which shortens the training time and further expands deep learning models [24]. Improving the quality of training samples, accelerating the training process of deep networks, and optimizing the performance of deep learning models are still the key issues of deep learning algorithms. In this regard, the deep learning-based data analysis needs to solve the following problems: improve the number and quality of big data training samples to solve problems such as lack of samples, insufficient sample quality, improve the deep learning model itself to shorten the training of deep networks Time, combined with other models to improve the training model to improve the performance of the algorithm, etc.

3 Data Visualization Data visualization technology can display the results of data analysis according to human cognitive rules and analysis methods to assist human-computer interaction. Since about 80% of the information that humans get from the outside comes from the visual system, when data can be visually displayed to the analysts in the form of graphics through data visualization technology, analysts can clearly understand the information hidden behind the data [2]. 3.1 Visual Color Design In order for the hidden information in the data to be fully perceived and understood by the analyst, it is necessary to appropriately select and combine visual elements, among which color is one of the most important elements [25]. The use of color coding can realize data information storage, and at the same time, according to human visual habits, colors can be used to highlight key information and display data associations and differences. In order to describe color mathematically, the concept of color space is proposed. Color space is also called color model or color coordinate system. According to the nature of the color space, the commonly used color spaces can be classified as follows: RGB,

Review of Intelligent Data Analysis and Data Visualization

369

Fig. 1. Commonly used color spaces:RGB, HSV, HSL, CIE [25]

CMYK and other device-dependent color spaces, HSL, HSV and other user-oriented color spaces, CIEL *a*b*, CIEL *u*v* perception uniform color space (Fig. 1). When designing data visualization, it is necessary to pay attention to the visual stimulation and visual perception of the analyst by color. Emphasizing colors on the main data can guide analysts to important information. By using different colors to distinguish different data information, it is easy for analysts to distinguish and compare. Ensure the legibility of colors, avoid using high-contrast colors, fill large-size areas with low-saturation colors and fill small-size areas with high-saturation colors, and control the number of colors used at the same time to less than 7 for easy understanding by analysts. Finally, in order to avoid color illusion and mislead the analyst’s understanding, it is necessary to adopt a combination of color and semantics in the analyst’s common sense. 3.2 Data Visualization Tools With the explosion of data information and the advent of the era of cloud and big data, data visualization must not only quickly collect, screen, analyze, summarize, and display data, but also need to have the ability to be real-time, simple to operate, rich in display methods, and support multiple data integration methods [26]. In response, many universities, such as MIT, Stanford University, Peking University, and Zhejiang University have set up data visualization laboratories. At the same time, major academic conferences have also been carried out, such as IEEE VIS, the most authoritative international academic conference in the field of data visualization, composed of IEEE Visual Analytics Science and Technology (VAST), IEEE Information Visualization (InfoVis), and IEEE Scientific Visualization (SciVis) [27]. Commonly used open source visualization tools include: Processing, ColorBrewer, R, Google Chart, Impact, Unfolding, D3.js, Envision.js, RAWGraphs, Google Fusion Tables, etc. The R language is mainly used for static mapping, which is suitable for data visualization of statisticians and academia. Interactive tools such as Processing and D3.js are more suitable for interactive data display and public-facing scenarios. However, open source data visualization tools usually require a certain programming ability, which is not conducive to resource sharing. Therefore, many companies’ teams have launched commercial data visualization tools, such as Tableau, Excel, CartoDB, Visual.ly, Splunk, etc. Excel is one of Microsoft’s office software which can conveniently display data and create visual charts, but its styles and colors are relatively single. Tableau is a professional business data visualization tool with rich templates and easy operations, which is good for business promotion.

370

K. Xie et al.

3.3 Data Visualization Methods For different types of data, different types of data visualization techniques can be used. For example: For unstructured comics or natural language, semantic visualization can be combined with data charts (Fig. 2) [28]. The space relationship method, node link graph, heat map, adjacency matrix, etc. can be used to reflect the network relationship of the data structure [29], and force-directed layout is one of the most common layouts for visual representation of node links (Fig. 3) [30]. For widely existing multi-dimensional data, such as spatio-temporal data, business data, data in traditional relational databases, etc., multi-dimensional data can be visualized using parallel coordinates (Fig. 4), scatter plots, projections, etc., to reveal the data distribution rules and the implicit relationship between attributes of different dimensions.

Fig. 2. Visualization by Galán [28]

Fig. 3. Node-link tree graph visualization [31]

Review of Intelligent Data Analysis and Data Visualization

371

Fig. 4. Time-series parallel coordinates of multidimensional data for infectious disease transmission [32]

When the size of the processed data is too large and the data association is complex, the phenomenon of aggregation, clutter, and coverage will occur during data visualization. In order to facilitate the analyst’s identification, the data visualization method needs to be improved. One method is to simplify the graph, such as the method of edge bundling of edges to make the network data visualization clear (Fig. 5), or transform large-scale graphs into hierarchical tree structures through hierarchical clustering, and visualize graphs at different levels through multiscale interaction (Fig. 6). At the same time, multiple types of data visualization methods can be combined to data visualize from different dimensions (Fig. 7).

Fig. 5. Stream map with edge-bundled time event stream [33]

372

K. Xie et al.

Fig. 6. Data visualization based on hierarchical clustering and multi-scale interactions [34]

Fig. 7. NodeTrix model: node links show global structure, adjacency matrix shows local [29]

4 Challenges and Open Problems In the past years, intelligent data analysis and data visualization technology have become more and more hotspots of research. Many researches have been successfully proposed, but there are still many challenges. For intelligent data analysis, when big data provides people with unprecedented large-scale data, it also challenges people in extracting key information from mass data and processing the association between information. The complexity of data information has greatly restricted people’s ability to design efficient computing models and methods for big data. The characteristics of multi-source heterogeneity, large scale, and fast changeability of big data continue to impact on previous machine learning and deep

Review of Intelligent Data Analysis and Data Visualization

373

learning algorithms. In addition, the sample size of big data is sufficient, the internal relationships are close and complex, and the value density distribution is extremely uneven. These characteristics provide opportunities and challenges for studying the computability of big data and establishing a new computing paradigm. It is the focus of research to propose algorithms with reasonable cost and accuracy in the face of computational complexity. At the same time, the compatibility of big data analysis in different fields and platforms and the huge scale, complex structure, and sparse value of data characteristics have caused great research challenges for the performance evaluation and optimization of big data processing systems. Through the iterative improvement of design, implementation, and verification, the ultimate realization of high-throughput data acquisition, low data storage energy consumption, and high data calculation efficiency in big data computing systems is the important goal of research. For data visualization technology, ensure the integration and interface design of multi-source, heterogeneous, non-complete, non-consistent, and inaccurate data, maximize the advantages of human and machine and optimize cooperation with humanmachine interaction, designed frameworks, tools, and their extensibility, will become a challenge for data visualization technology. Therefore, ensuring that filtering and sampling are just right, designing deep learning methods for faster training, ensuring visual data and displaying high accuracy while improving real-time visual response will be the focus of data visual design. At the same time, with the maturity of 3D modeling and AR and VR technologies, how to combine data visualization technology with these emerging technologies and propose corresponding algorithms is likely to bring more opportunities for data visualization technology research.

5 Conclusion Based on artificial intelligence technologies, such as machine learning and deep learning, the performance of data analysis can be greatly improved. Besides, the data visualization technology presents the analysis results through human-computer interaction and conforms to human cognitive rules. Facing the opportunities brought by technological updates, the optimization of data analysis algorithms and the improvement of data visualization effects will become the focus of future research.

References 1. Genyuan, D., Genyuan, Z.: Introduction to Information Technology. Wuhan University Press, (in Chinese), 09, p. 222 (2015) 2. Lei, R., Yi, D., Shuai, M.: Review of big data visual analysis. J. Softw. 25(09), 1909–1936 (2014). (in Chinese) 3. Pang, S.: Analysis of multidimensional data de-duplication clustering algorithm based on large data mining. Mod. Electron. Tech. 42(23), 150–153 (2019) 4. Xiaohui, H., Cheng, W., Liyan, X.: A weighting k-means clustering approach by integrating intra-cluster and inter-cluster distances. Chin. J. Comput. 42(12), 2836–2848 (2019) 5. Ezhilvathani, A., Raja, K.: Implementation of parallel apriori algorithm on Hadoop cluster. Int. J. Comput. Sci. Mobile Comput. 2(4), 513–516 (2013)

374

K. Xie et al.

6. Qiu, H., Gu, R., Yuan, C.: Yafim: a parallel frequent itemset mining algorithm with spark. In: 2014 IEEE International Parallel & Distributed Processing Symposium Workshops, Washington, D.C., USA, pp. 1664-1671. IEEE (2014) 7. Jie, W., Qinghao, D., Zeng, Y.: Optimization of parallel frequent modulus growth algorithm in cloud manufacturing environment. Comput. Integr. Manuf. Syst. 18(9), 2124–2129 (2012). (in Chinese) 8. Zhiyong, L.: Research on Parallel Algorithm for Mining Association Rules. Southeast University, Nanjing (2016). (in Chinese) 9. Singh, S., Garg, R., Mishra, P.: Review of apriori based algorithms on mapreduce framework, pp. 593–604. IEEE, Washington, D.C., USA (2014) 10. Lopez, V., Del R.S., Benitez, J.M.: Cost sensitive linguistic fuzzy rule based classification systems under the MapReduce framework for imbalanced big data. Fuzzy Sets Syst. 258, 5–38 (2015) 11. Huang, G., Huang, G.B., Song, S.: Trends in extreme learning machines: a review. Neural Netw. 61, 32–48 (2015) 12. Kamal, S., Ripon, S.H., Dey, N.: A MapReduce approach to diminish imbalance parametersfor big deoxyribonucleic acid dataset. Comput. Methods Programs Biomed. 131, 191–206 (2016) 13. Fernandez-Delgado, M., Cernadas, E., Barro, S.: Do we need hundreds of classifiers to solve real world classification problems. J. Mach. Learn. Res. 15(1), 3133–3181 (2014) 14. Huiying, Q., Yuhe, J.: Predicting breast cancer survival length with multi-omics data fusion. Data Anal. Knowl. Discov. 3(8), (2019). (in Chinese) 15. Shiming, T., Taorong, G., Xiaoqing, H.: Forecasting regional E-GDP value using power big data. Electric Power Autom. Equipment, 39(11) (2019). (in Chinese) 16. Jihua, Y., Changbin, Y.: Wear predication of tunnel boring machine cutters based on in-situ measured data. J. Southwest Jiaotong Univ. 54(6), 1283–1292 (2019). (in Chinese) 17. Wanling, W.: Artificial Intelligence and its Applications (3rd ed). Higher Education Press, Beijing (2016). (in Chinese) 18. Zhang, K., Chen, X.W.: Large-scale deep belief nets with mapreduce. IEEE Access 2, 395–403 (2014) 19. Zhang, H.J., Xiao, N.F.: Parallel implementation of multilayered neural networks based on MapReduce on cloud computing clusters. Soft. Comput. 20(4), 1471–1483 (2016) 20. Cao, J., Cui, H., Shi, H.: Big data: a parallel particle swarm optimization-back-propagation neural network algorithm based on MapReduce. PLoS ONE 11(6), e0157551 (2016) 21. Guojun, M., Dianjun, H., Songyan, X.: Big data classification model and algorithm based on distributed data flow. Chin. J. Comput. 40(1), 161–175 (2017). (in Chinese) 22. Jingyue, H., Bei, M.: Collaborative filtering recommendation algorithm using real-valued conditional Boltzmann machine with social relations. Chin. J. Comput. 39(1), 183–195 (2016) 23. Li, H., Su, P., Chi, Z.: Image retrieval and classification on deep convolutional SparkNet. In: 2016 IEEE International Conference on Signal Processing, Communications and Computing (ICSPCC), Washington, D.C., USA, pp. 1–6. IEEE (2016) 24. Yan, Y., Chen, M., Sadiq, S.: Efficient imbalanced multimedia concept retrieval by deep learning on spark clusters. Int. J. Multimedia Data Eng. Manage. 8(1), 1–20 (2017) 25. Huan, Y., Yinan, L., Kang, Y.: Proper use of color in visualization. J. Comput. Aided Des. Comput. Graph. 27(9), 1587 (2015) (in Chinese) 26. Ran, X.: TOP50 + 5 big data visualization analysis tool. Internet Wkly 17, 58–59 (2014). (in Chinese) 27. Xiaoyan, C., Liping, G., Wenpin, G.: Comparison and application of big data visualization tools. Comput. Educ. 06, 97–102 (2018). (in Chinese) 28. Kadembo, E.M.: Anchored in the story: the core of human understanding, branding, education, socialisation and the shaping of values. Mark Rev. 12(3), 221–231 (2012)

Review of Intelligent Data Analysis and Data Visualization

375

29. Yongchao, W., Shengwen, L., Yingbao, Y.: Review of knowledge graph visualization. J. Comput. Aided Des. Comput. Graph. 31(10), 1666–1676 (2019). (in Chinese) 30. Dudáš, M., Lohmann, S., Svátek, V., et al.: Ontology visualization methods and tools: a survey of the state of the art. Knowl. Eng. Rev. 33, e10 (2018) 31. Hussain, A., Latif, K., Rextin, A.T., et al.: Scalable visualization of semantic nets using power-law graphs. Appl. Math. Inf. Sci. 8(1), 355–367 (2014) 32. Sichen, J., Yubo, T., Yuyu, Y., et al.: Analysis of infectious disease patterns based on multidimensional spatiotemporal data visualization. J. Comput. Aided Des. Comput. Graph. 31(2), 241–255 (2019). (in Chinese) 33. Buchin, K., Speckmann, B., Verbeek, K.: Flow map layout via spiral trees. IEEE Trans. on Visual. Comput. Graph. 17(12), 2536–2544 (2011) 34. Abello, J., van Ham, F., Krishnan, N.: ASK-Graphview: a large scale graph visualization system. IEEE Trans. on Visual. Comput. Graph., 12(5), 669–676 (2006)

Data Analysis Based on Knowledge Graph Kang Xie1(B) , Qizhen Jia2 , Maohua Jing3 , Qilong Yu3 , Tao Yang1 , and Rourong Fan1 1 Key Lab of Information Network Security of Ministry of Public Security,

The Third Research Institute of the Ministry of Public Security, Shanghai, China [email protected] 2 Software College Northeastern University, Shenyang, China [email protected] 3 School of Computer and Communication Engineering, Northeastern University at Qinhuangdao, Qinhuangdao, China

Abstract. In the past few years, the automatic construction of large-scale knowledge graphs has received extensive attention from academia and industry. In this paper, we show the specific applications and typical cases of knowledge graphs in data analysis, and summarize the overall process of knowledge graphs from construction to data analysis. Moreover, some open problems and challenges are also presented for the future research. Keywords: Data analysis · Data visualization · Artificial intelligence · Data mining

1 Introduction The concept of the Knowledge Graph was firstly evolved from “The Semantic Web”, conceived by Berners-Lee, the father of the Web [1]. The purpose is to express the interrelationship between entities in the world in the form of graphs, so as to effectively search for entities accurately. In 2012, Google formally proposed the concept of Knowledge Graph, and announced that it would build a next-generation intelligent search engine. From the perspective of intelligent search with Knowledge Graph, the need is no longer a simple string matching, but making inferences based on the user’s query situation, thus producing more structured and hierarchical search results. With the booming of the concept of Knowledge Graph, the related research and applications gain rapid development, and more technologies such as knowledge answering, social networking, intelligent search are put forward. To this end, this paper summaries the state-of-the-art and technical content of Knowledge Graph in the field of data analysis, hoping to lay a solid foundation for the future research and provide some benefits to the applications. This work is supported by the National Key Research and Development Program of China (Project No. 2018YFC0806903, 2018YFC0820104) and the Key Lab of Information Network Security of Ministry of Public Security C19606 (The Third Research Institute of Ministry of Public Security). © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 376–385, 2021. https://doi.org/10.1007/978-3-030-61108-8_37

Data Analysis Based on Knowledge Graph

377

This paper first introduces the current applications of knowledge Graph, briefly describes the definition, performs simple analysis and displays of the technique processes. The state-of-the-art research of data analysis based on Knowledge Graph are then elaborated. Finally, some open problems and challenges are proposed for the future research.

2 Preliminaries of Knowledge Graph 2.1 Definitions Knowledge Graph/Vault is also called scientific knowledge graph. In the library and information world, it is called knowledge domain visualization or knowledge area projection map. Knowledge Graph is a series of various graphs showing the development process of knowledge and their relationship by applying the visualization technique to describe knowledge resources and their carriers [2]. The technique process of knowledge graph involves multiple aspects such as knowledge acquisition, knowledge processing, knowledge analysis, and knowledge reasoning. The general process is as follows: first, knowledge modeling, which is equivalent to the definition of the table structure of a relational database, second, knowledge extraction of unstructured and semi-structured data [3], third, extraction, alignment and merging knowledge from different sources to form a globally unified knowledge representation and correlation, and finally, knowledge reasoning for knowledge discovery, conflict and anomaly detection to complete the construction of the knowledge graph. 2.2 The State-of-the-Art Research Since the introduction of the Semantic Web by Tim Berners Lee in 1998, a large number of knowledge bases have appeared on the Internet. In 2010, Google acquired Freebase, an open shared one, and used it as one of the main data sources for the Knowledge Graph. In 2016, Google closed Freebase’s data and moved all its data to the current Wikidata [4]. Wikidata’s goal is to build a free, open, multilingual, large-scale connected knowledge base that anyone can edit and modify. As of 2018, Wikidata has more than 50 million knowledge entries [5]. Babel is a multilingual knowledge base similar to WordNet, which mainly solves the problem of lack of data in non-English languages; NELL is a knowledge base developed by Carnegie Mellon University, whose role is to give an ontology and a small number of samples to allow the machine to learn new knowledge by itself; Yago is a linked database of Max Planck research in Germany [6], integrating three databases of Wikpedia, WordNet and GeoNames. In China, OpenKG is a community project for the open knowledge graph, which is to promote the openness and interconnection of knowledge. Typical encyclopedias include Zhishi.me (Tarsatola Science and Technology, Southeast University), CNDBpedia (Fudan University), Xlore (Tsinghua University), Belief-Engine (Institute of Automation, Chinese Academy of Sciences), and PKUpie (Peking University). The knowledge graph of vertical domain is relative to the above-mentioned general encyclopedia ones [7]. It is a knowledge graph oriented to a specific field, such as law,

378

K. Xie et al.

medical, e-commerce, and so on. Compared with the general knowledge graphs, the one of vertical domain is larger, its structure is more complex, and the data sources are more diverse and more heterogeneous. Therefore, the construction of the knowledge graph of vertical domain is one of the hot topics in the research direction in the future. 2.3 Applications Knowledge Graph was first applied to intelligent search engines, and has now been widely applied to assisted search, knowledge answering, big data analysis and other scenarios, through the strong ability of processing and reasoning the unstructured, structured, and semi-structured data. Assisted Search. Traditional search is to perform string matching on webpage targets through the Internet and return matching results. The process in which knowledge graph is used to search the semantics, and the keywords entered by users are projected into the entity concepts of the knowledge graph, and then the reasoning and search are performed according to the concept hierarchy in the knowledge graph to return data to the user, not only directly displays of items, but also present the relationship between them [8]. For example, when you search for the keyword “Yao Ming” in the Google search engine, it will automatically help you to show Yao Ming’s identity, date of birth, works, achievements, wife, etc. through a knowledge graph with “Yao Ming” as the entity. Knowledge Answering. The knowledge question answering system is an intelligent system. First of all, it analyzes the user’s questions semantically and grammatically. Then it receives questions expressed by people in natural language through dialogues, understands people’s intentions, queries and analyzes related knowledge, and feeds it back to users through knowledge reasoning. Knowledge question answering systems can be divided into three types: Remark 1. Task-based: according to the user’s intention and information in the mobile phone, it will help the user to complete the operation. For example, with the current smart phone combined with voice recognition system, users can easily complete online booking through voice input, or combined with smart home to complete the control of water heater (Fig. 2). Remark 2. Question-based: It is mainly to retrieve and answer the natural language questions raised by users. For example: for the question who Yao Ming’s wife is, the system will automatically help users answer through the knowledge graph. Remark 3. Chat-based: Simple human-computer conversation can meet users’ emotional needs and shorten the distance between humans and computers. Baidu Knowledge Q & A platform has introduced the knowledge graph technology [9]. as well as Apple’s intelligent voice assistant Siri. Big Data Analysis. The knowledge graph enhances the association between data, making is convenient for users to more intuitively observe the relationship between the data.

Data Analysis Based on Knowledge Graph

379

There are some specific applications such as: implementing a recommendation system based on big data analysis of knowledge graphs [10]. When people shop online, the system uses big data analysis to recommend people suitable products and services.

3 Data Analysis Based on Knowledge Graph The data analysis through knowledge graph includes the following processes as shown in Fig. 1. First is the modeling of knowledge data, that is building a complete system of domain knowledge. Then the knowledge data representation should be done using OWL language to describe: Individual, Property, Functional Property, Inverse Functional Property, Transitive Property, Symmetric Property, Object Property, Datatype Property, Annotation Property and Class. Knowledge extraction and data fusion are then conducted to solve the heterogeneous problem. Generally, relational database is used for data storage. Afterwards, different kinds of data analysis approaches of knowledge graph are carried out as described in Sect. 3.F. Finally, the reasoning of knowledge data is performed to infer the hidden relationship.

Fig. 1. The process of data analysis based on Knowledge Graph

3.1 Knowledge Data Modeling The role of knowledge data modeling is to build a complete system of domain knowledge or open domain knowledge. Knowledge data modeling is an iterative update process, generally in two ways: one is a top-down construction method, that is, the pattern in which the knowledge graph is first defined, the pattern then is constructed from the top-level concepts [11]. and it will be gradually refined to form a hierarchical structure using structured data sources such as encyclopedia websites to define the ontology or data model for the knowledge base, and then add it. The other is a bottom-up method, which first summarizes the entities [12]. The method of summary is to extract useful entities from public information by means

380

K. Xie et al.

of certain methods, and then gradually classify upwards and build a hierarchy structure, abstracting new patterns, adding them to the knowledge base, and finally forming a top-level ontology model. 3.2 Knowledge Data Representation Knowledge data representation is an essential foundation for natural language understanding. The inconsistent way of knowledge representation and the lack of strict semantic theoretical models and formal semantic definitions have become the most crucial problems to be solved at present [13]. In the 1990s, R. Davis of MITAI Lab defined five major uses of knowledge representation: knowledge first needs to define the machine representation or reference of objective entities; knowledge data representation also is supposed to define the concepts and category systems used to describe objective things; it also needs to provide models and methods for machine reasoning; it is also a data structure for efficient computing; it must also be close to human cognition and be a machine that humans can understand [14]. 3.3 Knowledge Data Extraction Knowledge data extraction is a critical technique for constructing large-scale knowledge graphs. The key problem of knowledge data extraction is how to automatically extract information from heterogeneous data sources, extract knowledge and store it in the knowledge graph. There are three types of data sources for knowledge extraction: structured data (such as databases), semi-structured data (tables, lists, etc.) and unstructured data (that is, plain text data). As shown in Table 1, for different data source types, different type of extractions needs to be performed to address different problems [15]. Table 1. Different types of data extractions Extraction type

Description

Entity extraction

Entity extraction, also known as Named Entity Recognition (NER), refers to extracting named entities from text and dividing them according to predefined categories, such as person names, address names, and time

Relation extraction

The purpose of relation extraction is to end the problem of semantic connection between entities and to establish their relationship

Attribute extraction Attribute extraction is to solve the problem of entity attributes, so the events can be displayed through the entity attributes and the relationship between entities

3.4 Knowledge Data Fusion In real life, people define an ontology that meets their own characteristics according to their needs. In this way, a large number of ontologies will be generated in the same field.

Data Analysis Based on Knowledge Graph

381

These ontologies are often overlapping or related, but the used ontologies are different in language and model. So, here’s the heterogeneity. Then between different systems, if the same ontology is heterogeneous, they cannot interact normally. In addition, a large number of instances in the knowledge graph also have heterogeneity problems, such as different instances represent the same entity, and different entities represent the same instance. Therefore, the knowledge graph needs to solve the problem of ontology layer heterogeneity and instance layer heterogeneity. Knowledge fusion is the resolution by establishing the connection between heterogeneous ontology or heterogeneous instances, so that heterogeneous knowledge graphs can be connected to each other, thus achieving interoperability. 3.5 Knowledge Data Storage With the increasing data, the problem of knowledge storage has also become one of the key issues of the knowledge graph. Knowledge storage problems include knowledge preservation, knowledge query, knowledge reasoning, and knowledge analysis. As shown in Table 2, knowledge data storage databases can be divided into three categories: relational databases, RDF-oriented triples databases, and native graph databases, where the relational databases are the most commonly used ones [4]. 3.6 Data Analysis (Case Study Based on Risk Control Knowledge Graph) For risk control knowledge graph, it is particularly important to analyze the risk of fraud in the data. The main identification of anti-fraud is the person. First, we must construct a knowledge graph for the borrower and open up all its data sources. In this way, we can perform data analysis on the knowledge graph. From the perspective of algorithms, data analysis can be divided into two types: rule-based methods and probability-based methods [18]. Table 2. Different types of data extractions Relational database

RDF oriented triplet database

Native graph database

Specific introduction It refers to a database that uses a relational model to organize data. It stores data in rows and columns. It supports SPARQL query [16]

The RDF-oriented triples database is a database specially developed for storing large-scale RDF data and supports SPARQL queries [16]

Its storage efficiency is far superior to relational databases. The disadvantage is that when the graph data exceeds a certain scale, the system performance will decrease.

Common schemes

RDF4J, RDF-3X, gStore, Virtuoso, AllegroGraph

Neo4j [17]

Vertical partitioning, six-fold indexing, DB2RFD

382

K. Xie et al.

Rule-Based Methods Remark 1. Task-based: according to the user’s intention and information in the mobile phone, it will help the user to complete the operation. For example, with the current smart phone combined with voice recognition system, users can easily complete online booking through voice input, or combined with smart home to complete the control of water heater. Remark 2. Features based on rules: As shown in Fig. 3, we can analyze from the knowledge graph whether the applicant has direct or indirect contact with people on the blacklist. After these features are extracted, they are generally used as risk model inputs. Remark 3. Pattern-based judgment: As shown in Fig. 4, this method can effectively identify group fraud, the core of which is to further analyze the knowledge graph to find the information shared between entities which can basically be regarded as a group and be further analyzed. Rule-Based Methods. Probability-based methods: As shown in Fig. 5, the group mining algorithm mainly mines some groups from the knowledge graph data [19]. For those groups, we can intuitively understand that the density of internal relationships between groups is much greater than that of interrelationships between groups. The advantage of probability-based methods is that there is no need to manually define rules. For a complex network, defining rules is a very cumbersome thing. Knowledge Data Reasoning. The function of knowledge reasoning is to complete the knowledge graph. The basic method is to infer a new entity relationship and attribute relationship from the already established entity relationship network through computer deduction [20]. Knowledge reasoning is a significant part of the process of constructing a knowledge graph by helping complete knowledge, even attain new knowledge, and understand the current knowledge graph, and giving us more enlightment. Entity relationship reasoning. For example, Xiao Ming’s wife is Xiao Hong, and Xiao Hong’s son is Xiao Wang. Through knowledge reasoning, it can be inferred that Xiao Wang is Xiao Ming’s son. Attribute relationship reasoning. For example, Xiaoming’s birth date is in 1996, and it can be inferred that Xiaoming’s zodiac is rat.

Fig. 2. Inconsistent validation

Data Analysis Based on Knowledge Graph

383

Fig. 3. Extracting features

4 Open Problems and Challenges The concept of knowledge graph was proposed since 2012. Although Microsoft has shown the prospects of knowledge graph technology, it is not difficult to make a conclusion from the above analysis that in some aspects, we are still facing huge challenges in vertical domain. Knowledge extraction for the open domain is still in its infancy, the accuracy and recall of the algorithm are low, and the scalability is far from satisfaction. In terms of knowledge representation, the current way of knowledge storage is still in the form of triples, and the ability to express is limited faced with complex types of knowledge. In the field of knowledge application, the application scenarios of large-scale knowledge graphs are very limited. We can also see that applications such as intelligent search, knowledge answering, and social networking are still in their infancy, which requires to continuous learning and exploration. Knowledge graph-based data analysis also has many open problems and challenges in structured data. How to solve the problem of accurate and complete extraction of the required data from unstructured data is the starting and crucial step in data analysis.

Fig. 4. Pattern-based judgment

384

K. Xie et al.

Fig. 5. Group mining

5 Conclusion This paper elaborates the current research status and specific applications of data analysis based on knowledge graphs, summarizes the current process and core links of data processing for knowledge graph. Finally, the possible challenges are also presented that should be gradually solved in the future research. Moreover, this paper also hopes that the concluded insights can do a favor to those who are interested in the application of knowledge graphs.

References 1. Liu, E., Yang, L., Hong, D., Yao, L., Zhiguang, Q.: Overview of knowledge map construction technology. Comput. Res. Dev. 53(03), 582–600 (2016) 2. Zhu, L., et al.: Intelligent graph review system based on knowledge map. In: 2019 18th International Symposium on Distributed Computing and Applications for Business Engineering and Science (DCABES), Wuhan, China, pp. 100–103 (2019) 3. Guan, Q., Zhang, F., Zhang, E.: Application prospect of knowledge graph technology in knowledge management of oil and gas exploration and development. In: 2019 2nd International Conference on Artificial Intelligence and Big Data (ICAIBD), Chengdu, China, pp. 161–166 (2019) 4. He, L., Jiang, P.: Manufacturing knowledge graph: a connectivism to answer production problems query with knowledge reuse. IEEE Access 7, 101231–101244 (2019) 5. Wang, R., Wang, M., Liu, J., Yao, S., Zheng, Q.: Graph embedding based query construction over knowledge graphs. In: 2018 IEEE International Conference on Big Knowledge (ICBK), Singapore, pp. 1–8 (2018) 6. Fensel, A.: Keynote: building smart cities with knowledge graphs. In: 2019 International Conference on Computer, Control, Informatics and its Applications (IC3INA), Tangerang, Indonesia, p. 1 (2019) 7. Leijie, F., Yv, B., Zhenyuan, Z.: Constructing a vertical knowledge graph for non-traditional machining industry. In: 2018 IEEE 15th International Conference on Networking, Sensing and Control (ICNSC), Zhuhai, pp. 1–5 (2018) 8. Huang, H., Yu, J., Liao, X., Xi, Y.: A review of knowledge graph studies. Appl. Comput. Syst. 28(06), 1–12 (2019)

Data Analysis Based on Knowledge Graph

385

9. Zhao, S.: Baidu knowledge graph: the intellectual heart of artificial intelligence. In: China Computer Conference, Fuzhou (2017) 10. Luo, C., Liu, X., Zhang, K., Chang, Q.: A recommendation system for cloud services based on knowledge graph. In: 2018 IEEE 9th International Conference on Software Engineering and Service Science (ICSESS), Beijing, China, pp. 1–4 (2018) 11. Pan, L.H., et al.: Mapping of knowledge in the field of coal mining. Comput. Appl. Softw. 36(8), 47–54, 59 (2019). https://doi.org/10.3969/j.issn.1000-386x.2019.08.009 12. Liu, Q., et al.: Overview of knowledge map construction technology. Comput. Res. Dev. 53(3), 582–600 (2016). https://doi.org/10.7544/issn1000-1239.2016.20148228 13. Wen, L., Li, J., Liu, Z., et al.: Knowledge representation and ontology modeling based on concept level network. Chin. J. Inf. 32(4), 66–73 (2008). https://doi.org/10.3969/j.issn.10030077.2018.04.008 14. Davis, R., Shrobe, H., Szolovits, P.: What is a knowledge representation? AI Mag. 14(1), 17 (1993) 15. Luo, Z., Wang, H., Xie, R.: Extract domain terminologies for knowledge graph construction using domain feature vectors. In: 2017 IEEE 2nd International Conference on Big Data Analysis (ICBDA), Beijing, pp. 53–57 (2017) 16. Elfaki, A., Aljaedi, A., Duan Y.: Mapping ERD to knowledge graph. In: 2019 IEEE World Congress on Services (SERVICES), Milan, Italy, pp. 110–114 (2019) 17. Wei, J., Liu, R.: An approach of constructing knowledge graph of the hundred schools of thought in ancient China. In: 2019 ACM/IEEE Joint Conference on Digital Libraries (JCDL), Champaign, IL, USA, pp. 335–336 (2019) 18. Bellomarini, L., Fakhoury, D., Gottlob, G., Sallinger, E.: Knowledge graphs and enterprise ai: the promise of an enabling technology. In: 2019 IEEE 35th International Conference on Data Engineering (ICDE), Macao, Macao, pp. 26–37 (2019) 19. Ying. L., Wen-xiang, G.: A plan recognition algorithm based on the probabilistic goal graph. In: 2011 International Conference on Network Computing and Information Security, Guilin, pp. 359–362 (2011) 20. Li, Z., Jin, X., Guan, S., Wang, Y., Cheng, X.: Path reasoning over knowledge graph: a multiagent and reinforcement learning based method. In: 2018 IEEE International Conference on Data Mining Workshops (ICDMW), Singapore, Singapore, pp. 929–936 (2018)

Integration of Software-Defined Network and Fuzzy Logic Approaches for Admission Control in 5G Wireless Networks: A Fuzzy-Based Scheme for QoS Evaluation Phudit Ampririt1(B) , Seiji Ohara1 , Ermioni Qafzezi1 , Makoto Ikeda2 , Leonard Barolli2 , and Makoto Takizawa3 1

Graduate School of Engineering, Fukuoka Institute of Technology, 3-30-1 Wajiro-Higashi, Higashi-Ku, Fukuoka 811-0295, Japan [email protected], [email protected], [email protected] 2 Department of Information and Communication Engineering, Fukuoka Institute of Technology, 3-30-1 Wajiro-Higashi, Higashi-Ku, Fukuoka 811-0295, Japan [email protected], [email protected] 3 Department of Advanced Sciences, Faculty of Science and Engineering, Hosei University, Kajino-Machi, Koganei-Shi, Tokyo 184-8584, Japan [email protected]

Abstract. The Fifth Generation (5G) network is expected to be flexible to satisfy user requirements and the Software-Defined Network (SDN) with Network Slicing will be the constructive paradigm for admission control. The Quality of Service (QoS) is influential point in the 5G wireless networks. In this paper, we propose a Fuzzy-based scheme to evaluate the QoS considering three parameters: Slice Throughput, Slice Delay and Slice Loss. We carried out the simulation for evaluating the performance of our proposed scheme. From simulation results, we conclude that the considered parameters have different effect on the QoS performance. When Slice Throughput (ST) is increasing, the QoS parameter is increased. But, when Slice Delay (SD) and Slice Loss (SL) are increasing, the QoS is decreased.

1

Introduction

Recently, the growth of wireless technologies and user’s demand of services are increasing rapidly. Especially in 5G networks, there will be billions of new devices with unpredictable traffic pattern which provide high data rates. With the appearance of Internet of Things (IoT), these devices will generate Big Data to the Internet, which will cause to congest and deteriorate the QoS [1]. The 5G network will be expected to be better than 4G. The 5G network will provide users with new experiences such as Ultra High Definition Television c The Editor(s) (if applicable) and The Author(s), under exclusive license  to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 386–396, 2021. https://doi.org/10.1007/978-3-030-61108-8_38

Integration of Software-Defined Network and Fuzzy Logic Approaches

387

(UHDT) on Internet and support a lot of IoT devices with long battery life and high data rate on hotspot areas with high user density. In the 5G technology, the routing and switching technologies aren’t important anymore or coverage area is shorter than 4G because it uses high frequency for facing higher device’s volume for high user density [2–4]. Therefore, there are many research work that try to build systems which are suitable to 5G era. The SDN is one of them [5]. For example, the mobile handover mechanism with SDN is used for reducing the delay in handover processing and improve QoS. Also, by using SDN the QoS can be improved by applying Fuzzy Logic (FL) on SDN controller [6–8]. In our previous work [9], we presented a Fuzzy-based system for admission control in 5G Wireless Networks considering three parameters: Grade of Service (GS), User Request Delay Time (URDT), Network Slice Size (NSS). In this paper, we propose a fuzzy-based scheme for evaluation of QoS in 5G wireless networks. The rest of the paper is organized as follows. In Sect. 2 is presented an overview of SDN. In Sect. 3, we present application of Fuzzy Logic for admission control. In Sect. 4, we describe the proposed fuzzy-based system and its implementation. In Sect. 5, we explain the simulation results. Finally, conclusions and future work are presented in Sect. 6.

2

Software-Defined Networks (SDNs)

The SDN is a new networking paradigm that decouples the data plane from control plane in the network. In traditional networks, the whole network is controlled by each network device. However, the traditional networks are hard to manage and control since they rely on physical infrastructure. Network devices must stay connected all the time when user wants to connect other networks. Those processes must be based on the setting of each device, making controlling the operation of the network difficult. Therefore, they have to be set up one by one. In contrast, the SDN is easy to manage and provide network software based services from a centralised control plane. The SDN control plane is managed by SDN controller or cooperating group of SDN controllers. The SDN structure is shown in Fig. 1 [10,11]. • Application Layer builds an abstracted view of the network by collecting information from the controller for decision-making purposes. The types of applications are related to: network configuration and management, network monitoring, network troubleshooting, network policies and security. • Control Layer receives instructions or requirements from the Application Layer and control the Infrastructure Layer by using intelligent logic. • Infrastructure Layer receives orders from SDN controller and sends data among them.

388

P. Ampririt et al.

The SDN can manage network systems while enabling new services. In congestion traffic situation, management system can be flexible, allowing users to easily control and adapt resources appropriately throughout the control plane. Mobility management is easier and quicker in forwarding across different wireless technologies (e.g. 5G, 4G, Wifi and Wimax). Also, the handover procedure is simple and the delay can be decreased.

Fig. 1. Structure of SDN.

3

Outline of Fuzzy Logic

A Fuzzy Logic (FL) system is a nonlinear mapping of an input data vector into a scalar output, which is able to simultaneously handle numerical data and linguistic knowledge. The FL can deal with statements which may be true, false or intermediate truth-value. These statements are impossible to quantify using traditional mathematics. The FL system is used in many controlling applications such as aircraft control (Rockwell Corp.), Sendai subway operation (Hitachi), and TV picture adjustment (Sony) [12–14]. In Fig. 2 is shown Fuzzy Logic Controller (FLC) structure, which contains four components: fuzzifier, inference engine, fuzzy rule base and defuzzifier. • Fuzzifier is needed for combining the crisp values with rules which are linguistic variables and have fuzzy sets associated with them. • The Rules may be provided by expert or can be extracted from numerical data. In engineering case, the rules are expressed as a collection of IF-THEN statements.

Integration of Software-Defined Network and Fuzzy Logic Approaches

389

Fig. 2. FLC structure.

• The Inference Engine infers fuzzy output by considering fuzzified input values and fuzzy rules. • The Defuzzifier maps output set into crisp numbers. 3.1

Linguistic Variables

A concept that plays a central role in the application of FL is that of a linguistic variable. The linguistic variables may be viewed as a form of data compression. One linguistic variable may represent many numerical variables. It is suggestive to refer to this form of data compression as granulation. The same effect can be achieved by conventional quantization, but in the case of quantization, the values are intervals, whereas in the case of granulation the values are overlapping fuzzy sets. The advantages of granulation over quantization are as follows: • it is more general; • it mimics the way in which humans interpret linguistic values; • the transition from one linguistic value to a contiguous linguistic value is gradual rather than abrupt, resulting in continuity and robustness. For example, let Temperature (T) be interpreted as a linguistic variable. It can be decomposed into a set of Terms: T (Temperature) = {Freezing, Cold, Warm, Hot, Blazing}. Each term is characterised by fuzzy sets which can be interpreted, for instance, “Freezing” as a temparature below 0 ◦ C, “Cold” as a temparature close to 10 ◦ C. 3.2

Fuzzy Control Rules

Rules are usually written in the form “IF x is S THEN y is T” where x and y are linguistic variables that are expressed by S and T, which are fuzzy sets. The x is a control (input) variable and y is the solution (output) variable. This

390

P. Ampririt et al.

rule is called Fuzzy control rule. The form “IF ... THEN” is called a conditional sentence. It consists of “IF” which is called the antecedent and “THEN” is called the consequent. 3.3

Defuzzificaion Method

There are many defuzzification methods, which are showing in following: • • • • •

4

The Centroid Method; Tsukamoto’s Defuzzification Method; The Center of Are (COA) Method; The Mean of Maximum (MOM) Method; Defuzzification when Output of Rules are Function of Their Inputs.

Proposed Fuzzy-Based System

In this work, we use FL to implement the proposed system. In Fig. 3, we show the overview of our proposed system. Each evolve Base Station (eBS) will receive controlling order from SDN controller and they can communicate and send data with User Equipment (UE). On the other hand, the SDN controller will collect all the data about network traffic status and controlling eBS by using the proposed fuzzy-based system. The SDN controller will be a communicating bridge between eBS and 5G core network. The proposed system is called Fuzzy-based System for Admission Control (FBSAC) in 5G wireless networks. The structure of FBSAC is shown in Fig. 4. For the implementation of our system, we consider four input parameters: Quality of service (QoS), User Request Delay Time (URDT), Slice Priority (SP), Slice Overloading Cost (SOC) and the output parameter is Admission Decision (AD). In this paper, we apply FL to evaluate the QoS for FBSAC system. For QoS evaluation, we consider three parameters: Slice Throughput (ST), Slice Delay (SD), Slice Loss (SL) and the output parameter is Quality of Service (QoS). The membership functions are shown in Fig. 5. We use triangular and trapezoidal membership functions because they are more suitable for real-time operations [15–18]. We show parameters and their term sets in Table 1. The Fuzzy Rule Base (FRB) is shown in Table 2 and has 27 rules. The control rules have the form: IF “condition” THEN “control action”. For example, for Rule 1:“IF ST is Lo, SD is Sh, SL is Lw, THEN QoS is QoS5”.

Integration of Software-Defined Network and Fuzzy Logic Approaches

391

Fig. 3. Proposed system overview.

Fig. 4. Proposed system structure. Table 1. Parameter and their term sets. Parameters

Term set

Slice Throughput (ST)

Low (Lo), Moderate (Mo), Fast (Ft)

Slice Delay (SD)

Short (Sh), Intermediate(In), Long (Ln)

Slice Loss (SL)

Low (Lw), Medium (Md), High (Hg)

Quality of Service (QoS) QoS1, QoS2, QoS3, QoS4, QoS5, QoS6, QoS7

5

Simulation Results

In this section, we present the simulation result of our proposed scheme. The simulation results are shown in Fig. 6, Fig. 7 and Fig. 8. They show the relation of QoS with ST, SD and SL. We consider the ST as constant parameter. We

392

P. Ampririt et al.

Fig. 5. Membership functions.

Integration of Software-Defined Network and Fuzzy Logic Approaches

393

Table 2. Fuzzy rule base Rule ST SD SL

QoS

1

Lo

Sh Lw QoS5

2

Lo

Sh Md QoS4

3

Lo

Sh Hg QoS3

4

Lo

In

Lw QoS4

5

Lo

In

Md QoS3

6

Lo

In

Hg QoS2

7

Lo

Ln Lw QoS3

8

Lo

Ln Md QoS2

9

Lo

Ln Hg QoS1

10

Mo Sh Lw QoS6

11

Mo Sh Md QoS5

12

Mo Sh Hg QoS4

13

Mo In

Lw QoS5

14

Mo In

Md QoS4

15

Mo In

Hg QoS3

16

Mo Ln Lw QoS4

17

Mo Ln Md QoS3

18

Mo Ln Hg QoS2

19

Ft

Sh Lw QoS7

20

Ft

Sh Md QoS6

21

Ft

Sh Hg QoS5

22

Ft

In

Lw QoS6

23

Ft

In

Md QoS5

24

Ft

In

Hg QoS4

25

Ft

Ln Lw QoS5

26

Ft

Ln Md QoS4

27

Ft

Ln Hg QoS3

change the SD value from 0.1 to 0.9 and the SL from 0 to 1. In Fig. 6, we consider the ST value as 0.1. When SL increased form 0.1 to 0.9, we see that QoS is decreasing. When SL is 0.6, the QoS is increased by 12.2% and 15% when SD is decreased form 0.9 to 0.5 and form 0.5 to 0.1, respectively. We compare Fig. 6 with Fig. 7 to see how ST has affected QoS. When SL is 0.2 and SD is 0.1, QoS is increased 15% by increasing ST from 0.1 to 0.5. This is because a higher ST value means the slice has more throughput and can provide better QoS. In Fig. 7, when SD is 0.1, all QoS values are higher than 0.5. This means that the system will have acceptable QoS performance.

P. Ampririt et al. ST=0.1 1

SD=0.1 SD=0.5 SD=0.9

0.9 0.8

QoS [unit]

0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0

0.1

0.2

0.3

0.4

0.5 0.6 SL [unit]

0.7

0.8

0.9

1

0.9

1

0.9

1

Fig. 6. Simulation results for ST = 0.1. ST=0.5 1

SD=0.1 SD=0.5 SD=0.9

0.9 0.8

QoS [unit]

0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0

0.1

0.2

0.3

0.4

0.5 0.6 SL [unit]

0.7

0.8

Fig. 7. Simulation results for ST = 0.5. ST=0.9 1 0.9 0.8 0.7 QoS [unit]

394

0.6 0.5 0.4 0.3 0.2 SD=0.1 SD=0.5 SD=0.9

0.1 0 0

0.1

0.2

0.3

0.4

0.5 0.6 SL [unit]

0.7

0.8

Fig. 8. Simulation results for ST = 0.9.

Integration of Software-Defined Network and Fuzzy Logic Approaches

395

In Fig. 8, we increase the value of ST to 0.9. We see that the QoS value is increased much more compared with the results of Fig. 6 and Fig. 7.

6

Conclusions and Future Work

In this paper, we proposed and implemented a Fuzzy-based scheme for evaluation of QoS. The evaluated QoS parameter will be used as input parameter for Admission Control in 5G Wireless Networks. We evaluated the proposed scheme by simulation. From simulation results, we conclude as follows. • When ST parameter is increased, the QoS parameter is increased. This means that QoS performance will be high. • When SD, SL parameters are increased, the QoS parameter is decreased. Thus, the QoS performance will be low. In the future, we would like to evaluate the Admission Control system by considering QoS parameter and other parameters.

References 1. Navarro-Ortiz, J., Romero-Diaz, P., Sendra, S., Ameigeiras, P., Ramos-Munoz, J.J., Lopez-Soler, J.M.: A survey on 5G usage scenarios and traffic models. IEEE Commun. Surv. Tutor. 22(2), 905–929 (2020) 2. Hossain, S.: 5G wireless communication systems. Am. J. Eng. Res. (AJER) 2(10), 344–353 (2013) 3. Giordani, M., Mezzavilla, M., Zorzi, M.: Initial access in 5G mmwave cellular networks. IEEE Commun. Mag. 54(11), 40–47 (2016) 4. Kamil, I.A., Ogundoyin, S.O.: Lightweight privacy-preserving power injection and communication over vehicular networks and 5G smart grid slice with provable security. Internet Things 8, 100–116 (2019) 5. Hossain, E., Hasan, M.: 5G cellular: key enabling technologies and research challenges. IEEE Instrum. Meas. Mag. 18(3), 11–21 (2015) 6. Yao, D., Su, X., Liu, B., Zeng, J.: A mobile handover mechanism based on fuzzy logic and MPTCP protocol under SDN architecture∗. In: 18th International Symposium on Communications and Information Technologies (ISCIT-2018) 7. Lee, J., Yoo, Y.: Handover cell selection using user mobility information in a 5G SDN-based network. In: 2017 Ninth International Conference on Ubiquitous and Future Networks (ICUFN-2017), pp. 697–702, July 2017 8. Moravejosharieh, A., Ahmadi, K., Ahmad, S.: A fuzzy logic approach to increase quality of service in software defined networking. In: 2018 International Conference on Advances in Computing, Communication Control and Networking (ICACCCN2018), pp. 68–73, October 2018 9. Ampririt, P., Ohara, S., Liu, Y., Ikeda, M., Maeda, H., Barolli, L.: A fuzzy-based system for admission control in 5G wireless networks considering software-defined network approach. In: International Conference on Emerging Internetworking, Data & Web Technologies, pp. 73–81. Springer (2020) 10. Li, L.E., Mao, Z.M., Rexford, J.: Toward software-defined cellular networks. In: 2012 European Workshop on Software Defined Networking, pp. 7–12, October 2012

396

P. Ampririt et al.

11. Mousa, M., Bahaa-Eldin, A.M., Sobh, M.: Software defined networking concepts and challenges. In: 2016 11th International Conference on Computer Engineering & Systems (ICCES), pp. 79–90. IEEE (2016) 12. Jantzen, J.: “Tutorial on fuzzy logic,” Technical University of Denmark. Dept. of Automation, Technical Report (1998) 13. Mendel, J.M.: Fuzzy logic systems for engineering: a tutorial. Proc. IEEE 83(3), 345–377 (1995) 14. Zadeh, L.A.: Fuzzy logic. Computer 21, 83–93 (1988) 15. Norp, T.: 5G requirements and key performance indicators. J. ICT Stand. 6(1), 15–30 (2018) 16. Parvez, I., Rahmati, A., Guvenc, I., Sarwat, A.I., Dai, H.: A survey on low latency towards 5G: RAN, core network and caching solutions. IEEE Commun. Surv. Tutor. 20(4), 3098–3130 (2018) 17. Kim, Y., Park, J., Kwon, D.H., Lim, H.: Buffer management of virtualized network slices for quality-of-service satisfaction. In: 2018 IEEE Conference on Network Function Virtualization and Software Defined Networks (NFV-SDN), pp. 1–4 (2018) 18. Barolli, L., Koyama, A., Yamada, T., Yokoyama, S.: An integrated CAC and routing strategy for high-speed large-scale networks using cooperative agents. IPSJ J. 42(2), 222–233 (2001)

ICS Testbed Implementation Considering Dataset Collection Environment Eunseon Jeong(B) , Junyoung Park, Minseong Kim, Chanmin Kim, Soyoung Jung, and Kangbin Yim Department of Information Security Engineering, Soonchunhyang University, Asan, South Korea {gomdoli10,apple,ozrneam,tten15,jso0330,yim}@sch.ac.kr

Abstract. As there have been a growing number of introductions to integrated management service on the ICS (industrial control systems) and advancements of smart factories which link to various forms of network, cyber attacks targeting the ICS have been on the rise. To defend against them, a full set of data is needed when attacks take place in the actual ICS. Collecting data from attacks is a difficult task to perform in the actual environment due to the characteristic of the ICS that requires high-level security. To resolve this issue, some studies are in progress as to building ICS testbeds and collecting attack data, but they face a problem of not reflecting requirements for the collection of attack data set. This paper describes considerations to take when building testbeds for security research on the ICS and suggests subsequent implementations of the testbeds for collecting a set of data on the ICS.

1 Introduction As new cyberattacks targeting Industrial Control Systems (ICS) continue to take place, subsequent security requirements are also on the rise [1–7]. In particular, the ICS, which had used in closed networks separated from public networks, have recently been linked to multiple networks due to the development of Smart Factory and the operation of efficient integrated management system. On the other hand, this allowed attackers various invasion paths and facilitated the spread of malware by far. It also resulted in the growing cyberthreats to the ICS. A reduction in the cyber security threats to the industrial control systems can be achieved by improving systems against vulnerabilities beforehand through security tests such as a vulnerability analysis. However, it is very difficult to run a security test in actual ICS environments where high stability and availability are in need. Hence, this raises a need for ICS testbeds capable of validating attacks on industry control systems (ICS) and collecting the attack dataset. Led by affiliated research centers such as ORNL (Oak Ridge National Laboratory), iTrust and ETRI, robust studies are underway as to building ICS testbeds and collecting datasets for the purpose of security tests, providing constructed testbeds from validated attack datasets. However, previously conducted studies face problems, such as lack of full range datasets, no representation of the complexity of ICS and taking no consideration © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 397–406, 2021. https://doi.org/10.1007/978-3-030-61108-8_39

398

E. Jeong et al.

on environment for dataset collection. Hence, this paper aims to propose considerations to be taken in building testbeds for dataset collection from the ICS as well as how to build up focused ICS testbeds.

2 Related Work ICS testbed configuration for security testing and research on dataset collection are led by research institutes, such affiliated research centers as iTrust, Oak Ridge National Laboratory and ETRI. What they provide includes ICS datasets designed to detect known cyberattacks targeting the ICS. 2.1 ORNL (Oak Ridge National Laboratory) [8] To serve the purpose of tightening ICS security, Tommy Morris at ORNL (Oak Ridge National Laboratory) had constructed a variety of ICS testbeds, such as power systems, gas pipelines, water storage tank systems and energy management systems, from 2011 to 2015. So, validated attack datasets were provided from constructed testbeds. But, lack of information about constructed testbeds and faults found in part of provided datasets kept them from being applied actively to ICS security researches (Figs. 1 and 2).

Fig. 1. Power system framework

2.2 iTrust [9] iTrust retains three types of testbeds: Secure Water Treatment (SWaT), Water Distribution (WADI) and Electric Power Intelligent Control (EPIC). The world-class testbed host provides normal datasets and attack datasets for each testbed. Swat, the most well-known out of the three, is the water treatment system testbed built on requirements specified

ICS Testbed Implementation Considering Dataset Collection Environment

399

Fig. 2. SWaT framework

in ISA/IEC 62443 Industrial Network and System Security Standards. Various research activities are underway with respect to detecting attacks against the ICS through the verified attack dataset in the corresponding testbed by building environments similar to actual industrial water treatment system. 2.3 HAI (HIL Based Augmented ICS Testbed) [10] An ETRI-affiliated research institute provides HAI 1.0 (HIL based Augmented ICS testbed), ICS security dataset. The dataset, which is designed to develop ICS security technology with artificial intelligence (AI), builds testbeds industrial control systems, sensors and actuators from GE, Emerson and Siemens and provides attack datasets collected by replaying potential attacks on ICS (Fig. 3).

Fig. 3. HAI framework

3 Related Work Aimed to explore and defend types of attacks taking place in industry control system (ICS), a diversity of research activities is ongoing to configure testbeds similar to actual ICS environment and collect normal/attack datasets. However, datasets made publicized to date are restricted to ICS equipment and network packet data. They also fail to indicate a clear point of attack occurrence in some cases. Therefore, considering the

400

E. Jeong et al.

characteristics of malware attacks targeting ICS as described above, this paper defined four considerations to build testbeds capable of practical dataset collection as listed below. 3.1 ISA 95/99 Based ICS Framework Configuring the ICS in the actual industrial environment is advised to comply with International Society of Automation (ISA) standards, being represented by ISA95 and ISA99 standards [11, 12]. Serving as manufacturing and operating management system standards, ISA95 standards define each level of ICS, whereas ISA99 provide definitions of security and network environments requested in the ICS (Fig. 4 and Table 1).

Fig. 4. ICS architecture Table 1. ICS level. Level Description 0

Field network (sensor, robot, actuator)

1

Control network (PLC)

2

Supervisory/process network (workstation, HMI, OPC server)

3

Operations and control ICT/DMZ (historian, firewall)

4

Enterprise/IT network (web server, printer, remote access)

Actual industrial environments should be considered and configured to collect normal and attack datasets generated in the ICS, while certainly referring to ISA 95/99 standards as a base model. 3.2 Dataset Collection Range Normal and attack datasets in the ICS framework should demand a data collection from all sections (level 0–4) given the characteristics of ICS malware.

ICS Testbed Implementation Considering Dataset Collection Environment

401

Collect-to-be components should include: ➀ Sensor Data, ➁ Control System (PLC) Operation Log ➂ HMI Operation Log, ➃ Firewall Log ➄ O/S (Windows) Log ➅ Network Packet. 3.3 Ease of Collecting and Restoring Attack Dataset Datasets designed to detect attacks to the ICS can be applied to laying the research groundwork for Big Data such as Machine Learning and Deep Learning. It is necessary to collect datasets by executing various types of malware and performing simulated repeating attacks. In case attack datasets are collected repeatedly, previously performed attack datasets should not influence datasets to be collected. But, following the malware being attacked, the system has already been compromised, which is incapable of guaranteeing no system integrity, accordingly. Now that it is time-consuming to bring back to normal all the systems configured in testbeds such as PLC and workstations, testbeds should be designed in consideration of environments where systems can return back to normal. 3.4 Inter-dataset Time Synchronization A reference point where an attack has initiated should be present when analyzing types of attack in the collected attack dataset. As different types of attack data are simultaneously collected from each segment of the IC testbed, time for each dataset should be synchronized to specify a point of attack.

4 ICS Testbed Configuration Method This chapter proposes ICS testbeds configured on the basis of considerations, which were provided in Sect. 3, for collecting datasets from the ICS testbeds for the purpose of security testing. Figure 5 demonstrates the network architecture of the ICS testbed built to collect attack datasets and marked logs collected from each point. By applying the ISA 95-based network topology model in order to gather system log likely to occur in the ICS, the ICS testbed was designed to collect data from network environments similar to operational environments and being practiced in actual industrial sectors. In particular, Level3 testbed had virtual operation environment built using VMware ESXi so that a variety of PC nodes could be configured, and repeat attacks enabled in collecting attack datasets. To return back to normal before collecting attack data from the testbed built, it saves the virtual machines in normal condition using a snapshot feature in the virtual environment. Then, it performs attacks and restores again with a snapshot. Through this way, time and cost savings can be achieved (Fig. 6).

402

E. Jeong et al.

Fig. 5. Network architecture of testbed

Fig. 6. Hydro power generation and water purification system

ICS Testbed Implementation Considering Dataset Collection Environment

403

In order to collect data from actuators and sensors based on the architecture described above, WAGO and SIEMENS PLCs were used to build hydroelectric power systems and water purification systems (Fig. 7).

Fig. 7. Sensor and actuator connection diagram of WAGO PLC

WAGO PLC uses a motor pump for supplying water from the main water tank to water tank 1. To keep water tank 1 at reasonable water levels, the controller regulates the pump speed responsible for the influx of water depending on water levels measured by a 3-step water level sensor and determines whether to inject water to the power generation system. Water injected by the pump shifts to the power generation system and water used for power generation is stored in water tank 2 (Fig. 8).

404

E. Jeong et al.

Fig. 8. Sensor and actuator connection diagram of SIEMENS PLC

Fig. 9. ICS testbed

SIEMENS PLC is a controller used to discharge water into the water purification system through a 2-step water level sensor in water tank 2 and water purified through the water purification system is delivered to the main tank. The water delivered to the main tank is reused for power generation use. Figure 9 above shows the actually configured testbed for dataset collection.

ICS Testbed Implementation Considering Dataset Collection Environment

405

5 Conclusions Unlike the closed system in the past, ICS environments interconnect to various types of networks for the sake of work efficiency, which is bringing many different types of ongoing attacks against the ICS. The ICS uses controllers and industrial control protocols which are not used in general cyber environments. Taking advantage of such characteristics, ICS attackers set their targets and produce malware fitting to their ICS environments. For this reason, there is a growing need to collect and analyze types of attacks taking place in the actual operation environment unlike the existing cyber attacks. However, the ICS should continue to operate without stopping due to its unique characteristics, and it requires a high level of protection which is along with many problems, such as making it difficult to collect data and analyze attacks in actual operation environments. Hence, ICS testbeds are therefore needed to collect and analyze attack datasets created to defend attacks beforehand. This paper investigated existing ICS testbeds collecting attack datasets likely to occur in ICS. As well, it also provided a definition of considerations in building ICS testbeds and proposed measures to build testbeds based on the considerations. This paper expects to draw unfound vulnerabilities through proposed considerations, develop configuration methods for the ICS testbeds and collect practical datasets related to the vulnerabilities, ultimately applying it to IDS (Intrusion Detection System) development for attack detection. Acknowledgments. This research was supported by the MSIT (Ministry of Science and ICT), Korea, under the ITRC (Information Technology Research Center) support program (IITP-20202015-0-00403) supervised by the IITP (Institute for Information & communications Technology Planning & Evaluation) and the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (NRF-2018R1A4A1025632).

References 1. Spenneberg, R., Brüggemann, M., Schwartke, H.: PLC-BLASTER: A Worm Living Solely in the PLC. Black Hat Asia (2016) 2. ICS-CERT: Abb ac500 plc webserver codesys vulnerability (2013). https://ics-cert.us-cert. gov/advisories/ICSA-12-320-01 3. ICS-CERT: Password Transmission Vulnerability (2019). https://www.us-cert.gov/ics/adviso ries/icsa-19-213-04 4. ICS-CERT: Schneider electric modicon m340 buffer overflow vulnerability. https://ics-cert. us-cert.gov/advisories/ICSA-15-351-01 5. ICS-CERT: Rockwell automation micrologix 1100 plc overflow vulnerability (2016). https:// ics-cert.us-cert.gov/advisories/ICSA-16-026-02 6. DigitalBond: 3S CoDeSys, Project Basecamp (2012). http://www.digitalbond.com/tools/bas ecamp/3s-codesys/ 7. DigitalBond: WAGO IPC 758/870, Project Basecamp (2015). http://www.digitalbond.com/ tools/basecamp/wago-ipc-758870/ 8. Morris, T.H.: Industrial control system (ICS) cyber attack datasets. https://sites.google.com/ a/uah.edu/tommy-morris-uah/ics-data-sets. Accessed 30 Apr 2018

406

E. Jeong et al.

9. iTrust: Swat datasets. https://itrust.sutd.edu.sg/dataset/. Accessed 30 Apr 2018 10. Shin, H.-K., et al.: Implementation of programmable {CPS} testbed for anomaly detection. In: 12th {USENIX} Workshop on Cyber Security Experimentation and Test ({CSET} 2019) (2019) 11. American National Standard ANSI/ISA-95.00.01-2010 Enterprise control system integration—Part 1: models and terminology (2010) 12. American National Standard ANSI/ISA–99.00.01–2007 Security for Industrial Automation and Control Systems. Part 1: Terminology, Conce (2007)

A Study on Reducing Interest Misleading by Publisher Migration on Mobile Networks Taichi Iwamoto(B) and Tetsuya Shigeyasu Department of Management and Information System, Prefectural University of Hiroshima, Hiroshima, Japan [email protected], [email protected]

Abstract. Recently, NDN (Named Data Networking) [1] which is a new content communication architecture, attracts a lot of attentions from the network research community. NDN routes the Interest/Data according to both of its content name and route information on each relay router. Then, NDN routing is categorized as hop-by-hop routing. Incidentally, due to the rapid development of recent ICT technology, current mobile networks allow a mobile terminal to publish/delivery its content even if it is in migration. NDN, however, could not operate correctly because of old information of publisher’s location information on relay router when we implement original NDN on such mobile networks. The literature [2] has proposed the method for solving the problem, but it requires publisher’s destination information. Hence, in this paper, we propose new method solving the Interest misleading problem include by mobile publisher. In order to clarify the advances of our proposal, this paper reports the results of performance evaluation in terms of cache hit ratio and content acquisition ratio.

1

Introduction

Recent Internet traffic is growing continuously. The most part of then is consisted of sharable contents: movies, music and other media. So, it can be said that the Internet is utilized for content sharing. For the purpose of content sharing, most important thing is “how to obtain” (content centric manner) than “where it is obtained from” (location centric manner). The traditional Internet, however, only provides the location centric manner on the basis of IP (Internet Protocol) communications. Hence, development of a new network architecture based on content centric manner is strongly desired by network researchers in these days. The NDN (Named Data Networking), one of the content centric network architecture, has been proposed. The NDN utilizes relay routers, named, CR (Content Router) as the network buffer for caching contents. In NDN, forwarding of content acquisition requests (Interest) and returning contents (Data) are based on content name. CRs receiving Interest/Data forward those packets according c The Editor(s) (if applicable) and The Author(s), under exclusive license  to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 407–415, 2021. https://doi.org/10.1007/978-3-030-61108-8_40

408

T. Iwamoto and T. Shigeyasu

to information of its local routing entry. This characteristic realizes autonomous forwarding control on NDN and it makes NDN high scalable network architecture. Incidentally, current development of ICT enables high computing performance even if the mobile terminals. These mobile terminals can publish/deliver the content under mobile environment (mobile publisher). Meanwhile, it is assumed that NDN can not deliver requested content correctly when the network implements conventional NDN over mobile network because of old local routing entry induced by mobile publishers. This paper called the problem as “mobile publisher problem”. The literature [2] has proposed a method PMC (Publisher Mobility support protocol in CCN) to solve the mobile publisher problem. But the method proposed in the literature requires the destination information of mobile publisher for its correct operation. This control lacks of autonomy of NDN. Hence, in this paper, we propose a new method operating autonomously without destination information of mobile publisher. In our proposal, mobile publisher transfers its contents to neighboring fixed node in advance its network migration. This paper also clarifies the advantages of our proposal than the conventional NDN in terms of content acquisition ratio and cache hit ratio on the basis of computer evaluations. The rest of this paper is organized as follows: Sect. 2 describes the related works, and Sect. 3 describes the problem of content delivery on NDN due to migration of mobile publisher. Section 4 proposes our method solving mobile publisher problem without any information of location of publisher. Section 5 describes the evaluation condition, and Sect. 6 reports results of performance evaluation conducted by computer simulation. Section 6 concludes the paper and mentions our future tasks.

2

Related Works

For the solution to the publisher migration problem on NDN, PMC (Publisher Mobility support protocol in CCN) has been proposed in [2]. In the PMC, in order to cope with the publisher migration, a publisher selects a HomeNode. The publisher registers its new location to the HomeNode when it migrates to the other place. By the HomeNode, Interests arrived at the old publisher’s location due to the old/incorrect FIB entry, can be correctly forwarded to the new publisher’s location. Figure 1 shows the procedure of Interest forwarding on PMC. As the figure shows, in advance a publisher migration, the publisher reports its destination information to the HomeNode as MR (Mobility Report) request. The HomeNode returns a MR response to the mobile publisher. At this time, nodes receiving MR response update its FIB entry according to the MR.

A Study on Reducing Interest Misleading by Publisher Migration

409

Fig. 1. PMC

3

Publisher Migration Problem

This chapter describes the publisher migration problem which generates undeliverable message due to old and incorrect forwarding information on each relay router. Figure 2 show the problem. In this figure, node 0 is a mobile publisher, and it migrates from the side of node 1 to node 4. The figure supposed the situation that the content request/delivery for “/host0/1” took place among user (node 3) and the mobile publisher (node 0). By the packets exchange, FIB entry on routers belonging to the forwarding path, are registered as shown in Fig. 2. After the registration, mobile publisher can not receive Interests from node 3 after user its migration to the side of node 4 even if user sends Interests repeatedly because node 2 forwards all Interests to 1 according to the old/incorrect FIB entry.

4

Proposal

This paper proposes a method for solving undeliverable messages induced by publisher migration problem. The proposal consists of two parts: 1) contents transfer in advance publisher migration, and 2) elimination of old and incorrect FIB entries on relay routers. Implementation of 1) reduces the mishits of requested contents due to unable to connect the publisher during the migration, and 2) reduces undeliverable messages due to mis routing by old/incorrect FIB entry. 1) is called advance migration.

410

T. Iwamoto and T. Shigeyasu

Fig. 2. Publisher migration problem

When the mobile publisher migrates its location to new location, it transfers its original contents to neighboring relay node. The transferred contents is marked as it advanced migrated contents by set M (Migrated) flag in it header. The neighbor node received advanced migrated contents, and it stores them into its buffer and deliver them as same as usual for content requests. The relay nodes receiving the contents with “M ” flag, however, delete corresponding FIB entry to avoid mis routing for future request. The reason for that is, advance migration is supposed to use for tentative content deliveries. On the other hand, if relay routers received content without M flag, it set FIB entry according to the receiving face number when the corresponding FIB entry is not currently registered. At the relay nodes receive the Interest from its downstream node, it checks its buffer and returns cached content if it exists. Otherwise, it forwards the Interest according to its FIB entry. At this time, if the corresponding FIB entry does not exist, the relay router forwards toward its upstream relay routers in broadcast fashion (Fig. 3).

5

Performance Evaluation

This chapter describe the simulation environment for performance evaluations of proposal.

A Study on Reducing Interest Misleading by Publisher Migration

411

Fig. 3. Proposal

5.1

Simulation Environment

Figure 4 shows the network topology used for the performance evaluation. As the figure shows, node 0 is a mobile publisher, and it migrates from the side of node 1 to the side of node 5 during evaluation. Node 3 and 4 are the user groups consist from multiple users each other. Each group generates 100 random Interests during every 1 s, for generated contents at mobile publisher. On the other hand, mobile publisher generates 100 original contents during every 1 s. For more realistic evaluation, each user group generates random Interest with biased probabilities in our evaluations. For making the biased random probabilities, we divide generated contents into 10 classes, and it identified 1, 2, . . ., 10 in the ascending order of elapsed time. Under the above assumption, probability of selection of class x content is as Eq. (1). P (x) =

10  1 10k

(1)

k=x

In the following section, we report the results of performance evaluations in terms of cache hit ratio and ratio of successful content acquisition. For the evaluations, we assumed that mobile publisher selects advance migration contents based on elapsed time since its generation. It means that more younger contents likely to be selected (Table 1).

412

T. Iwamoto and T. Shigeyasu

Fig. 4. Network topology Table 1. Simulation parameters Parameter

Value

# of nodes

6

Interest generation rate 100 [pkt/sec] Cache capacity

5.2

Infinity

Evaluation Result

Figure 5 and Fig. 6 show the cache hit ratio and content acquisition ratio under varying amount of advanced migrated contents. For the results of contents acquisition ratio, values are the total obtained contents divided by total amount of requests. As the Fig. 5 shows, cache hit ratio is improved by 10% than conventional when the ratio of advanced migrated contents is 10%. The cache hit ratio of proposal is, however, saturated over the ratio of advanced migrated contents is 20%. The reason for the above is that newborn contents are more likely to requested than older contents according to the simulation condition described in Sect. 4. Hence, for increasing cache hit ratio, it is enough to migrate a certain amount of newborn contents. Figure 6 confirms the same characteristics with the Fig. 5. As the figure shows, our proposal improves by approx. 10% than conventional method. Figure 7 and 8 show the results of performance evaluation in terms of RTT (Round Trip Time) and Number of hops for round trip exchange. These figures show that our proposal achieves higher performance than conventional over the 10% of advanced migrated contents. However, improvement ratio is not so high

A Study on Reducing Interest Misleading by Publisher Migration

Fig. 5. Cache hit ratio

Fig. 6. Contents acquisition success ratio

413

414

T. Iwamoto and T. Shigeyasu

Fig. 7. Round Trip Time

Fig. 8. Number of hops

A Study on Reducing Interest Misleading by Publisher Migration

415

compared with the case of the evaluation of cache hit ratio. The reason for that is due to the topology of the evaluation network. Distances among mobile publisher and users requesting contents are same regardless the publisher migration. Therefore, the number of hops and RTT are not improved significantly though the proposal improves these performances little bit by the effects of cache hit at the node 1 by advanced migration at the publisher migration. According to the entire evaluations, we can conclude that our proposal effectively works for continuing communication among the mobile publisher and users against the disconnection due to publisher migration.

6

Conclusion

In this paper, we have discussed mobile publisher problem inducing communication disconnection on NDN on mobile networks. By the results of computer simulations, our proposed method well educes performance of NDN by advance contents migration of mobile publisher in advance its physical location migration. In the future, we will discuss further improvement of our proposal in terms of adequate amount of migration of contents on the basis of network condition and topology.

References 1. Soniya, M., Kumar, K.: A survey on named data networking. In: Proceedings of 2015 2nd International Conference on Electronics and Communication Systems (ICECS), pp. 1515–1519, Coimbatore (2015) 2. Han, D., Lee, M., Cho, K., Kwon, T., Choi, Y.: Publisher mobility support in content centric network. In: Proceedings of IEEE International Conference on Information Networking (ICOIN), pp. 214–219 (2014)

Cyber Attack Scenarios in Cooperative Automated Driving Insu Oh(B) , Eunseon Jeong, Junyoung Park, Taeyoung Jeong, Junghoon Park, and Kangbin Yim Department of Information Security Engineering, Soonchunhyang University, Asan, South Korea {catalyst32,gomdoli10,apple,jtyworld,skyzami,yim}@sch.ac.kr

Abstract. As V2X system has developed, some vehicles that are recently released have been designed to meet the requirements of cooperative automated driving technology. Cooperative automated driving is the method of communication between the car and the surrounding infrastructure, and it predicts the surrounding situation through the V2X communication system to prevent unexpected accidents and situations that external sensors of the car are not aware of. In this environment, external security threats are also possible through each contact points while V2X (Vehicle to Everything) system communicates with Ethernet-based WAVE (Wireless Access in Vehicle Environment) through antenna, GPS module, and OBU (On Board Unit). Those all can generate false messages, for example, which are modulating location data by disturbing the GPS, manipulating the WAVE messages inside the OBU, or injecting internal CAN messages to cause propagation of wrong situation. Based on these potential threats, this paper presents possible cyber-attack scenarios in V2X environment. Keywords: Cooperative automated driving · V2X · Security threats · Cyber-attack scenarios · WAVE

1 Introduction While extending the performance envelope of vehicles, ADAS (Advanced Driver Assistance System) and autonomous vehicles [1, 2] are emerging trends in the automotive space. They bring a number of benefits, including improved safety, reduction of collision risks and road congestions, and higher productivity for drivers. However, it should not be forgotten that autonomous vehicles depend on the data collected through external sensors and therefore possesses potential attacks that can occur by unexpected external sensor values, which means security threats to the autonomous vehicles [3]. Therefore, if an autonomous vehicle communicates only through external sensors, it has difficulties judging the surrounding situation. On the other hand, if they use V2X (Vehicle to Everything) technology, they can send and receive real-time information for driving, such as traffic information, between vehicles and vehicles or between vehicles and surrounding objects. In this case, the autonomous vehicles can deal with unexpected situations © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 416–425, 2021. https://doi.org/10.1007/978-3-030-61108-8_41

Cyber Attack Scenarios in Cooperative Automated Driving

417

or messages from the surrounding infrastructure to overcome physical limitations in detecting distances and obstacles that may cause accidents. Cooperative Automated driving shares current vehicle information and traffic conditions with nearby vehicles or devices through OBU (On-Board Unit) on the vehicle. There are three types of communication for Cooperative automated driving vehicles: (1) V2V (Vehicle to Vehicle), (2) V2I (Vehicle to infrastructure) and (3) V2X (Vehicle to Everything). V2I communication is the wireless exchange of data between vehicles and road infrastructure, such as RSU (Road-Side Unit), and V2X communication is between vehicles and its surroundings [4]. V2X messages are transmitted based on the WAVE standard through an external contact point in the wireless communication section, and there are various possible security threats [5]. For security of V2X communication, it needs additional measures and solutions. This paper will show many types of scenarios about security threats in V2X communication systems, a cooperative automated driving technology, such as triggering emergency notification events through arbitrary CAN messages, transmitting false location information by disturbing GPS and injecting noises into external contact sensors, and transmitting false data through WSM (Wave Short Message) manipulation.

2 Related Work 2.1 Structure of V2X Communication V2X communication technology, a cooperative automated driving technology, is composed of V2V and V2I, depending on the target of the communication. In order to perform V2X communication, it is required to have an antenna and a GPS receiver for collecting data, and an information display unit for presenting the processed data from the OBU. V2X supported vehicles can receive and transmit traffic data and situational events from/to the RSU on the roads and immediately predict a specific situation and make a decision for it. Currently, vehicles do not directly support V2X communication and third party provides additives to add V2X capability into existing vehicles as shown Fig. 1.

Fig. 1. The structure of V2X communication system not inherent to car manufactures

418

I. Oh et al.

In this approach, an internally mounted OBU continuously collects information from the vehicle through the OBD (On Board Diagnostics) interface, transfers it to other vehicles and/or infrastructures nearby, and gathers environmental information from them [6]. However, latest vehicles tend not to expose normal CAN messages except diagnosisrelated ones through the OBD interface. To cover this problem, V2X communication system should be embedded into vehicles in the near future in order for them to share internal situation without OBD interface. There are two communication standards in V2X technology: WAVE (Wireless Access in Vehicle Environment) and C-V2X (Cellular V2X). WAVE is a technology that improves coverage and access time from wireless LAN. C-V2X improves direct communication and resource allocation from LTE in order to make them suitable for communication among vehicles. In the case of LTE-based C-V2X, it is expected to accelerate its development due to recent growth of 5G with low latency [7]. WAVE has been led by the Ministry of Transportation in most countries including the United States and now the Ministry of Land, Infrastructure and Transportation leads pilot services using it. In recent years, researches are also undergoing on hybrid V2X for transferring traffic safety messages and providing high-precision maps for relative services that require a large amount of bandwidth through WAVE [8]. 2.2 WAVE Standard WAVE is defined in IEEE 1609.x standard. It performs based on IEEE 802.11p, a physical and MAC layer standard, which is a derivative of WLAN modified specific for vehicle communication [9]. IEEE 1609.x is a standard for system structure, communication model, management structure, and security mechanism for low-latency communication through WAVE between vehicles [10, 11]. IEEE 1609.2 defines security services for the WAVE application layer and management messages based on crypto standards such as elliptic curve encryption, wireless access certificate in vehicle environment, and hybrid encryption method [12] (Table 1). Table 1. List of Standards in IEEE 1609.x Standard

Define

IEEE 1609.0/D10

WAVE system structure

IEEE 1609.2-2016

Security service of application layer and management message

IEEE 1609.3-2016

WAVE networking service

IEEE 1609.4-2016

WAVE multi-channel service

IEEE 1609.12-2016

WAVE allocation identity

IEEE 1609.3 defines WSMP (WAVE Short Message Protocol) as networking services and supports two different types of V2X communication, which transports WSA (WAVE Service Advertisement) or SAE 2735 DSRC (Dedicated Short-Range Communications) messages (Fig. 2).

Cyber Attack Scenarios in Cooperative Automated Driving

419

Fig. 2. WAVE protocol stack

SAE 2735 DSRC message set includes BSM (Basic Safety Message), PVD (Probe Vehicle Data), EVA (Emergency Vehicle Alert), RSA (Road-Side Alert), TIM (Traveler Information Message) etc., which provides communication among vehicles and between vehicles and infrastructures on the road.

3 Security Threats to Cooperative Automated Driving Security threats to V2X communication exist in various types, from nick on integrity that forgery/modify WAVE messages or sensor information, to hack on availability such as radio interference or DDoS attacks. In particular, vehicles or infrastructure that are targets for V2X communication may receive incorrect information through transmission of false WAVE messages through insertion of unexpected CAN messages, OBU attacks through external contact errors, and manipulation of location information through GPS signal modulation. Because of this, vehicles on V2X communication such as guiding the wrong route and receiving false messages may face security threats [13, 14]. 3.1 Security Threats to OBU by External Error Injection The OBU is the most important component in V2X communication because it transmits/receives messages for communication with external infrastructure. An OBU checks the current location information through the GPS module and the current status information of the vehicle through CAN messages. Based on the collected data, it sends a new WAVE message to the neighboring nodes about the current state and surroundings. Therefore, there is a possibility that the OBU may not work or an incorrect WAVE message may be transmitted due to the influence or security threats on the OBU connected through the external contacts [15].

420

I. Oh et al.

Fig. 3. CAN message propagation through external error injection and OBU security threats

The First Scenario: Because the CAN network is broadcast communication, one CAN message generated is transmitted to various ECU modules. Therefore, as shown in Fig. 3, abnormal CAN messages can be received from the OBU through the OBDII port due to the error injected traffic generated through an external contact. Because of this, there is a possibility of abnormal operation such as generating an error in the OBU and sending incorrect WAVE messages. Abnormal WAVE messages can provide false information to nearby RSU or OBU on other vehicles.

3.2 Generating Fake WAVE Messages by CAN Message Injection CAN messages are transmitted between ECUs to check the vehicle’s functional operation or status. CAN messages do not guarantee encryption or integrity and can be an attack point because they are transmitted through broadcast communication. Therefore, an attacker can manipulate and transmit CAN messages or also retransmit collected CAN messages. The OBU determines the status of the vehicle based on the CAN messages received through the OBDII port. Therefore, when a retransmitted or injected CAN message is received. the OBU will determine the state of the vehicle based on the abnormal information [16].

Cyber Attack Scenarios in Cooperative Automated Driving

421

Fig. 4. Inserting fake CAN messages into OBU by CAN message injection tool

The Second Scenario: As shown in Fig. 4, a CAN message can be injected and transmitted simply by connecting to CAN High and CAN Low. Therefore, it is connected inside the victim vehicle to periodically inject CAN messages generated and collected in advance in the event of a sudden braking situation. Although the brake is not actually activated, the OBU judges only with the CAN messages, so it wrongly recognizes that the vehicle has suddenly stopped, and incorrect WAVE message may be transmitted to neighboring nodes.

3.3 Sending False Events by Manipulating WAVE Messages WAVE communication transmits WAVE messages signed by itself through a certificate built into the OBU based on WSMP (WAVE Short Message Protocol). Therefore, an attacker cannot inject or transmit a message in the middle. However, an attacker can obtain a certificate and modify messages or create a message targeting the one that does not contain authentication information (Fig. 5). The Third Scenario: Accessing the vehicle’s internal system allows collecting WSM messages on the OBU and analyzes SAE J2735 messages. Because of this, the attacker understands the structure of the transmitted messages and there is a possibility of malicious message transmission through message manipulation.

3.4 Sending False Events by Manipulating WAVE Messages Since OBU’s are installed one by one for each vehicle, a method to identify them is required. Meanwhile, in V2X communication, only authenticated messages are allowed, and PKI-based certificates are used to identify from which vehicle each message came. Only certificates issued by a correct subject can be used, and certificates can be extracted through OBU and copied to a compatible OBU device.

422

I. Oh et al.

Fig. 5. Inserting fake messages into OBU by false events

Fig. 6. Fake WAVE message generation and injection testbed

The Fourth Scenario: As shown in Fig. 6, you can compose an environment where WAVE message transmission is possible for V2X communication. Since CAN messages are collected for the information needed in an actual vehicle, a replay is possible based on the CAN messages collected from the vehicle through the OBDII port. Because of this, the duplicated OBU transmits WAVE messages, and neighboring nodes can recognize it as the WAVE messages from the original OBU though it is not a legitimate OBU installed in the actual vehicle.

3.5 Location Information Disturbance by GPS Spoofing In V2X communication, the GPS signal is used to check the location of the current vehicle and nearby OBU or RSU to exchange information. It is possible to attack by

Cyber Attack Scenarios in Cooperative Automated Driving

423

manipulating the vehicle’s location from GPS signal information to provide abnormal locale through modulation of the Local Dynamic Map (LDM), a precision map containing real-time information. The GPS signal calculates the current location by measuring the time difference received through the radio signals of three or more GPS satellites. The GPS signals can be generated through simulation through an open source called “GPS-SDR-SIM”. This enables GPS spoofing simply by entering the latitude, longitude, and altitude of the desired location. In addition, GPS can generate a GPS signal in a moving state by using a CSV file containing the coordinates of a moving object as a list or a data file containing information such as time, position, and orientation called NMEA (National Marine Electronic Association) 0183 [17–20].

Fig. 7. Location information disturbance by GPS spoofing

The Fifth Scenario: It is possible to attack a moving vehicle as shown in Fig. 7. Vehicles attacked through GPS spoofing may confuse drivers because the information displayed on the navigation does not work properly, and incorrect guidance may cause traffic congestion or speeding. There is also the possibility of transmitting a WAVE message in the wrong location from a car that is not actually in that location.

3.6 Location Information Disturbance by GPS Spoofing For the five security threats presented above, we have summarized whether confidentiality, integrity, and availability are infringed, which are the three elements of security. For the security of cooperative automated driving, the conditions for the three security

424

I. Oh et al.

elements must be satisfied. As shown in Table 2. Security threats to the three elements of security, the integrity and availability are violated through most security threats to cooperative automated driving, and WAVE Message transmission by duplicated OBU violates all three. Therefore, there is a need for a countermeasure against attacks that duplicate OBU adding security elements to the OBU, which is the core of V2X technology. Table 2. Security threats to the three elements of security Security threats

The three elements of security Confidentiality

Integrity

Availability

External contact error injection

✓

✓

CAN message injection

✓

✓

Manipulating WAVE message

✓

✓

✓

✓

✓

✓

OBU terminal duplication GPS spoofing

✓

4 Conclusions It is the purpose of this paper to show various security threat scenarios that can occur in the cooperative automated driving environment and encourage researchers think how they can be used for V2X security-related standards or development of countermeasures. It is expected to be helpful for ongoing V2X projects and related researches. Acknowledgements. This research was supported by the MSIT (Ministry of Science and ICT), Korea, under the ITRC (Information Technology Research Center) support program (IITP-20202015-0-00403) supervised by the IITP (Institute for Information & communications Technology Planning & Evaluation) and the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (NRF-2018R1A4A1025632).

References 1. Paul, A., Chauhan, R., Srivastava, R., Baruah, M.: Advanced driver assistance systems (No. 2016-28-0223). SAE Technical paper (2016) 2. Shaout, A., Colella, D., Awad, S.S.: Advanced driver assistance systems-past, present and future. In: International Computer Engineering Conference (ICENCO 2011), pp. 72–82 (2011) 3. Amoozadeh, M., Raghuramu, A., Chuah, C.N., Ghosal, D., Zhang, H.M., Rowe, J., Levitt, K.: Security vulnerabilities of connected vehicle streams and their impact on cooperative driving. IEEE Commun. Mag. 53(6), 126–132 (2015) 4. MacHardy, Z., Khan, A., Obana, K., Iwashina, S.: V2X access technologies: regulation, research, and remaining challenges. IEEE Commun. Surv. Tutor. 20(3), 1858–1877 (2018)

Cyber Attack Scenarios in Cooperative Automated Driving

425

5. Alnasser, A., Sun, H., Jiang, J.: Cyber security challenges and solutions for V2X communications: a survey. Comput. Netw. 151, 52–67 (2019) 6. Jung, H.G., Lim, K.T., Shin, D.K., Yoon, S.H., Jin, S.K., Jang, S.H., Kwak, J.M.: Reliability verification procedure of secured V2X communication for autonomous cooperation driving. In: 2018 International Conference on Information and Communication Technology Convergence (ICTC), pp. 1356–1360. IEEE, October 2018 7. Ghafoor, K.Z., Guizani, M., Kong, L., Maghdid, H.S., Jasim, K.F.: Enabling efficient coexistence of DSRC and C-V2X in vehicular networks. IEEE Wirel. Commun. 27(2), 134–140 (2019) 8. Qi, W., Landfeldt, B., Song, Q., Guo, L., Jamalipour, A.: Traffic differentiated clustering routing in DSRC and C-V2X hybrid vehicular networks. IEEE Trans. Veh. Technol. 69, 7723–7724 (2020) 9. Kenney, J.B.: Dedicated short-range communications (DSRC) standards in the United States. Proc. IEEE 99(7), 1162–1182 (2011) 10. Ahmed, S.A., Ariffin, S.H., Fisal, N.: Overview of wireless access in vehicular environment (WAVE) protocols and standards. environment, vol. 7, p. 8 (2013) 11. Uzcátegui, R.A., De Sucre, A.J., Acosta-Marum, G.: Wave: A tutorial. IEEE Commun. Mag. 47(5), 126–133 (2009) 12. ETSI, TS. ETSI TS 102 867 v1. 1.1-intelligent transport systems (ITS); security; stage 3 mapping for IEEE 1609.2. Standard, TC ITS (2012) 13. Ghosal, A., Conti, M.: Security issues and challenges in V2X: a survey. Comput. Netw. 169, 107093 (2020) 14. Hasan, M., Mohan, S., Shimizu, T., Lu, H.: Securing Vehicle-to-Everything (V2X) communication platforms. IEEE Trans. Intell. Veh. (2020) 15. Woo, S., Jo, H.J., Lee, D.H.: A practical wireless attack on the connected car and security protocol for in-vehicle CAN. IEEE Trans. Intell. Transp. Syst. 16(2), 993–1006 (2014) 16. Oh, I., Kim, T., Yim, K., Lee, S.Y.: A novel message-preserving scheme with formatpreserving encryption for connected cars in multi-access edge computing. Sensors 19(18), 3869 (2019) 17. National Marine Electronics Association. NMEA 0183–Standard for interfacing marine electronic devices. NMEA (2002) 18. Shepard, D.P., Bhatti, J.A., Humphreys, T.E., Fansler, A.A.: Evaluation of smart grid and civilian UAV vulnerability to GPS spoofing attacks. In: Radionavigation Laboratory Conference Proceedings (2012) 19. Ebinuma, T.: Gps-sdr-sim (2018) 20. Ferreira, R., Gaspar, J., Sebastião, P., Souto, N.: Effective GPS Jamming Techniques for UAVs using low-cost SDR platforms. Wirel. Pers. Commun., 1–23 (2020)

Implementation of a User Finger Movement Capturing Device for Control of Self-standing Omnidirectional Robot Kenshiro Mitsugi1 , Keita Matsuo2(B) , and Leonard Barolli2 1

2

Graduate School of Engineering, Fukuoka Institute of Technology (FIT), 3-30-1 Wajiro-Higashi, Higashi-Ku, Fukuoka 811-0295, Japan [email protected] Department of Information and Communication Engineering, Fukuoka Institute of Technology (FIT), 3-30-1 Wajiro-Higashi, Higashi-Ku, Fukuoka 811-0295, Japan {kt-matsuo,barolli}@fit.ac.jp

Abstract. Convenient systems and equipments to support humans are in great need for ever-growing populations of the elderly and those with disabilities caused by illness or injury. One of these system is the wheelchair, which can provide the user with many benefits such as maintaining mobility, continuing or broadening community and social activities, conserving strength and energy, and enhancing quality of life. However, when users use wheelchairs, they have to frequently stand and sit. This increases the physical burden on the user. For this reason, we proposed a self-standing omnidirectional robot. In order to support the user, the robot body must be able to flexibly make different movements and should be capable to deal with various control methods to meet diverse needs. In this paper, we present implementation of a user finger movement capturing device for control of self-standing omnidirectional robot.

1

Introduction

Robots are being steadily introduced into modern everyday life and are expected to play a key role in the near future. Typically, the robots are deployed in situations where it is too dangerous, expensive, tedious, and complex for humans to operate. One of the main features of world population in the 20th century is the increment of elderly people. According to WHO (World Health Organization) by 2025, the increase of population over aged 60 is predicted to reach 23% in North America, 17% in East Asia, 12% in Latin America and 10% in South Asia. There are 1 billion disabled persons in the world constituting nearly 14% of the global population. In recent years, convenient facilities and equipments have been developed in order to satisfy the requirements of elderly people and disabled people. Among them, robot technologies are a common one which is used widely. They can c The Editor(s) (if applicable) and The Author(s), under exclusive license  to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 426–435, 2021. https://doi.org/10.1007/978-3-030-61108-8_42

Implementation of a User Finger Movement Capturing Device for Control

427

provide the user with many benefits such as maintaining mobility, continuing or broadening community and social activities, conserving strength and energy, and enhancing quality of life. In this paper, we present implementation of a user finger movement capturing device for control of self-standing omnidirectional robot. The structure of this paper is as follows. In Sect. 2, we introduce the related work. In Sect. 3, we propose the self-standing omnidirectional robot and present our new control device. In Sect. 4, we evaluate the control device for self-standing omnidirectional robot. Finally, conclusions and future work are given in Sect. 5.

2

Related Work

Most of the work for mobile robots has be done for improving the quality of life of disabled people. One of important research area is robotics. The persons having physical impairment often find it difficult to navigate the moving robot themselves. The reduced physical function associated with the age or disability make independent living more difficult. Many research works have been undertaken to reduce the problem of navigation faced by the physically and mentally challenged people and also older age persons. One of the suggestive measures is the development of a Brain Control Interface (BCI), that assist an impaired person to control the robot using his own brain signal. The research proposes a high-frequency SSVEP-based asynchronous BCI in order to control the navigation of a mobile object on the screen through a scenario and to reach its final destination [2]. This could help impaired people to navigate a robotic wheelchair. The BCIs are systems that allow to translate in real time the electrical activity of the brain in commands to control devices, provide communication and control for people with devastating neuromuscular disorders, such as the Amyotrophic Lateral Sclerosis (ALS), brainstem stroke, cerebral palsy, and spinal cord injury [3]. One of the key issue in designing wheelchairs is to reduce the caregiver load. Some of the research works deal with developing prototypes of moving robots that helps the caregiver by lifting function or which can move with a caregiver side by side [4,5]. The lifting function equipment facilitates easy and safe transfer from/to a bed and a toilet stool by virtue of the opposite allocation of wheels from that for a usual wheelchair. The use of lifting function and the folding of frames makes it more useful in indoor environments. Moving care Robots based on observations of people using integrated sensors can move with a caregiver side by side. This is achieved by a visual-laser tracking technique, where a laser range sensor and an omnidirectional camera are integrated to observe the caregiver. Another important issue for the design of robot is the collision detection mechanism. The omnidirectional robots with collaborative controls ensures better safety against collisions. Such robots possess high level of ability when moving over a step, through a gap or over a slope [6,7]. To achieve omnidirectional motion, vehicles are generally equipped with an omniwheel consisting of a large number of free rollers or a spherical ball wheel. The development of such omniwheels attempts to replace the conventional wheel-type mechanism.

428

K. Mitsugi et al.

There are also other works which deal with vision design of robotic wheelchairs by equipping the wheelchair with camera for monitoring wheelchair movement and obstacle detection and pupil with gaze sensing [8,9]. Prototype for robots have been suggested in various research works, which are exclusively controlled by eye and are used by different users, while proving robust against vibration, illumination change, and user movement [10,11]. To enable older person to communicate with other people the assisting devices have been developed. They can improve the quality of life for the elderly and disabled people by using robot technologies. The head gesture recognition is performed by means of real time face detection and tracking techniques. They developed a useful human-robot interface for RoboChair [12].

3

Proposed Self-standing Omnidirectional Robot

In this section, we describe the implementation of a self-standing Omnidirectional Robot. In our previous work, we implemented an omnidirectional wheelchair as shown in Fig. 1. This wheelchair can move omnidirectionly while keeping the direction, which is very good for different tasks. However, when users use the wheelchair, they have to frequently stand and sit. This increases the physical burden on the user. For this reason, we proposed a self-standing omnidirectional robot (see Fig. 2). In order to confirm the movement of self-standing omnidirectionl robot, we implemented a test model as shown in Fig. 3. This figure shows three omniwheels with small tires. With this structure, the wheel can rotate in front, back, right and left, so it is able to move in all directions. 3.1

Kinematics

For the control of the omnidirectional robot are needed the omniwheel speed, omnidirectional robot movement speed and direction. Let us consider the movement of the omnidirectional robot in 2 dimensional space. In Fig. 4, we show the onmiwheel model. In this figure, there are 3 onmiwheels which are placed 120◦ with each other. The omniwheels are moving in clockwise direction as shown in the figure. We consider the speed for each omniwheel M1, M2 and M3, respectively. As shown in Fig. 4, the axis of the omnidirectional robot are x and y and the ˙ In this case, the moving speed of speed is v = (x, ˙ y) ˙ and the rotating speed is θ. the omnidirectional robot can be expressed by Eq. (1). ˙ V = (x, ˙ y, ˙ θ)

(1)

Based on the Eq. (1), the speed of each omniwheel can be decided. By considering the control value of the motor speed ratio of each omniwheel as linear and synthesising the vector speed of 3 omniwheels, we can get Eq. (2) by using

Implementation of a User Finger Movement Capturing Device for Control

429

Fig. 1. Movement of omnidirectional wheelchair.

Fig. 2. Image of self-standing type omnidirectional robot.

Fig. 3. Test model of self-standing type omnidirectional robot.

Reverse Kinematics, where (d) is the distance between the center and the omniwheels. Then, from the rotating speed of each omniwheel based on Forward Kinematics, we get the omnidirectional robot moving speed. If we calculate the

430

K. Mitsugi et al.

Fig. 4. Model of omniwheel.

inverse matrix of Eq. (2), we get Eq. (3). Thus, when the omnidirectional robot moves in all directions (omnidirectional movement), the speed for each motor (theoretical values) is calculated as shown in Table 1.    M1   1 0          1 √3  M2  =  − − 2    2       √  M3   − 1 3 2 2

  d x˙        d y˙      d  θ˙ 

(2)

Table 1. Motor speed ratio. Direction (degrees)

Motor speed ratio Motor1 Motor2 Motor3

0

0.00

−0.87

0.87

30

0.50

−1.00

0.50

60

0.87

−0.87

0.00

90

1.00

−0.50

−0.50

120

0.87

0.00

−0.87

150

0.50

0.50

−1.00

180

0.00

0.87

−0.87

210

−0.50

−1.00

−0.50

240

−0.87

0.87

0.00

270

−1.00

0.50

0.50

300

−0.87

0.00

0.87

330

−0.50

−0.50

1.00

360

0.00

−0.87

0.87

Implementation of a User Finger Movement Capturing Device for Control

431

Fig. 5. Control system for the motor.

  2   x˙   3 − 13 − 13  M1                 y˙  =  0 − √1 √1  M2      3 3           1  M   θ˙   1 1 3 3d 3d 3d 3.2

(3)

Control System of the Proposed Self-standing Omnidirectional Robot

For the control of the proposed self-standing omnidirectional robot, we considered motor control system. We used brushless motor (BLHM015K-50) and Raspberry Pi3 B+ as a controller. In Fig. 5 is shown the control system for the motor. The implemented motor control system can connect to any control devices with TCP/IP and Bluetooth as shown in Fig. 6.

Fig. 6. Block diagram of control system.

432

4

K. Mitsugi et al.

Implementation of User Finger Capturing Device

In this section, we present our proposed control device using small ([mm] scale) movements of user finger. So far, we have been studying the wheelchair vision and IoT sensors [13,14]. Therefore, we tried to control the robot by following the human finger with a camera using our previous research. Some people with disabilities have a small range of their arm and fingers and can not operate devices like the joystick. Thus, the aim is to operate the robot by capturing small finger movements with a camera. In this way, the operation is easier than the joystick. By using video recognition as controller, the physical burden for users can be reduced [15]. We show the image of the control device using small movement of user finger in Fig. 7. The proposed system can capture the user finger top movement and track the finger orbit. We used this orbit to decide directions for controlling the robot. However, it is difficult to determine user directions, because each user has different habits of their finger movement. In order to predict the user direction, we used machine learning. In Fig. 8 is shown the captured finger top and in Fig. 9 the finger movements to each direction such as top, down, left and right. The system can track the orbit of user finger. We used these data to recognize the user direction.

Fig. 7. Image of control device using small movements of user finger.

Fig. 8. Image of capturing finger at center position.

Implementation of a User Finger Movement Capturing Device for Control

5

433

Experimental Results

In Fig. 10 are shown the user finger orbits for each direction (up, down, left and right). Our proposed control device with image sensor can track the small user finger movement correctly as shown in Fig. 10.

Fig. 9. Image of capturing a finger movement.

Fig. 10. Results of tracing small finger movement by camera.

434

K. Mitsugi et al.

Fig. 11. Experiment results of loss function.

In order to decide the user direction, we used machine learning tool (Scikitlearn). The Scikit-learn is an open source software for machine learning. The Scikitlearn has a number of algorithms that support vector machine, random forest, k-means clustering and neural network. We used neural network for predicting the directions. We show the experimented results of loss function in Fig. 11. When the Time Step is around 18, the Loss value is almost 0. The time step means the number of validations. In this case, our proposed system can predict the user direction. The accuracy of the system for predicting the direction is over 98%.

6

Conclusions and Future Work

In this paper, we proposed a self-standing omnidirectional robot and presented the implementation of user finger movement capturing device. We introduced some of the previous works and discussed their problems and issues. Then, we presented in details the kinematics and proposed self-standing robot. In addition, we have shown the implementation of user finger movement capturing device for the control of self-standing omnidirectional robot. The evaluation results have shown that the proposed system can decide the user directions with a good accuracy. In the future work, we would like to enhance the accuracy and consider other control devices for users.

References 1. Lu, T., Yuan, K., Zhu, H.: Research status and development trend of intelligent wheelchair. Appl. Technol. Robot 2, 1–5 (2008)

Implementation of a User Finger Movement Capturing Device for Control

435

2. Diez, P.F., Mut, V.A., Perona, E.M.A., Leber, E.L.: Asynchronous BCI control using high-frequency SSVEP. J. NeuroEng. Rehabil.D 8(39), 8 (2011). https:// doi.org/10.1186/1743-0003-8-39 3. Grigorescu, S.M., Luth, T., Fragkopoulos, C., Cyriacks, M., Graser, A.: A BCIcontrolled robotic assistant for quadriplegic people in domestic and professional life. Robotic 30(3), 419–431 (2012). Cambridge University PressCambridge University Press 4. Mori, Y., Sakai, N., Katsumura, K.: Development of a wheelchair with a lifting function. Adv. Mech. Eng. 2012, 9 (2012). https://doi.org/10.1155/2012/803014 5. Kobayashi, Y., Kinpara, Y., Shibusawa, T., Kuno, Y.: robotic wheelchair based on observations of people using integrated sensors. In: Proceedings of IEEE/RSJ International Conference on Intelligent Robots and Systems, 11–15 October USA (2009) 6. Ishida, S., Miyamoto, H.: Collision detecting device for omni directional electric wheelchair. Robotics 2013, 8 (2013) 7. Carlson, T., Demiris, Y.: Robotic wheelchair with collaborative control. In: Proceedings of IEEE International Conference on Robotics and Automation, pp. 5582– 5587 (2010) 8. Jia, P., Hu, H.H., Lu, T., Yuan, K.: Head gesture recognition for hands-free control of an intelligent wheelchair. Ind. Robot: Int. J. 34(1), 60–68 (2007). https://doi. org/10.1108/01439910710718469 9. Arai, K., Mardiyanto, R.: Electric wheelchair controlled by eye-only for paralyzed user. J. Robot. Mechatron. 23(1), 66–74 (2011) 10. Escobedo, A., Spalanzani, A., Laugier, C.: Multimodal control of a robotic wheelchair: using contextual information for usability improvement. In: Proceedings of IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS-2013), pp. 4262–4267 (2013). https://doi.org/10.1109/IROS.2013.6696967 11. Gonzalez, J., Munoz, A.J., Galindo, C., Fernandez-Madrigal, J.A., Blanco, J.L.: A description of the SENA robotic wheelchair. In: Proceedings of IEEE Mediterranean Conference (MELECON-2006), pp. 437–440 (2006) 12. Wang, H., Grindle, G.G., Candiotti, J., Chung, C., Shino, M., Houston, E., Cooper, R.A.: The Personal Mobility and Manipulation Appliance (PerMMA): a robotic wheelchair with advanced mobility and manipulation. In: Proceedings of IEEE Engineering in Medicine and Biology Society, pp. 3324–3327 (2012). https://doi. org/10.1109/EMBC.2012.6346676 13. Matsuo, K., Barolli, L., Implementation of omnidirectional wheelchair vision with small reflect mirrors: performance evaluation for tennis ball tracking considering different mirror angles. In: Proceedings of International Conference on Complex Intelligent and Software Intensive Systems (CISIS-2018), pp. 136–148 (2018) 14. Matsuo, K., Kurita, T., Barolli, L.: A new system for management of IoT sensors considering agile-kanban. In: Proceedings of Workshops of the International Conference on Advanced Information Networking and Applications (WAINA-2019), pp. 604–612 (2019) 15. Mitsugi, K., Matsuo, K., Barolli, L.: A comparison study of control devices for an omnidirectional wheelchair. In: Proceedings of the Workshops of the International Conference on Advanced Information Networking and Applications (WAINA2020), pp. 651–661 (2020)

Implementation of Control Interfaces for Moving Omnidirectional Access Point Robot Atushi Toyama1 , Kenshiro Mitsugi1 , Keita Matsuo2(B) , and Leonard Barolli2 1

Graduate School of Engineering, Fukuoka Institute of Technology (FIT), 3-30-1 Wajiro-Higashi, Higashi-Ku, Fukuoka 811-0295, Japan {mgm20105,mgm20108}@bene.fit.ac.jp 2 Department of Information and Communication Engineering, Fukuoka Institute of Technology (FIT), 3-30-1 Wajiro-Higashi, Higashi-Ku, Fukuoka 811-0295, Japan {kt-matsuo,barolli}@fit.ac.jp

Abstract. Recently, various communication technologies have been developed in order to satisfy the user requirements. Especially, mobile communication technology continues to develop rapidly and Wireless Mesh Networks (WMNs) are attracting attention from many researchers in order to provide cost efficient broadband wireless connectivity. The main issue of WMNs is to improve network connectivity and stability in terms of user coverage. In this paper, we introduce a moving omnidirectional access point robot (called MOAP robot) and implementation of control interfaces for the robot. In order to realize moving Access Points (APs), the MOAP robot should move omnidirectional in 2 dimensional space. It is important that the MOAP robot moves to an accurate position in order to have a good connectivity. Thus, MOAP robot can provide good communication and stability for WMNs. The experimental evaluation show that the MOAP robot was controlled correctly by the implemented interfaces.

1 Introduction Recently, communication technologies have been developed in order to satisfy the user requirements. Especially, mobile communication technologies continue to develop rapidly and has facilitated the use of laptops, tablets and smart phones in public spaces [4]. In addition, Wireless Mesh Networks (WMNs) [1] are becoming on important network infrastructure. These networks are made up of wireless nodes organized in a mesh topology, where mesh routers are interconnected by wireless links and provide Internet connectivity to mesh clients. WMNs are attracting attention from many researchers in order to provide cost efficient broadband wireless connectivity. The main issue of WMNs is to improve network connectivity and stability in terms of user coverage. This problem is very closely related to the family of node placement problems in WMNs [5, 8, 10]. In these papers are assumed that routers move by themselves or by using network simulator moving models. In our work, we consider a moving robot as network device. In order to realize a moving Access Point (AP), we implemented a moving omnidirectional access point c The Editor(s) (if applicable) and The Author(s), under exclusive license  to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 436–443, 2021. https://doi.org/10.1007/978-3-030-61108-8_43

Implementation of Control Interfaces

437

robot (called MOAP robot). It is important that the MOAP robot moves to an accurate position in order to have a good connectivity. Thus, the MOAP robot can provide good communication and stability for WMNs. In this paper, we implement two control interfaces for MOAP robot. The rest of this paper is structured as follows. In Sect. 2, we introduce the related work. In Sect. 3, we present our implemented moving omnidirectional access point robot. In Sect. 4, we described implementation of interfaces for moving omnidirectional access point robot. In Sect. 5, we show the implementation results. Finally, conclusions and future work are given in Sect. 6.

2 Related Work Many different techniques are developed to solve the problem of position detection. One of important research area is indoor position detection, because the outdoor position can be detected easily by using GPS (Global Positioning System). However, in the case of indoor environment, we can not use GPS. So, it is difficult to find the target position. Asahara et al. [2] proposed to improve the accuracy of the self position estimation of a mobile robot. A robot measures a distance to an object in the mobile environment by using a range sensor. Then, the self position estimation unit estimates a self position of the mobile robot based on the selected map data and range data obtained by the range sensor. Wang et al. [12] proposed the ROS (Robot Operating System) platform. They designed a WiFi indoor initialize positioning system by triangulation algorithm. The test results show that the WiFi indoor initialize position system combined with AMCL (Adaptive Monte Carlo Localization) algorithm can be accurately positioned and has high commercial value. Nguyen et al. [9] proposed a low speed vehicle localization using WiFi fingerprinting. In general, these researches rely on GPS in fusion with other sensors to track vehicle in outdoor environment. However, as indoor environment such as car park is also an important scenario for vehicle navigation, the lack of GPS poses a

Fig. 1. Implemented MOAP robot.

438

A. Toyama et al.

serious problem. They used an ensemble classification method together with a motion model in order to deal with the issue. Experiments show that proposed method is capable of imitating GPS behavior on vehicle tracking. Ban et al. [3] proposed indoor positioning method integrating pedestrian Dead Reckoning with magnetic field and WiFi fingerprints. Their proposed method needs WiFi and magnetic field fingerprints, which are created by measuring in advance the WiFi radio waves and the magnetic field in the target map. The proposed method estimates positions by comparing the pedestrian sensor and fingerprint values using particle filters. In [6, 7], Matsuo et al. implemented and evaluated a small size omnidirectional wheelchair.

3 Moving Omnidirectional Access Point Robot In this section, we describe the MOAP (Moving Omnidirection Access Point) robot [11]. We show the implemented MOAP robot in Fig. 1. The MOAP robot can move omnidirectionaly keeping the same direction and can provide access points for network devices. In order to realize the MOAP robot, we used omniwheels which can rotate omnidirectionaly in front, back, left and right. The movement of the MOAP robot is shown in Fig. 2. We would like to control the MOAP robot to move accurately in order to offer a good environment for communication. 3.1

Overview of MOAP Robot

Our implemented MOAP robot has 3 omniwheels, 3 brushless motors, 3 motor drivers and a controller. The MOAP robot 24 V battery to move 5 V battery for the controller. We show the specification of MOAP robot in Table 1.

Fig. 2. Movement of our implemented MOAP robot.

Implementation of Control Interfaces

439

Table 1. Specification of MOAP robot. Item

Specification

Length

490.0 [mm]

Width

530.0 [mm]

Height

125.0 [mm]

Brushless Motor BLHM015K-50 (Orientalmotor corporation) Motor Driver

BLH2D15-KD (Orientalmotor corporation)

Controller

Raspberry Pi 3 Model B+

Power Supply

DC24V Battery

PWM Driver

Pigpio (The driver can generate PWM signal with 32 line)

3.2 Motor Control System We designed the motor control system for operation of MOAP robot, which is shown in Fig. 3. We are using brushless motors as main motor to move the robot, because the motor can be controlled by PWM (Pulse Width Modulation). We used Rasberry Pi as a controller. However, the controller has only 2 PWM hardware generators. But, we need to use 3 generators, so we decided to use the software generator to get a square wave for the PWM. As software generator, we use the Pigpio which can generate better signal than other software generators and make PWM signals with 32 lines. Figure 4 shows the square signal generated by Pigpio.

4 Implementation of Interfaces for MOAP Robot In order to control the MOAP robot, we implemented 2 interfaces: user and robot side interfaces, which are shown in Fig. 5. In our previous work, for robot control system, we used RS232C module, which is good to control the motor. However, the communication speed is slow to get information of the brushless motor in real time. Therefor, we used TCP socket communication to drive the brushless motor. The implemented control interfaces consist of Python language on Raspbian OS. The robot side program needs two import files: importFUNCTION.py and importSOCKET.py (see, Fig. 5). The importFUNCTION.py calculates the number of rotation of brushless motor using the pulse which is generated by brushless motor. While, the importFUNCTION.py can monitor the connection state between user and robot side. The importSOCKET.py establishes the connection to user interface and manage the error detection for communication. We show user side and robot side interfaces in Fig. 6 and Fig. 7 respectively. In Fig. 8 is shown the flowchart of robot interfaces. We used thread function which can use many processes in a program at the same time and each process is not affected by other threads. For instance, when one of the thread process is stopped, the other thread process can work continuously. We used 3 threads as F SocketConnect, F MakeScreen and F LedCtrl. The F SocketConnect thread is for managing TCP connection, F MakeScreen thread is able to make control screen and F LedCtrl is in charge of correcting information on the brushless motor and controller.

440

A. Toyama et al.

Fig. 3. Motor Control system for MOAP robot.

Fig. 4. Square signal by using Pigpio.

Fig. 5. Interfaces of user side and robot side.

Implementation of Control Interfaces

441

Fig. 6. User side interface to control MOAP robot.

5 Experimental Evaluation We carried out some experiments to evaluate the implemented interfaces. By user side interface, we can control the motor speed and show the state of motor controller such as CPU temperature, CPU voltage, CPU clock frequency and memory usage. The user interface can change the robot speed and direction by using a slider. By robot side interface can be controlled directly the MOAP robot. The experiments show that the robot is controlled smoothly by using the implemented interfaces.

Fig. 7. Robot side interface to control MOAP robot.

442

A. Toyama et al.

Fig. 8. Flowchart of robot interfaces.

6 Conclusions and Future Work In this paper, we introduced our implemented MOAP robot. We showed some of the previous works and discussed the related problems and issues. Then, we presented in details the kinematics and the control system for our implemented MOAP robot. We implemented two interfaces for MOAP robot. The implemented interfaces can control the MOAP robot smoothly. In the future work, we would like to carry out extensive experiments to evaluate quantitatively the implemented interfaces.

References 1. Akyildiz, I.F., Wang, X., Wang, W.: Wireless mesh networks: a survey. Comput. Netw. 47(4), 445–487 (2005) 2. Asahara, Y., Mima, K., Yabushita, H.: Autonomous mobile robot, self position estimation method, environmental map generation method, environmental map generation apparatus, and data structure for environmental map, US Patent 9,239,580, 19 Jan 2016 3. Ban, R., Kaji, K., Hiroi, K., Kawaguchi, N.: Indoor positioning method integrating pedestrian dead reckoning with magnetic field and WiFi fingerprints. In: 2015 Eighth International Conference on Mobile Computing and Ubiquitous Networking (ICMU), pp. 167–172, January 2015

Implementation of Control Interfaces

443

4. Hamamoto, R., Takano, C., Obata, H., Ishida, K., Murase, T.: An access point selection mechanism based on cooperation of access points and users movement. In: 2015 IFIP/IEEE International Symposium on Integrated Network Management (IM), pp. 926–929, May 2015 5. Maolin, T.: Gateways placement in backbone wireless mesh networks. Int. J. Commun. Netw. Syst. Sci. 2(01), 44–50 (2009) 6. Matsuo, K., Barolli, L.: Design and implementation of an omnidirectional wheelchair: control system and its applications. In: Proceedings of the 9th International Conference on Broadband and Wireless Computing, Communication and Applications (BWCCA-2014), pp. 532–535 (2014) 7. Matsuo, K., Liu, Y., Elmazi, D., Barolli, L., Uchida, K.: Implementation and evaluation of a small size omnidirectional wheelchair. In: Proceedings of the IEEE 29th International Conference on Advanced Information Networking and Applications Workshops (WAINA-2015), pp. 49–53 (2015) 8. Muthaiah, S.N., Rosenberg, C.: Single gateway placement in wireless mesh networks. Proc. ISCN 8, 4754–4759 (2008) 9. Nguyen, D., Recalde, M.E.V., Nashashibi, F.: Low speed vehicle localization using WiFi fingerprinting. In: 2016 14th International Conference on Control, Automation, Robotics and Vision (ICARCV), pp. 1–5, November 2016 10. Oda, T., Barolli, A., Spaho, E., Xhafa, F., Barolli, L., Takizawa, M.: Performance evaluation of WMN using WMN-GA system for different mutation operators. In: 2011 14th International Conference on Network-Based Information Systems, pp. 400–406, September 2011 11. Toyama, A., Kenshiro, M., Matsuo, K., Barolli, L.: Implementation of a moving omnidirectional access point robot and a position detecting system. In: Proceedings of the 14th International Conference on Innovative Mobile and Internet Services in Ubiquitous Computing (IMIS-2020), pp. 203–212 (2020) 12. Wang, T., Zhao, L., Jia, Y., Wang, J.: WiFi initial position estimate methods for autonomous robots. In: 2018 WRC Symposium on Advanced Robotics and Automation (WRC SARA), pp. 165–171, August 2018

Proposal and Experimental Results of an Ambient Intelligence for Training on Soldering Iron Holding Yuto Hirota1(B) , Tetsuya Oda1 , Nobuki Saito1 , Aoto Hirata2 , Masaharu Hirota3 , and Kengo Katatama1 1

2

3

Department of Information and Computer Engineering, Okayama University of Science (OUS), 1-1 Ridaicho, Kita-ku, Okayama 700–0005, Japan {t17j069hy,t17j033sn}@ous.jp, {oda,katayama}@ice.ous.ac.jp Engineering Project Course, Okayama University of Science (OUS), 1-1 Ridaicho, Kita-ku, Okayama 700–0005, Japan [email protected] Department of Information Science, Okayama University of Science (OUS), 1-1 Ridaicho, Kita-ku, Okayama 700–0005, Japan [email protected]

Abstract. In Japan, vocational schools, technical high schools and junior high schools offer classes on soldering with a soldering iron. The use of a soldering iron can be difficult and dangerous for first-time learners. Teachers are limited in their ability to keep track of each practice of students. So, there is a need for a support system for hands-on practice in Japan. In this paper, we propose an ambient intelligence based support system to reduce the risk of soldering practice in the educational field.

1 Introduction In Japan, vocational schools, technical high schools and junior high schools offer classes for circuit implementation. For circuit implementation, the students use soldering iron. However, many students are using the soldering iron for the first time, which can cause burns, expose other students to the soldering iron, or cause circuits to fail. Teachers monitor students to prevent accidents and mistakes, but it is difficult to completely prevent the aforementioned mistakes. One of the reasons why students make soldering mistakes is because they are not holding the soldering iron correctly. If the soldering iron is not held correctly, the soldering iron will not be stable, which will increase the likelihood of mistakes. In this paper, we propose a ambient intelligence [1– 3] based support system for training to hold the soldering iron.

c The Editor(s) (if applicable) and The Author(s), under exclusive license  to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 444–453, 2021. https://doi.org/10.1007/978-3-030-61108-8_44

Proposal of an Ambient Intelligence for Training on Soldering Iron Holding

445

Fig. 1. Structure of proposed system.

2 Proposed System In this section, we present the proposed system. In Fig. 1 and Fig. 2 are shown the structure of proposed system and flow chart.

446

Y. Hirota et al.

Fig. 2. Flowchart of the proposed system.

2.1

Object Recognition

From the streaming video acquired from the USB Video Class (UVC) camera, the coordinates of the key points of the hand and the coordinates of the tip in soldering iron, which is the soldering iron hazard area are obtained. For detection of the tip in soldering iron, we train the cascade classifier using the image of tip in soldering iron. The OpenCV cascade classifier was used to detect the tip in soldering iron. The xml classifier file for the cascade classifier is created with opencv traincascade. The Local Binary Pattern (LBP) [5] based cascade classifier is used in proposed system. The positive image set including the tip in soldering iron and the negative image set including everything we did not want to detect except for the tip in soldering iron were actually prepared by taking pictures with a UVC camera.

Proposal of an Ambient Intelligence for Training on Soldering Iron Holding

447

We developed the highest weight value extracting algorithm for more than one object is detected in cascade classifier, and the proposed system use this algorithm. To determine the dangerous grip of the soldering iron, we consider the hybrid method which the tip in soldering iron is detected by the cascade classifier and the finger is recognized by the Open Source Computer Vision Library (OpenCV) [4] Deep Neural Network (DNN) module. For recognition of finger, the DNN recognize the key points of the finger joints using obtained. The finger joints are touching each other finger joints, the key points of the finger joints of the touching each other are paired, otherwise calculate the difference between coordination of key points in fingertip and finger root joints pair the key points. 2.2 Automation Surprise For automation supplies with voice output near the soldering iron user, the computer for computational intelligence and the computer for voice output are connected wirelessly. The Secure Shell (SSH) connection was based on the Paramiko and the sound in automation surprise was generated using Open JTalk [6]. The computer for computational intelligence obtains an image from the UVC camera and recognizes the object to determine if the finger is close to the tip in soldering iron, if the soldering iron is not dangerous to hold, or if it is not dangerous to hold the iron, or if it is neither. For computer for voice output, we use Google AIY Voice Kit V2. When it is decision that the fingers are close to the tip in soldering iron or the way to hold the soldering iron is dangerous, the voice output for warnings using Open JTalk is executed according to the decision. If a student has a dangerous grip on the soldering iron, an audio output based on automation surprise will be provided. When the coordinates of the key points of the index finger and the coordinates of the tip in soldering iron are at a certain distance, the proposed system emits the voice as automation surprise. We describes the examples of the warning messages in automation surprise. When the distance between the finger and the tip in soldering iron and the coordinate of the is closer than a certain value, the kit emits a voice message “The finger is close to the tip in soldering iron and dangerous”, and when the distance between the finger and the tip in soldering iron is farther than a certain value, the kit emits a voice message “The way you hold the tip in soldering iron is dangerous”.

3 Experimental Results In this section, we describe the experimental results. A picture of the mounting environment is shown in Fig. 3. Table 1 shows the configuration of the experimental environment. 3.1 Experimental Settings As shown in the Fig. 4, a rectangle is generated to identify the object, and if the tip in soldering iron is within the rectangle, it is decision to be the correct answer (Fig. 4(a)).

448

Y. Hirota et al.

Fig. 3. Experimental environment.

If a rectangle is generated as shown in the Fig. 4(b), but the tip in soldering iron is not included in the rectangle or multiple parts are recognized as shown in the Fig. 4(c), it is decision to be a false detection. If the rectangle identifying the object is not generated as shown in the Fig. 4(d), it is decision to be undetected. The opencv traincascade parameters are shown in the Table 2. We describes the opencv traincascade parameters in following: The featureType is the type of features. The numStages is number of cascade stages to be trained. The minHitRate is minimal desired hit rate for each stage of the classifier. The maxFalseAlarmRate is

Proposal of an Ambient Intelligence for Training on Soldering Iron Holding

449

Table 1. Composition of the experimental environment. OS

Ubuntu 18.04.5 LTS

CPU

Intel Core i5-9400F

GPU

GeForce GTX 1050 Ti

Camera

Logicool UVC camera c310h

Voice Output Device

Google AIY Voice Kit V2

Programing Languages Python 3.6.0 OpenCV

Version 4.2.0

Fig. 4. Examples of detection types of tip in soldering iron.

maximal desired false alarm rate for each stage of the classifier. The sampleWidth and sampleHeight is height and width of the sample created by the training samples. All parameters are the same except for the number of images. 3.2 Experimental Scenario 1: Detection of Tip in Soldering Iron Using Cascade Classifier As shown in the Fig. 5, the image with the tip in soldering iron is the positive image (Fig. 5(a)) and the image that contains everything except the tip in soldering iron is a

450

Y. Hirota et al.

Fig. 5. Examples of positive and negative images.

negative image (Fig. 5(b)). The positive images and negative images in a ratio of 1:2, and positive and negative images were combined into a dataset of 1500 [images], 3000 [images], 4500 [images] and 6000 [images]. An evaluation experiment was conducted on 100 [images] containing the tip in soldering iron. In Fig. 6, we show the experimental results not using the highest weight value extracting algorithm and using the highest weight value extracting algorithm. Figure 6(a) shows the experimental result of cascading the tip in soldering iron. Figure 6(b) shows the experimental result of using the highest weight extracting algorithm which determines one object when multiple objects are detected. The experimental results showed that the algorithm to determine one object improved the detection rate by 18 [%] with up to 1500 samples. In both experimental results, the undetected rate increased as the number of samples increased, suggesting that the present experiment is overtrained. Using the example with 1500 [images], which had the highest detection rate, we tested whether the finger coordinates and the coordinates of the center of gravity of the tip in soldering iron are considered dangerous if they fall below a certain value. Table 2. Parameters of the opencv traincascade. featureType

LBP

numStages

15

minHitRate

0.99

maxFalseAlarmRate 0.1 sampleWidth

72

sampleHeight

72

Proposal of an Ambient Intelligence for Training on Soldering Iron Holding

(a) Result not using the highest weight value extracting algorithm.

(b) Result using the highest weight value extracting algorithm.

Fig. 6. Experimental results of the cascade classifier.

451

452

Y. Hirota et al.

Fig. 7. Example of estimation of the state between hand and tip in soldering iro.

3.3

Experimental Scenario 2: Estimation of the State Between Finger and Tip in Soldering Iron Using Cascade Classifier and DNN Hybrid Method

We present the experimental results of estimation of the state between hand and tip in soldering iron using cascade classifier and DNN hybrid method. The cascade classifier and DNN hybrid method get the a finger near the tip in soldering based on the coordinates of the key point of the finger and the center of gravity of the tip in soldering iron as dangerous. For experimental results, 100 [images] in which the index finger is close to the tip in soldering iron is prepared as shown in the Fig. 7. The experimental result show that 76 [%] of the images is found to be dangerous.

4 Conclusion In this paper, we proposed a support system for soldering iron training. In cascade classifier, we find that there is only one object to be detected in the image and developed the highest weight value extracting algorithm for more than one object is detected in cascade classifier. Also, we find that object recognition is possible even when training with few images by using an extracting algorithm that determines the highest weight value of the detected objects when multiple objects are detected. The experimental results show that the proposed system is capable of communicating the inherent dangers of the soldering iron to the user by voice. In the future, we will conduct in other scenarios, such as detection of other incorrectly held objects and support systems.

Proposal of an Ambient Intelligence for Training on Soldering Iron Holding

453

References 1. Acampora, G., Cook, D., Rashidi, P., Vasilakos, A.V.: A survey on ambient intelligence in health care. Proc. IEEE 101(12), 2470–2494 (2013) 2. Ramos, C., Augusto, J., Shapiro, D.: Ambient intelligence - the next step for artificial intelligence. IEEE Intell. Syst. 23, 15–18 (2008) 3. Obukata, R., Cuka, M., Elmazi, D., Oda, T., Ikeda, M., Barolli, L.: Design and evaluation of an ambient intelligence testbed for improving quality of life. Int. J. Space-Based Situ. Comput. 7(1), 8–15 (2017) 4. Bradski, G.: The OpenCV Library. Dr. Dobb’s J. Softw. Tools 120, 122–125 (2000) 5. Kadir, K., Kamaruddin, M., Nasir, H., Safie, S., Bakti, Z.: A comparative study between LBP and haar-like features for face detection using OpenCV. In: Proceedings of The 4-th International Conference on Engineering Technology and Technopreneuship (ICE2T-2014), pp. 335–339 (2014) 6. Open JTalk. http://open-jtalk.sourceforge.net/. Accessed 19 Aug 2020

Design of Education Tool for Reinforcement-Learning Agent Developers Takahiro Uchiya(B) , Kodai Shimano, and Ichi Takumi Nagoya Institute of Technology, Gokiso-cho, Showa-ku, Nagoya, Aichi 466-8555, Japan {t-uchiya,takumi}@nitech.ac.jp, [email protected]

Abstract. Because of the diversification of network service requirements, agentoriented computing is attracting attention as a means of realizing flexible systems. A learning agent adds “learnability”, which produces optimal behavior based on results of past behaviors. The learning agent can perform flexible and efficient actions by learning. Learning agents often deal with problems where uncertain factors such as stochastic events affect the outcome. This uncertain factor adversely affects the learning opportunities to educate the learnability of agents. Furthermore, such problems present obstacles to the training of learning agent system developers. Therefore, we investigated whether a tool that enables efficient education considering uncertain factors such as probabilistic events is effective for the education of learning agent developer.

1 Introduction With the spread of the internet, network services have developed into widely diverse fields. Furthermore, in addition to computers, various devices used in daily life are connected to networks. Services provided through cooperation are becoming increasingly popular. By linking devices over a network, services that meet the diverse needs of users can be provided. Particularly, agent-oriented computing can realize flexible systems that respond to diverse network service requirements. Agent-orientation is a method of creating agents able to operate autonomously by adding a function that changes its own parameters and procedures according to the object’s environment. Some learning agents possess added “learnability” to learn optimal behavior from results of past behaviors [1]. This learning property enables the learning agent to perform flexible and efficient operations. Many difficulties that learning agents solve by demonstrating their learning ability are often affected by uncertain factors such as stochastic events. This uncertain factor adversely affects the learning opportunities of agents. Opportunities for the education of learning abilities are lost, thereby hindering the development of learning agent system developers. The authors hold that “tools which can teach efficiently by considering uncertain factors such as probabilistic phenomena” are effective for training developers of learning agent systems. For this study, we propose an educational tool for training developers of learning agent systems. The proposed tool provides users with examples that incorporate uncertain factors when enhancing the learning ability of learning agents and provides users with applications that can perform education efficiently based on the examples. © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 454–462, 2021. https://doi.org/10.1007/978-3-030-61108-8_45

Design of Education Tool for Reinforcement-Learning Agent Developers

455

2 Proposed Method For this study, we propose an educational tool for training developers of learning agent systems to facilitate learning by system developers (Fig. 1). This tool includes an example implemented through the IDEAL [2, 3] agent development environment and an execution result monitor for checking the execution result of the example. The user obtains empirical knowledge of learning agent system development from the execution result monitor.

Fig. 1. Overview of proposed tool.

2.1 Empirical Knowledge When using a learning agent system, education for adjusting the learning rate, discount rate, action selection method, and other parameters is important for training learning agent developers because those parameters must be adjusted according to the environment in which they are used. However, because the learning agent system can be used in innumerable environments, the combinations of environment and learning rate, discount rate, action selection method, etc. are vastly numerous. It is inappropriate to teach all environments and combinations of parameter values that are appropriate for that environment. Therefore, for the learning rate, discount rate, and action selection method education, it is necessary to accumulate empirical knowledge by learning through actually operating the system. For this study, the agent’s action selection method is treated as empirical knowledge acquired by the user. 2.2 Action Selection Method Action selection in the learning agent system might be unable to select an optimum action because it converges to a local solution. A method exists for performing random action selection with a certain frequency in action selection methods. Using such an action selection method, a behavior other than good behavior is taken temporarily with

456

T. Uchiya et al.

a certain frequency by random selection. The system can escape from the local solution by seeking good behavior. However, if the search frequency is too high, then the number of random selections becomes excessive. The learning effect becomes weak. The local solution cannot be avoided if it is too low. Therefore, learning agent system development is important to determine the search frequency of the action selection method suitable for the environment. For this research, we strive to enable users of the proposed tool to understand the action selection method and to adjust the search frequency appropriately by working on examples. 2.3 Multi-armed Bandit Problem The multi-armed bandit problem is used as an example for learning the action selection method. The multi-armed bandit problem posits multiple slot machines with different winning probabilities. After the agent selects a slot machine, a lottery is performed based on the winning probability. Then a reward is paid if the player wins. The agent acts with the aim of maximizing the reward. In the multi-armed bandit problem, when a local solution is encountered, a platform with a low winning probability is selected such that the user can readily recognize the state of the local solution. Because the goodness or badness of the result can be judged in light of the acquired reward, performing parameter adjustment training of the action selection method is suitable as an example. 2.4 Local Solution and Escape of the Multi-armed Bandit Problem The greedy method, a method by which an agent continues to select actions that it considers to be a good condition, is the simplest action selection method. When this greedy method is used, the multi-armed bandit problem might fall into a local solution. When the greedy method is used for the multi-armed bandit problem, if a player with a low probability of winning is selected first and then wins, then that player with a lower probability of winning will continue to be drawn. The solution will converge to a local solution. Therefore, an action selection method called epsilon-greedy method is applied to avoid that local solution. The epsilon-greedy method is an action selection method that takes random actions with low probability rather than continuing to select actions that the agent regards as good conditions, unlike the greedy method. When applying this method for the multi-armed bandit problem, whether the agent misidentifies a platform with a low winning probability as having a high winning probability, it can correct the misidentification by random action selection. Therefore, it can escape from the local solution.

3 Design of Example on Proposed Tool 3.1 Consideration When using the proposed tool, the user works on examples to elucidate the role of exiting the local solution of the action selection method, to highlight the importance of adjusting the search frequency of the appropriate action selection method, and to learn

Design of Education Tool for Reinforcement-Learning Agent Developers

457

the adjustment method. For those purposes, when the user executes the example program, it tends to fall into a local solution when the search frequency is not adjusted. One must escape from the local solution and obtain good results by adjusting the search frequency. However, in many cases that involve environments for which learning agent systems must be flexible, included elements such as stochastic events cannot be controlled completely. Therefore, because complete control is impossible when using a learning agent system as an example, the possibility exists that results do not match the intention of the questioner, i.e., results do not match what one wants to learn. Even in the multi-armed bandit problem, the winning probability is set for each slot machine. The result changes according to the influence of the probability. Therefore, when the multi-armed bandit problem is used as an example, complete control of the result is impossible. Depending on the winning probability to be set, even if the user executes the multi-armed bandit problem, it does not fall into a local solution. The necessity of adjusting the search frequency is not conveyed: even if the search frequency is adjusted, it is not possible to escape from the local solution. Therefore, search frequency adjustment training cannot be performed. Consequently, the effectiveness of the education using the examples is reduced. This study assesses production of ideal results for education as frequently as possible by adjusting the learning rate and the winning probability setting. In doing so, one must study an example in which the user can learn a parameter adjustment method for escaping from a local solution and a search frequency for the action selection method. 3.2 Design The multi-armed bandit problem setting is described. First we create an example in which the user can learn how to adjust parameters affecting the escape frequency from the local solution and affecting the search frequency of the action selection method with high frequency. The following experiments were conducted to determine the winning probability and learning rate values set in the example. [Experiment 1] Determine the appropriate learning rate for the winning probability. [Experiment 2] Confirm that escape from the local solution occurs by changing the search frequency for the determined winning probability and learning rate. In the experiment, Q-learning is used as the reinforcement learning algorithm used in the learning agent system. The epsilon-greedy method is used as the action selection method. We use other parameters: number of slot machines, 3; reduction rate, 1.0; and number of learning iterations, 1000. From Experiments 1 and 2, as examples for users to work with the proposed tool, we use the multi-armed bandit problem with three slot machines. The respective winning probabilities are 60%, 10%, and 10%. The learning rate is 0.4. The discount rate is set to 1.0. Based on the results, the user can learn about the action selection method while working on examples and changing the search frequency.

458

T. Uchiya et al.

3.3 Execution Result Monitor The execution result monitor is designed to raise the efficiency of confirmation of execution results when the user works on an example. The main functions of the execution result monitor are presented below. (1) Winning probability display (Fig. 2) The winning probability set for each unit is displayed. (2) Display of unit selection rate (Fig. 2) Each unit displays the selected percentage. (3) Number of learning (Fig. 2) Display the number of learning times set for the multi-armed bandit problem. In addition, the execution result monitor displays information other than the graph while updating it for each learning count. The monitor also displays how many learning results are output to the current monitor. (4) Search rate (Fig. 2) The search rate, a parameter of the action selection method that handles the search frequency, is displayed. Because the example uses the epsilon-greedy method as the action selection method, the value of the parameter epsilon that handles the search frequency is displayed. (5) Table of results of selecting units (Fig. 2) A table is displayed showing the selected platform, lottery results for that platform, and whether the factors that selected that platform were attributable to the priority obtained through reinforcement learning or attributable to random selection by the action selection method. When a table other than that with the highest winning probability is selected, the cell of the selected table is displayed in red. When the agent’s action selection is random, the selection method cell is displayed as gray. (6) SKIP function (Fig. 2) A button outputs the result at the end of learning by omitting the update for each learning count of the execution result monitor. (7) Acquisition reward graph (Fig. 3) A graph is displayed with the vertical axis representing the average earned reward per trial and the horizontal axis representing the number of learning iterations.

Design of Education Tool for Reinforcement-Learning Agent Developers

459

(8) Graph of the number of selections for each unit (Fig. 3) Graphs for respective platforms are displayed with the vertical axis representing the number of platform selections and the horizontal axis representing the number of learning iterations. (9) Graph switching function (Fig. 3) Button to switch the graph display of earned rewards and a graph of the number of selections of each unit. (10) Execution result history table (Fig. 2) Among the executions of the past examples, the execution probability was set to the same winning probability as that currently set for each machine. The search rate set at

Fig. 2. Execution result monitor: part I.

460

T. Uchiya et al.

that time and the average of the rewards earned per trial at the end of learning were executed. The history is displayed as a table.

Fig. 3. Execution result monitor: part II.

Design of Education Tool for Reinforcement-Learning Agent Developers

461

(11) Average of the past 10 earned rewards (Fig. 2) The average of earned rewards per trial is displayed by averaging 10 times from the newest result in the execution result history. In addition, the average of the target earned rewards per trial is shown. The average of the earned rewards per trial, which is averaged over 10 times from the newest result history, exceeds the target value. This is a guideline for working on the example.

4 Experiment and Evaluation We conducted an experiment to evaluate the proposed educational tool with five university students as subjects. The questionnaire results were evaluated on a scale of 5, with 5 being the highest value. The evaluation items are shown below. (1) (2) (3) (4) (5)

Convenience of table of selection results of units Convenience of execution result history table Convenience of graph of average earned reward for one trial Convenience of the graph of the number of selections of each unit Convenience of displaying acquired rewards

Table 1 presents results of the questionnaire on the execution result monitor of the proposed tool. Table 1. Questionnaire results Question

Average

Variance

Convenience of table of selection results of units

4.6

0.64

Convenience of execution result history table

4.2

0.56

Convenience of graph of average earned reward for one trial

3.8

0.56

Convenience of the graph of the number of selections of each unit

3.2

2.56

Convenience of displaying acquired rewards

3

0

The table of the selection results of the machines shows an average of 4.6. The table of the history of execution results shows an average of 4.2: a high value. The table of table selection results shows the table (table A, table B, table C) selected by the agent, the lottery result (win/lose), and the table selection method (random/priority). The value of epsilon set in the past execution and the average of the rewards earned per trial are displayed in the history table. Because both tables provide information that cannot be obtained from other sources, they are highly evaluated because they are used by many subjects and because they are of high importance. The average earned reward graph for one trial has an average evaluation of 3.8. This graph is displayed in the initial state. It is easy to see. One can roughly judge

462

T. Uchiya et al.

whether learning is successful. However, the evaluation might not be so high because obtaining detailed information was impossible. The graph of the number of selections for each unit shows the average of evaluations as 3.2. The variance is 2.56: a large variance. This result suggests that large individual differences exist in the handling of this information. The drop or escape to the local solution can be judged by checking the table of the platform selection results, but if falling into a local solution at the middle of learning, one can confirm the drop or escape to a local solution. Checking the table of the selection results of the units requires much work such as scrolling the table. However, when checking the graph of the number of selections for each machine, it can be judged quickly and intuitively. Therefore, the evaluation was considered to be divided mainly between subjects who used such a graph and those who did not. The average of the evaluations of the display related to earned rewards is 3. The variance is 0. Because this information is mainly a guideline for working on examples, the evaluation was regarded as low for all subjects because the learning was not connected directly to the action selection.

5 Conclusion This study was conducted to support efficient education using examples that incorporate uncertain factors when ascertaining the learning ability of learning agents. To achieve this purpose, we proposed and evaluated an educational tool for training developers of learning agent systems. We administered a questionnaire to evaluate the execution result monitor. Components that provided information which was unobtainable from other components were highly evaluated, but evaluation values of some components exhibited high variance. Results show that large individual differences arise in how they are used. Future studies will therefore address a new multi-armed bandit problem suitable for adjusting the search frequency.

References 1. Hibino, M., Uchiya, T., Takumi, I., Kinoshita, T.: Development tool of nash-Q learning agent for intelligent system. In: 18th International Conference on Network-Based Information Systems (NBiS), pp. 581–585 (2015) 2. Uchiya, T., Itazuro, S., Takumi, I., Kinoshita, T.: IDEAL: interactive design environment for agent system with learning mechanism. In: Proceedings of the 12th IEEE ICCI*CC, pp. 153– 160 (2013) 3. Uchiya, T., Maemura, T., Li, X., Kinoshita, T.: Design and implementation of interactive design environment of agent system, IEA/AIE 2007. LNAI 4570, 1088–1097 (2007)

Author Index

A Ampririt, Phudit, 11, 386 B Barolli, Admir, 11 Barolli, Leonard, 1, 11, 233, 296, 386, 426, 436 Bylykbashi, Kevin, 233 C Chellappan, Sriram, 76 D Dey, Arup Kanti, 76 Duolikun, Dilawaer, 341 E Enokido, Tomoya, 22, 34, 44, 330, 341 F Fan, Rourong, 365, 376 Fan, Yao-Chung, 265 Fujisaki, Kiyotaka, 304 Funabiki, Nobuo, 154 G Gao, Qiang, 97 Goel, Bharti, 76 Guo, Yinzhe, 22 H Han, Linshan, 365 Hayashibara, Naohiro, 87 Henmi, Kenta, 188 Hirata, Aoto, 67, 321, 355, 444

Hirota, Masaharu, 67, 321, 355, 444 Hirota, Yuto, 321, 444 Huang, Tz-Yuan, 244 Huda, Samsul, 154 Hung, Li-Ling, 275 I Ikeda, Makoto, 1, 233, 296, 386 Ishihara, Nobuya, 154 Itokazu, Takuya, 166 Iwamoto, Taichi, 407 J Jeong, Eunseon, 397, 416 Jeong, Taeyoung, 416 Jia, Qizhen, 376 Jing, Maohua, 365, 376 Jung, Soyoung, 397 K Kanahara, Kazuho, 211 Kao, Wen-Chun, 154 Katatama, Kengo, 444 Katayama, Kengo, 67, 211, 321, 355 Kim, Chanmin, 397 Kim, Minseong, 397 Kohata, Masaki, 204 Kojima, Rikuhiro, 284 Koyama, Akio, 188 Kuribayashi, Minoru, 154 Kurokawa, Takeru, 87 L Lai, Sen-Tarng, 254 Leu, Fang-Yie, 244, 254, 265, 275

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 L. Barolli et al. (Eds.): BWCCA 2020, LNNS 159, pp. 463–464, 2021. https://doi.org/10.1007/978-3-030-61108-8

Author Index

464 Li, Lei, 22 Lin, Frank Yeong-Sung, 223 Lin, Wen-Yao, 223 Luan, Jingmin, 365 M Maeda, Hiroshi, 313 Matsuo, Keita, 233, 426, 436 Mimura, Mamoru, 176 Mitsugi, Kenshiro, 426, 436 Miyake, Takafumi, 211 Mukae, Keigo, 330 N Nagai, Yuki, 67 Nakamura, Shigenari, 22, 34, 330 Ngoc, Phung Minh, 176 Nguyen, Linh Vu, 120 Nimura, Kazuaki, 284 Nishii, Daisuke, 296 Noaki, Naomichi, 341 Nomi, Masaki, 143

Sakamoto, Shinji, 11 Sakuraba, Akira, 131 Sato, Goshi, 131 Shibata, Masahiro, 120 Shibata, Yoshitaka, 131 Shigeyasu, Tetsuya, 97, 407 Shimano, Kodai, 454 Shimoyama, Takeshi, 284 Sudibyo, Rahardhita Widyatra, 154 Sugawara, Shinji, 166 T Tada, Yoshiki, 1 Tai, Kuang-Yen, 223 Takano, Keisuke, 204 Takizawa, Makoto, 11, 22, 34, 44, 199, 233, 330, 341, 386 Takumi, Ichi, 454 Tomita, Etsuji, 211 Toyama, Atushi, 436 Tsai, Kun-Lin, 244 Tsuru, Masato, 120

O Oda, Tetsuya, 67, 204, 321, 355, 444 Ogiela, Lidia, 199 Ogiela, Urszula, 199 Oh, Insu, 416 Ohara, Seiji, 11, 386 Okada, Yoshihiro, 143 Ozaki, Ryo, 204

U Uchida, Noriki, 131 Uchiya, Takahiro, 454 Uehara, Minoru, 55 Uejima, Akira, 204

P Park, Park, Park, Park,

X Xie, Kang, 365, 376

Hyunhee, 108 Junghoon, 416 Junyoung, 397, 416 Seunghyun, 108

Q Qafzezi, Ermioni, 233, 386 S Saito, Nobuki, 67, 321, 355, 444 Saito, Takumi, 22, 34, 330, 341

W Wu, Ting-Huan, 223

Y Yamamoto, Dai, 284 Yang, Hao-En, 244 Yang, Tao, 365, 376 Yasaki, Kouichi, 284 Yim, Kangbin, 397, 416 Yoong, Siang Yun, 265 Yoshigai, Yuki, 304 Yu, Qilong, 376