Learning strategic group formation for coordinated behavior in adversarial multi-agent with double DQN

Elhadji Amadou Oury Diallo, Toshiharu Sugawara

    Research output: Chapter in Book/Report/Conference proceedingConference contribution

    2 Citations (Scopus)

    Abstract

    We examine whether a team of agents can learn geometric and strategic group formations by using deep reinforcement learning in adversarial multi-agent systems. This is a significant point underlying the control and coordination of multiple autonomous and intelligent agents. While there are many possible approaches to solve this problem, we are interested in fully end-to-end learning method where agents do not have any prior knowledge of the environment and its dynamics. In this paper, we propose a scalable and distributed double DQN framework to train adversarial multi-agent systems. We show that a large number of agents can learn to cooperatively move, attack and defend themselves in various geometric formations and battle tactics like encirclement, guerrilla warfare, frontal attack, flanking maneuver, and so on. We finally show that agents create an emergent and collective flocking behaviors by using local views from the environment only.

    Original languageEnglish
    Title of host publicationPRIMA 2018
    Subtitle of host publicationPrinciples and Practice of Multi-Agent Systems - 21st International Conference, 2018, Proceedings
    EditorsNir Oren, Yuko Sakurai, Itsuki Noda, Tran Cao Son, Tim Miller, Bastin Tony Savarimuthu
    PublisherSpringer-Verlag
    Pages458-466
    Number of pages9
    ISBN (Print)9783030030971
    DOIs
    Publication statusPublished - 2018 Jan 1
    Event21st International Conference on Principles and Practice of Multi-Agent Systems, PRIMA 2018 - Tokyo, Japan
    Duration: 2018 Oct 292018 Nov 2

    Publication series

    NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
    Volume11224 LNAI
    ISSN (Print)0302-9743
    ISSN (Electronic)1611-3349

    Other

    Other21st International Conference on Principles and Practice of Multi-Agent Systems, PRIMA 2018
    CountryJapan
    CityTokyo
    Period18/10/2918/11/2

    Keywords

    • Collective intelligence
    • Cooperation
    • Coordination
    • Deep reinforcement learning
    • Multi-agent systems

    ASJC Scopus subject areas

    • Theoretical Computer Science
    • Computer Science(all)

    Fingerprint Dive into the research topics of 'Learning strategic group formation for coordinated behavior in adversarial multi-agent with double DQN'. Together they form a unique fingerprint.

  • Cite this

    Diallo, E. A. O., & Sugawara, T. (2018). Learning strategic group formation for coordinated behavior in adversarial multi-agent with double DQN. In N. Oren, Y. Sakurai, I. Noda, T. Cao Son, T. Miller, & B. T. Savarimuthu (Eds.), PRIMA 2018: Principles and Practice of Multi-Agent Systems - 21st International Conference, 2018, Proceedings (pp. 458-466). (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 11224 LNAI). Springer-Verlag. https://doi.org/10.1007/978-3-030-03098-8_30