TY - GEN
T1 - Learning strategic group formation for coordinated behavior in adversarial multi-agent with double DQN
AU - Diallo, Elhadji Amadou Oury
AU - Sugawara, Toshiharu
N1 - Funding Information:
This work is partly supported by JSPS KAKENHI Grant Number 17KT0044.
PY - 2018
Y1 - 2018
N2 - 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.
AB - 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.
KW - Collective intelligence
KW - Cooperation
KW - Coordination
KW - Deep reinforcement learning
KW - Multi-agent systems
UR - http://www.scopus.com/inward/record.url?scp=85056447233&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85056447233&partnerID=8YFLogxK
U2 - 10.1007/978-3-030-03098-8_30
DO - 10.1007/978-3-030-03098-8_30
M3 - Conference contribution
AN - SCOPUS:85056447233
SN - 9783030030971
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 458
EP - 466
BT - PRIMA 2018
A2 - Oren, Nir
A2 - Sakurai, Yuko
A2 - Noda, Itsuki
A2 - Cao Son, Tran
A2 - Miller, Tim
A2 - Savarimuthu, Bastin Tony
PB - Springer Verlag
T2 - 21st International Conference on Principles and Practice of Multi-Agent Systems, PRIMA 2018
Y2 - 29 October 2018 through 2 November 2018
ER -