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

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.