A Multi-group Multi-agent System Based on Reinforcement Learning and Flocking

In this paper, we present an inter-group confrontation and intra-group cooperation method for a predator group and prey group, and construct a multi-group multi-agent system. We model the motion of the prey group using the flocking control algorithm. The prey group can cooperatively avoid predators and maintain the integrity of the group after the predators have been detected. The autonomous decision-making of the predator group is implemented based on the distributed reinforcement learning algorithm. To efficiently share the learning experience among agents in the predator group, a distributed cooperative reinforcement learning algorithm with variable weights is proposed to accelerate the convergence of the learning algorithm. Simulations show the feasibility of this proposed method.