Collective behavior in multiagent systems based on Reinforcement Learning

Abstract

In this work would be considered multiagent approach to solve intellectual tasks based on Reinforcement Learning. The multiagent approach involves agent teamwork for the solution of the problem. But classical RL supports for single agent learning. Therefore, the key question is how to modify the reinforcement learning for a group of interacting agents. In this work would be considered a modified algorithm for supporting training for a group of agents. As an example of the problem we will take a coordinated movement in the space group of agents. As a result of training were observed interesting patterns of behavior of groups, such as «leader», «chain of action», «clustering».