Distributed multi-agent deep reinforcement learning for cooperative multi-robot pursuit

As a popular research topic in the area of distributed artificial intelligence, the multi-robot pursuit problem is widely used as a testbed for evaluating coordinated and cooperative strategies in multi-robot systems. This study the problem of multi-robot pursuit game using reinforcement learning (RL) techniques is studied. Unlike most existing studies that apply fully centralised deep RL methods based on the centralised-learning and decentralised-execution scheme, the authors propose a fully decentralised multi-agent deep RL approach by modelling each agent as an individual deep RL agent that has its own individual learning system (i.e. individual action-value function, individual leaning update process, and individual action output). To realise coordination among agents, the limited information of other environmental agents is used as input of the learning process. Experimental results show that both distributed and centralised approaches can ultimately solve the pursuit-evasion problem in different dimensions, but the learning efficiency and coordination performance of the proposed distributed approach are much better than the traditional centralised approach.