Expert knowledge based multi-agent reinforcement learning and its application in multi-robot hunting problem

Abstract

We propose a novel reinforcement learning algorithm based on expert knowledge to further solve curse of dimensionality and accelerate the convergence when solving multi-robot hunting problem. Two kinds of expert knowledge are designed: one is the multi-agent joint state abstraction method based on dynamic ID, which dramatically reduces the number of state space; the other is the multi-agent Q-learning algorithm based on artificial potential field (APF). In this method, the artificial potential field is utilized to initialize the Q value according to the environmental prior knowledge, hence the robot can acquire a better learning foundation and accelerate convergence. Finally, the proposed algorithm is validated by the multi-robot hunting problem. The results show that the method can reduce the number of state space and accelerate convergence, thus improve the performance of the algorithm.