Competitive Evolution Multi-Agent Deep Reinforcement Learning

Abstract

As an effective method to solve the optimal policy in multi-agent systems, multi-agent deep reinforcement learning (MADRL) has achieved impressive results in many applications. However, the training of a deep neural network used in MADRL is time-consuming and laborious. In order to improve the training efficiency and accelerate the training speed of the neural network, this paper proposed a sharing learning framework to support the effectively training of multi-agent state-action value function neural network through making full use of sharing diverse experience from different agents to train a single shared network. Besides, a novel learning algorithm combined with competition and evolution idea was developed to accelerate the training process and improve the performance of the trained neural network. Finally, the experiment results prove the effectiveness of the proposed framework and the fact that the proposed algorithm with an appropriate update rate can indeed accelerate the training process and improve the performance of the trained network by providing a well-matched opponent.