中国科学院自动化研究所机构知识库

Non-stationary environment is general in real environment, including adversarial environment and multi-agent problem. Multi-agent environment is a typical non-stationary environment. Each agent of the shared environment must learn a efficient interaction for maximizing the expected reward. Independent reinforcement learning (InRL) is the simplest form in which each agent treats other agents as part of environment. In this paper, we present Max-Mean-Learning-Win-or-LearnFast (MML-WoLF), which is an independent on-policy learning algorithm based on reinforcement clustering. A variational autoencoder method based on reinforcement learning is proposed to extract features for unsupervised clustering. Based on clustering results, MML-WoLF uses statistics and the dominated factor to calculate the values of the states that belong to a certain category. The agent policy is iteratively updated by the value. We apply our algorithm to multi-agent problems including matrixgame, grid world, and continuous world game. The clustering results are able to show the strategies distribution under the agent's current policy. The experiment results suggest that our method significantly improves average performance over other independent learning algorithms in multi-agent problems.