Reducing Q-Value Estimation Bias via Mutual Estimation and Softmax Operation in MADRL

Abstract

With the development of electronic game technology, the content of electronic games presents a larger number of units, richer unit attributes, more complex game mechanisms, and more diverse team strategies. Multi-agent deep reinforcement learning shines brightly in this type of team electronic game, achieving results that surpass professional human players. Reinforcement learning algorithms based on Q-value estimation often suffer from Q-value overestimation, which may seriously affect the performance of AI in multi-agent scenarios. We propose a multi-agent mutual evaluation method and a multi-agent softmax method to reduce the estimation bias of Q values in multi-agent scenarios, and have tested them in both the particle multi-agent environment and the multi-agent tank environment we constructed. The multi-agent tank environment we have built has achieved a good balance between experimental verification efficiency and multi-agent game task simulation. It can be easily extended for different multi-agent cooperation or competition tasks. We hope that it can be promoted in the research of multi-agent deep reinforcement learning.