Abstract

 
Chinese chess has long been viewed as one of the most popular board games in China. It has a larger state and action space than chess, hence greater difficulty for AI to conquer. Many previous work focused on search based algorithm or simple TD learning to tackle Xiangqi. However, in this project, we propose a deep reinforcement learning based algorithm inspired by AlphaGo. We first used supervised learning to initialize player agent and use reinforcement learning algorithms to update the players against a commercial Xiangqi agent called Elephant Eye. We are able to achieve a consistent 56\% win rate over Elephant Eye evaluated in 100 games.

DeepShuai: Deep Reinforcement Learning based Chinese Chess Player