A Supervised Learning Method for Chinese Chess Programs

Abstract

The key factor in strong game-playing programs is the quality of the evaluation function which is usually computed based upon a large number of features. The tradition method is to manually set and adjust each feature weight by masters, incurring the drawback: consume a lot of time and difficultly get precise weights. Numerous high-quality game records of Chinese chess have been collected from the Internet, and can be used for adjusting the present features or even discovering new ones for the evaluation function. In this paper, we design an adjustment system for Chinese chess to automatically improve feature values based on the method of computing Elo rating with Minorization-Maximization algorithm. Furthermore, while using move selections combined with quiescence search, the judgment of the piece mobility becomes more accurate. We also designed new composite features whose weights were tuned by our system. The experiment results showed that our system effectively enhanced the strength of Chinese chess programs, reaching the win rate of 61.7% using self-play by CHIMO, which won the silver medal in the Chinese chess tournament in the 16th Computer Olympiad.