象棋程式之特徵學習

Abstract

電腦象棋的審局是對電腦象棋棋力有重大的影響。然而審局是由大量的特徵計算加總而成，每個特徵都需要給予一個權重值。傳統的作法是以人工不斷調整特徵權重，費時而難以準確。 本篇論文研究象棋審局之特徵權重，利用Minorization-Maximization演算法與Elo等級分的方式，設計一套適用於象棋程式之自動調整適當特徵權重的系統。並實作新的特徵使用此系統做參數的調整。再利用對戰的方式，觀察棋力的變化。實驗的結果顯示，本論文提出的方式能有效提升電腦象棋的棋力。

The evaluation function has a significant effect on the playing strength of a Chinese chess program. However the evaluation function is usually computed by a large number of features with different feature weight values. The traditional approach is to manually adjust the weight values by a trial and error process, but it consumes a lot of time. This paper is a study of the feature weight of evaluation function in Chinese chess program. We designed a system to automatically adjust the feature weight values which are in Chinese chess program by computing Elo rating with a Minorization-Maximization algorithm. We also implemented new features and assigned the weight values by this system. Then we estimate the performance by playing against the original version. And the result comes that the method this paper presents can enhance the playing strength of the Chinese chess program effectively.