麻將人工智慧之研究

Abstract

本論文提出一套改良過的演算法，主要利用麻將轉張的概念來改進程式。我們先提出一個對手只會摸一張打一張的簡化遊戲模式，並在此模型下設計含有轉張概念的期望最大值演算法來考慮轉張的情況以解決問題。我們分別實驗在此遊戲模式下程式的強度並與過去程式的強度比較。在實驗中，轉張能有效的提升胡牌率。在實際比賽中，新程式與過去程式進行了192局勝負之後，贏得14000分。此外，此新程式獲得了2015年TCGA電腦麻將程式比賽冠軍及2015年ICGA電腦麻將程式比賽冠軍。

In this paper, we purpose a modified algorithm. We use the relaxed-move concept in Mahjong to improve the strength. We purpose a simplified model. In this simplified model, all opponents can only take tiles and discard tiles in a Mahjong game. In this simplified model, we design the expectimax algorithm which contain relax concept. We experiment the strength of different AI in this simplified model and compete with the previous AI program. In the experimental result, the program relaxed-move can effectively raises winning rates. In contests, our new AI program won 14000 points after 192 games. Beside, our new AI won 2015 TCGA gold and 2015 ICGA gold.