结合先验知识与蒙特卡罗模拟的麻将博弈研究

Research on mahjong game based on prior knowledge and Monte Carlo simulation

针对内陆麻将缺乏统一平台和大量牌谱数据,难以设计出基于监督学习的博弈算法的问题,本文设计了一系列将规则、经验与蒙特卡罗方法相结合的博弈算法。首先,分别针对麻将博弈的弃牌模块、听牌模块、吃牌模块提出了弃牌优先级、听牌有效数、吃牌优先级的方法,完善了麻将AI的知识体系,设计了基础版博弈算法Fanfou_ba和优化版博弈算法Fanfou_op;其次,提出了利用蒙特卡罗方法模拟听牌对手手牌来降低己方点炮概率的提升版博弈算法Fanfou_mc;最后,将3种博弈算法进行对比实验。实验结果显示Fanfou_op相比Fanfou_ba胜率提高了9.76%,Fanfou_mc相比Fanfou_op胜率提高了0.13%且点炮率降低了0.47%,表明本文所提出的改进策略是可行并有效的。

In view of the difficulty in designing game algorithms based on supervised learning due to the shortage of a unified platform and a large amount of card score data for inland mahjong,,this paper designs a series of game algorithms that combine rules,experience and the Monte Carlo method for inland mahjong game.Firstly,the fold priority,effective number of draws and the eating priority are proposed for the discard module,draw module,and card eating module of the mahjong game,respectively.The mahjong AI knowledge system is improved,and the basic game algorithm Fanfou_ba and the optimized game algorithm Fanfou_op are designed.Secondly,the game algorithm Fanfou_op is proposed that reduces the probability of firing a shot by using the Monte Carlo method to simulate the waiting opponent's hand.Finally,comparative experiments are conducted on these three kinds of game algorithms.The experimental results show that compared with Fanfou_ba,the Fanfou_op algorithm improves the win rate by 9.76%,and that compared with the Fanfou_op algorithm,the Fanfou_mc algorithm enhances win rate by 0.13%and reduces the shot rate by 0.47%,which proves that the improvement strategy proposed is feasible and effective.