改进深度神经网络在爱恩斯坦棋中的应用研究

Enhanced application of deep neural networks in Einstein Chess research

爱恩斯坦棋作为一种附带随机性的完美信息博弈,其难度在于每次投掷骰子导致的结果不确定性,这增加了策略设计和局面的评估难度。针对爱恩斯坦棋的游戏规则,提出了一种改进的深度学习方法。对Alpha(go)Zero神经网络模型进行改进和设计,使其能精确地评估各种棋盘状态,生成有效的游戏策略。通过结合改进的残差神经网络和蒙特卡洛树搜索,提取棋局特征并进行局面评估,动态生成策略和进行决策。结合强化学习,以期望胜率为准则,通过自我对弈不断优化权重,改进策略生成效果。实验结果表明:改进的深度学习方法优于全国计算机博弈大赛冠军组算法,进一步验证了深度学习方法在爱恩斯坦棋随机性完美信息博弈中的有效性和可行性。

As a perfect information game with added randomness,Einstein Chess poses challenges due to the uncertainty brought by the dice rolls,which increases the difficulty of strategy design and position evaluation.This paper proposes a modified deep learning approach in response to the rules of Einstein Chess.First,improvements and designs are made to the Alpha(go)Zero neural network model,enabling it to accurately evaluate various board states and generate effective game strategies.Then,by combining an enhanced residual neural network and Monte Carlo tree search,chessboard features are extracted and position evaluation is conducted to dynamically generate strategies and make decisions.Reinforcement learning is employed,using expected win rate as the criterion,to continually optimize the weights through self-play and improve the effectiveness of strategy generation.Our experimental results indicate the improved deep learning approach outperforms the algorithms used by the champion team in the National Computer Games Tournament,further validating the effectiveness and feasibility of deep learning methods in the context of Einstein Chess as a random perfect information game.