Bluefir/escampe-dataset
收藏资源简介:
该数据集包含数百万个棋盘游戏Escampe的评估游戏状态,旨在训练深度神经网络(如BandDPER架构)以准确预测任何给定棋盘状态的启发式值。Escampe是一款确定性的、完全信息的6x6棋盘游戏,使用独角兽和圣骑士进行。数据集通过Java中的Alpha-Beta搜索引擎程序化生成,该引擎自我对弈Escampe游戏,采用随机开局和半随机走子来保证棋盘状态的高多样性。数据集中的每个位置都通过minimax引擎使用多种启发式配置(默认、SPSA、贝叶斯和单纯形)进行深度评估,以分配连续得分。数据字段包括表示白方和黑方圣骑士位置的64位整数位棋盘掩码、独角兽的索引、当前玩家必须开始移动的所需带子约束、轮到哪方移动的布尔标志,以及归一化到[-1.0, 1.0]范围的得分(从当前移动玩家的视角)。数据集生成过程强调多样性,通过混合最佳走子和随机走子(70%探索)来暴露子最优、混沌和终局位置,并将原始启发式得分钳位并归一化。使用数据时需注意得分字段严格基于移动玩家的视角,并建议将位棋盘转换为6x6空间张量以更好地处理空间约束。
This dataset contains millions of evaluated game states for the board game Escampe. It is designed to train Deep Neural Networks (such as the BandDPER architecture) to accurately predict the heuristic value of any given board state. Escampe is a deterministic, perfect-information 6x6 board game played with Unicorns and Paladins. The dataset was procedurally generated using an Alpha-Beta search engine in Java that plays games of Escampe against itself, taking random opening positions and semi-random moves to guarantee high diversity in board states. Every position in this dataset has been deeply evaluated by the minimax engine using various heuristic configurations (Default, SPSA, Bayes, and Simplex) to assign a continuous score. The data fields include bitboard masks for Whites and Blacks Paladins, indices for Unicorns, an integer for the required band constraint, a boolean for which player is to move, and a normalized score in the range [-1.0, 1.0] from the perspective of the current player to move. The creation process emphasizes diversity through a mix of best-moves and random moves (70% exploration), and scores are clamped and normalized. Considerations include inverting the score when white_to_move is false for absolute advantage models and translating bitboards into spatial tensors for neural network input.




