Hybrid Ensemble Neural Networks for Chess Position Evaluation

Evaluation of chess positions represents one of the oldest challenges in artificial intelligence, a field that has captivated mankind for replicating human-like strategic thinking. Traditionally, this domain has been dominated by rule-based engines that operate on a complex predefined evaluation function to assess a given chess position. In contrast, deep NNs learn from vast datasets of games, allowing them to evaluate the chess position with a level of sophistication that is often comparable to, and sometimes better than, traditional approaches. In this work, we present a hybrid ensemble neural network model to evaluate a given chess position. We combine two commonly used architectures: feedforward neural networks (FNNs) and convolutional neural networks (CNNs), leveraging their respective strengths. The model has been trained on a dataset of around 80 million unique chess positions evaluated using the current best chess engine, Stockfish. To evaluate the effectiveness of our model in assessing chess positions, we use it as the baseline evaluation function within a minimax algorithm enhanced with alpha-beta pruning. The top predicted moves are then compared to those generated by Stockfish.