A Self-Trained Chess Engine for Crazyhouse Chess

Abstract

This thesis investigates the development of a self-trained chess engine for the Crazy house chess variant, inspired by the AlphaZero framework. Crazyhouse extends classical chess with a drop mechanic, where captured pieces can be collected and reintroduced onto the board, fundamentally altering strategic evaluation and significantly increasing the branching factor compared to classical chess. The absence of established opening theory makes Crazyhouse a particularly suitable domain for self-learning approaches, as chess engines cannot rely on accumulated human knowledge and must instead discover effective strategies entirely through self-play. The engine was implemented using a C-based game engine for efficient move generation and rule enforcement, combined with a Python-based Monte Carlo Tree Search (MCTS) algorithm guided by a neural network. Four distinct convolutional neural network architectures were designed and compared, a shallow non-residual baseline, a shallow residual network, a deep residual network and a wide shallow residual network. All architecture were trained through an iterative self-play process and evaluated using a round-robin tournament. The results showed that residual architectures consistently outperformed the non residual baseline, confirming that skip connections provide significant benefits for gradient flow and learning stability in the context of Crazyhouse position evaluation. Among the residual architectures, the shallow residual network, using 128 channels and 6 residual blocks, demonstrated the strongest learning trajectory, having a more upwards going regression trend compared to the other neural network models. The deep residual architecture underperformed relative to its theoretical capacity, which suggests that the project’s computational and time constraints resulted in insufficient training data. While all trained architectures demonstrated clear strategic intent beyond random play, none were able to win consistently against Stockfish at any tested skill level, indicating that the scale of training required to reach competitive strength exceeds what was achievable within a single academic semester.