MinervaZero: A Self-Trained Chess Engine for Atomic Chess - An Empirical Study of Neural Network Size, Transfer Learning, and Rule Acquisition in Atomic Chess

Abstract

Chess engines emerged shortly after the invention of modern computing. Traditional engines relied on handcrafted heuristics and evaluation functions derived from human expertise. In 2017, AlphaZero defeated Stockfish, the strongest chess engine at the time, using reinforcement learning and self-play rather than manually designed evaluation strategies. This thesis presents MinervaZero, an AlphaZero-style engine developed for the Atomic Chess variant. It uses Monte Carlo Tree Search (MCTS) guided by a neural network to evaluate board positions and select promising moves. The primary objective of the project is to outperform a baseline MCTS rollout engine. Furthermore, the thesis examines the effect of transfer learning from standard chess, the impact of neural network size, and how well the network internalizes the rules of Atomic Chess. The engine was implemented in C++ and designed with extensive parallelization to utilize the available hardware resources on a compute cluster node. Multiple models were trained, and then evaluated through both head-to-head matches and controlled experiments across thousands of training iterations. The most extensively trained model completed 5000 training iterations, played 2.56 million matches, and trained on 81.9 million positions extracted from those games. The results show that MinervaZero substantially outperforms the baseline MCTS rollout engine after sufficient training. Transfer learning from standard chess provides no performance advantage. Larger networks achieve stronger performance during early training stages. Furthermore, the neural network quickly reduces illegal move probabilities when training, indicating strong rule acquisition.