Deep learning and computer chess (part 2)

• Main Author: Lee, Zachary Varella Zheyu
• Other Authors: He Ying
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering
• Online Access: https://hdl.handle.net/10356/166646
• Institution: Nanyang Technological University

The dominant approach to computer chess has typically been through the use of Minimax-based chess engines. In recent years, Monte Carlo Tree Search (MCTS) game engines have seen success, with the advent of AlphaZero and Leela Chess Zero. However, there is still much to explore regarding the use of MCTS in the domain of chess. This paper evaluates the efficacy of an MCTS-based engine in the area of chess. On top of the base MCTS engine, several enhancements were proposed and implemented, including early playout termination, progressive bias, progressive unpruning, decisive moves, epsilon-greedy search, score bounded Monte-Carlo tree search, and root parallelization. Each enhancement was implemented in stages, and the performance of the enhancement was measured by comparing it to the model from the previous stage. It was determined that early playout termination, progressive unpruning, score bounded search, and root parallelization were effective in improving the playing strength of the engine. However, decisive moves and epsilon-greedy search negatively impacted the engine’s performance. From the results, it appears that it is possible to adapt an MCTS model to the realm of chess through the aid of several enhancements such that can compete with the traditional Minimax approach, with much room for improvement available.