[P] gumbel-mcts, a high-performance Gumbel MCTS implementation

Developer olivkoch has released gumbel-mcts, a new high-performance implementation of Monte Carlo Tree Search (MCTS) algorithms built in Python with Numba for speed. The project fills a gap in the open-source ecosystem by providing a validated, efficient library for developers building game-playing AI or self-play environments. The core PUCT (Predictor + Upper Confidence Bound for Trees) implementation delivers identical policy results to a golden standard baseline but runs 2 to 15 times faster, a significant performance boost for training and inference.

The library also implements Gumbel MCTS in both dense and sparse variants. The sparse version is specifically designed for games with large action spaces, such as chess, where traditional methods struggle. Gumbel MCTS is noted for making much better use of low simulation budgets compared to PUCT, allowing for more effective decision-making with fewer computational resources. The developer used AI coding assistants during development but emphasized significant manual work to validate all results, ensuring reliability for the community.