darkforestGo is an early deep-reinforcement-learning Go engine that combined a convolutional policy/value network with Monte Carlo Tree Search (MCTS) to play the full 19×19 game at a strong amateur level. The system couples fast GPU policy inference with CPU or GPU-assisted tree search so priors from the network guide exploration while search refines local tactics. Training pipelines mix supervised learning from human professional games and self-play fine-tuning, allowing the model to learn opening patterns and endgame tactics beyond simple pattern libraries. The codebase includes tools for parsing classic Go formats, generating training examples, and evaluating models on standard test suites and online servers. A KGS/online client and match runner make it practical to stage controlled tournaments or continuous rating evaluation. Although later projects (like ELF OpenGo) surpassed it in strength, darkforestGo remains a historically important, clean reference for neural-MCTS Go systems.
Features
- Residual CNN policy/value network integrated with MCTS
- Supervised pretraining on human games plus self-play fine-tuning
- Data pipelines for feature extraction, example generation, and evaluation
- Match runner and online client for KGS or scripted tournaments
- Configurable search parameters and time controls for reproducible tests
- Tools to export, analyze, and compare model checkpoints
Project Samples
