MCTS search
PUCT with transposition table, parallel playouts, and heuristic or ONNX policy/value heads.
Gofer is a Go engine with Chinese rules, MCTS search, neural net evaluation, GTP for GUIs, and a self-play training loop. Built to be understood and extended.
Features
PUCT with transposition table, parallel playouts, and heuristic or ONNX policy/value heads.
Plug into Sabaki, Lizzie, or any GTP client. Think-time or fixed playout budgets.
Play, analyze, watch engine vs engine, and export SGF — no GUI required.
Self-play samples, PyTorch bootstrap trainer, arena gating, and ONNX export.
Get started
Clone the repo, build the binary, and play a game in your terminal or hook it up to a GUI.
Defaults are tuned for 9×9. ONNX inference runs in-process — no sidecar required.
# build make build # play in the terminal bin/gofer -play -size 9 -color b # GTP for Sabaki / Lizzie bin/gofer -gtp -size 9 -think-time 5s