API Reference
Environment
OpenSpiel wrapper, board encoder, and action masking
| env.game_wrapper | OpenSpiel backgammon wrapper with perspective-relative board views. |
| env.encoder | Encode a BoardView into a (26, 2, 12) tensor for the neural network. |
| env.actions | Action space utilities for OpenSpiel backgammon. |
Model
ResNet policy-value network
| model.network | ResNet policy-value network for backgammon. |
Search
MCTS with chance-node handling
| search.mcts | AlphaZero-style MCTS with chance-node handling for backgammon. |
Training
Self-play loop, replay buffer, and coach
| train.self_play | Self-play game generation for training data. |
| train.replay_buffer | Circular replay buffer for training positions. |
| train.coach | Training orchestrator: self-play -> replay buffer -> SGD -> checkpoint. |
Evaluation
Checkpoint arena and GNUBG benchmark
| eval.arena | Checkpoint vs checkpoint evaluation arena. |
| eval.gnubg_harness | Automated money-game sessions against GNUBG via the gnubg-nn package. |
Protocol
Raccoon Game Protocol (RGP)
| protocol.rgp | Raccoon Game Protocol (RGP) — text-based engine communication. |