keisei
done形成 — “to give form to, to shape.” Deep reinforcement learning for Shogi: PPO agents against a Rust engine.
Three further architectures sit behind the same registry as baselines, so results have something to be measured against.
The engine handles board representation, legal move generation and rule enforcement — drops, promotion, repetition and impasse included — and exposes a vectorised environment to Python through PyO3. The primary target is a multi-head architecture: a spatial policy over Shogi’s move semantics, a win/draw/loss value head, and score prediction as an auxiliary task. The registry also carries the ablations, and the harness is as much the deliverable as the agent. The SE-ResNet trunk, W/D/L value head and score head are adapted from KataGo (David Wu, 2019); the self-play approach originates with AlphaZero.
uv pip install -e ".[dev]"
uv run keisei-train keisei-katago.toml \
--epochs 100 --steps-per-epoch 256