DLManaka

GPU-first residual-network shogi engine with Gumbel AlphaZero MCTS. Training is PyTorch; inference/search integration is Rust and currently uses CUDA/ONNX Runtime, with Apple silicon as a longer-term target.

Read by topic

Design

• Architecture — current network and runtime structure
• Search design — implemented Gumbel MCTS and batching invariants
• Inference — inference execution path
• Network history — placeholder; no separate history page yet

Training and self-play

• Training — trainer-specific operation
• Teacher data — shared corpus used by both engine lines
• Self-play roadmap — implemented pieces and remaining candidates
• Generation loop — placeholder; loop details are currently covered by the roadmap/improvement records

Measurement

• DLManaka benchmark notes — engine-specific interpretation
• Common benchmark — canonical current speed figures
• Playing-strength measurement — external anchors and SPRT

Historical diagnosis / current priorities

• First self-play plateau — historical record; the decisive finding was that the original teacher search was too shallow
• Value-head decay — historical training diagnosis; value_weight=2 was not adopted
• Gen11 improvement plan — latest documented priority checkpoint

Placeholder pages exist to keep the DLManaka/Manaka navigation symmetric; “placeholder” means the topic has no dedicated write-up yet, not that the URL is missing.