AlphaZero-Style Chinese Chess AI
A Chinese chess system combining a policy–value network with enhanced Monte Carlo Tree Search, self-play data generation and model lifecycle management.
2026Inside the system
A closer look at the operating problem, implementation approach and concrete scope of delivery.
The challenge
Xiangqi has a large search space. The system needed complete rules and legal moves while connecting neural evaluation, tree search, experience reuse and external-engine comparison into a sustainable training loop.
System approach
MCTS applies UCB1-Tuned, RAVE, progressive widening and Zobrist transposition tables. Self-play creates policy/value data, model management handles checkpoints and promotion, and Pikafish is integrated through UCCI for comparison.
Delivery scope
Covered Xiangqi rules and state representation, policy–value integration, MCTS optimisation, self-play data, model management, opening books and cross-platform engine discovery.
Selected system views
Supporting interfaces are curated and anonymised. Open an image to view it at full size.
Algorithm, search and training system overview
The poster presents the real technical flow across board-state encoding, policy–value inference, enhanced MCTS, self-play, model management and external-engine comparison.
AlphaZero-style Xiangqi AI workflow
Board states and legal moves enter policy–value inference and enhanced MCTS; search selects moves, self-play data returns to training, and promotion plus Pikafish comparison completes evaluation.
Core deliverables
- 01
Xiangqi rules and state encoding
- 02
Policy–value inference
- 03
Enhanced MCTS search
- 04
Self-play and model lifecycle