This week, I focused on refining the architecture for the engine’s decision-making system by researching search-based algorithms and LLM orchestration. Instead of fully switching to a deep reinforcement learning approach as originally planned, I proposed a more practical and reliable baseline using Negamax search with alpha–beta pruning and fixed depth for efficient and deterministic move generation. I also planned an extension where machine learning is used only for position evaluation and move ordering, trained offline using Xiangqi datasets. This hybrid approach keeps the core engine stable and debuggable while still allowing performance improvements from learned heuristics.
I also looked into LLM-based explanation generation using pruned move sequences from the search as structured context for reasoning about multi-move strategies. Initial testing with vLLM and Gemini showed frequent hallucinations, including incorrect checkmate claims, so I plan to constrain the model to validated candidate moves and add rule-based verification to ensure explanations match the engine’s output. My progress is on schedule, and next week I will begin implementing the Negamax search, integrating it with the current move generator, and prototyping the evaluation interface for future ML integration.