1. How to design a multi-layer nested hierarchical set encoding architecture for complex turn-based strategy games to efficiently process variable-length state data.
2. LLM Semantic Clustering Pipeline Practice: Utilizing Large Language Models to transform skill texts into semantic embeddings and cluster them, achieving zero-annotation generalization for skills in new version updates.
3. Imperfect Information Game Modeling and Dual-Stream Historical Encoding: Combining the dual sequences of skill operations and deployments to enhance the AI's "memory" and "anticipation" capabilities.
4. Reward Engineering for Unique Game Mechanics: How to design targeted reward guidance strategies for exclusive game mechanics such as weather teams and resonance ritual, allowing a single architecture to cover hundreds of scenarios.
4. Industrial-Grade Practices from Training to Deployment: Covering CFR (Counterfactual Regret Minimization) lineup pool management, distributed training, and lightweight deployment. The system has already successfully supported billions of online matches.
