Abstract:

The development of artificial intelligence technology has greatly promoted the transformation of the solving paradigm of intelligent game decision problems. From optimal solution, equilibrium solution to adaptive variable solution, how to build an intelligent game adaptive decision agent based on generative large model is full of challenges. The force distribution and multi-entity coordination in the game strong confrontation environment are the core issues in the study of troop deployment and operational coordination. Based on the methods of strategy reinforcement learning, strategy game tree search and strategy preference voting based on skill, ranking and preference meta-game model construction, a large model agent architecture is designed to meet the planning at generation time. The architecture can align the commander's intention with feasibility, applicability and extensibility, and can provide interpretable strategy recommendation for adaptive decision-making process. Key technical requirements are analyzed from the base model construction, goal-guided game reinforcement learning and open meta-game strategy learning. It is expected to provide reference for the cross-research of reinforcement learning model, game learning model and generative large language model.

Key words: adaptive, force allocation, multi-entity coordination, multi-agent reinforcement learning, meta-game, large language model, chain of thought