Dynamic Decision-Making in Tank War Game: An LLM-Based AI Framework

This paper presents a novel framework for intelligent game AI in real-time 2D shooting games using large language models (LLMs), with Tank War as the testbed. Addressing the limitations of conventional rulebased systems and reinforcement learning approaches in dynamic combat scenarios. Leveraging the inherent interpretability and general-purpose nature of LLMs, our approach demonstrates superior cost-effectiveness and interpretability compared to reinforcement learning systems (requiring no specialized training) while outperforming rule-based architectures in adaptability and intelligence. To successfully and smoothly using LLMs in the game, we propose a hierarchical decision-making architecture that synergizes LLM-based strategic reasoning with real-time tactical execution. Specifically, In the strategic reasoning stage, a LLM is responsible for dispatching tanks to the most suitable destination in order to protect the base or attack the enemy or to wait for an opportunity. In the tactical execution stage, we construct a automatic attack module for attacking enemies in allowed distance, and a path planning module for generating a series of optimal moving actions to the destination and avoid obstacles according to the LLM's output. This work not only advances real-time game AI capabilities but also provides insights for applying LLMs in time-sensitive interactive systems.