CAS and Wang Jun's Team Release Large Model Playing StarCraft with Impressive Performance
-
Facing the immense challenge of StarCraft II, the team developed TextStarCraftII—a brand-new interactive environment. TextStarCraftII is based on the python-sc2 framework, mapping in-game state information and action spaces into a textual space. Macro-level strategic actions are transformed into specific semantic actions that the LLMAgent can understand and execute, while micro-level operations are handled by a set of fixed rule-based methods. The LLMagent is capable of competing against the game's built-in AI on entirely new battlefields.
Paper address: https://arxiv.org/pdf/2312.11865.pdf
Project address: https://github.com/histmeisah/Large-Language-Models-play-StarCraftII
On the battlefield of StarCraft II, making effective decisions requires timely processing of large amounts of complex information, conducting reasonable strategic analysis and long-term planning, and ultimately formulating macro strategic decisions. The team innovatively proposed the 'Chain of Summarization' method.
This method improves LLM's understanding and decision-making capabilities in complex environments through single-frame summarization and multi-frame summarization. To validate the effectiveness of the Chain of Summarization method, the team selected GPT-3.5-turbo-16k as the LLM. The results show that Chain of Summarization not only increased the interaction speed between LLM and the game client by ten times but also significantly enhanced the model's understanding of game scenarios and decision-making capabilities.
The team has carefully designed a sophisticated prompt system, including game state summaries, state analysis, strategic suggestions, and final decisions. The model can comprehensively understand the current game situation, analyze the strategies of both allies and enemies, and provide strategically profound suggestions, ultimately making multi-step reasonable decisions. This significantly improves the LLM's real-time decision-making and long-term planning capabilities, while also greatly enhancing the interpretability of decisions.