Starcraft II learning environment for Large Language Models

PDF file display entitled LLM-PYSC2: Learning Environment in Starcraft II for large language models, written by Zongyuan Li and 14 other authors

PDF HTML (experimental) view

a summary:This tremendous potential through LLMS models (LLMS) has shown problems in making smart decisions, with unprecedented capabilities offered through various applications ranging from AI games to complex strategic planning frameworks. However, the Starcraft II platform, which has been widely adopted to verify the health of decision -making algorithms in the past decade, has not provided great support for this emerging field. To address problems that LLMS cannot interact with hundreds of procedures from the PYSC2 back interface and the lack of original support for multi-factor cooperation (MA), we suggest the LLM-PYSC2 environment. This is the first environment that provides LLMS in full PYSC2 work space with sufficient multimedia information and the game Wiki. Through the inclusive query structure, the environment interacts efficiently with LLMS, which maintains a fixed continuity time regardless of the size of the population. In experiments, we evaluated the performance of LLMS decision -making in each of the scenarios of the total decision and the micro -process, with the traditional Starcraft II challenge tasks (SMAC) and a series of new proposal. The results indicate that LLMS has the ability to achieve victories in complex scenarios but cannot constantly generate correct decisions, especially in the recovered PYSC2 work space and MA settings. Without instructions related to the task, pre -trained models suffer from issues such as hallucinations and ineffective cooperation. The results we have reached indicate that Starcraft II still challenges in the era of large models, revealing that there is a lot to do to develop an advanced LLM decisions system, and the proposed LLM-PYSC2 environment will support future development solutions to LLM decision-making solutions.