Military reinforcement learning with large language model–based agents: a case of weapon selection

Autonomous agents have been the center of discussion for the future concept of operations. Reinforcement learning (RL) is the core machine learning area for developing those intelligent agents, particularly in complex and dynamic environments such as battlefields and afflicted areas. This study proposes the large language model (LLM)-based RL system to utilize the power of LLMs for military RL applications. Users can use the system through prompts, and three different types of prompting are tested with the weapon selection scenario. The proposed system helps and guides users not only in building an RL agent (optimal policy) quickly but also in providing related theories and other information. In comparison to the human-designed RL system, the proposed system also had some limitations, such as reproducibility and reliability. This study discussed and suggested some remedies for the limitations.