Abstract
With the imminent emergence of peer adversaries, the DoD must elevate its design and use of simulation-based training exercises so that they better represent the complexities of anticipated future conflict. It is widely understood that future conflicts will be multi-domain, and will require simulating hundreds of disparate actions and behaviors. To model these behaviors in the absence of human role-players, we rely on computer-generated forces (CGFs) – long a staple in constructive and virtual simulation environments. CGFs can be used to simulate teammates, OPFOR, and pattern-of-life behaviors that all contribute to realistic training scenarios.
While there is no shortage of techniques, architectures, and even graphical authoring tools for building and integrating these CGF agents, the time and expertise required to develop those behaviors is often cost-prohibitive. Large Language Models (LLMs) provide a compelling alternative to manual CGF behavior development and authoring. Optimally, a robust LLM – tuned with knowledge about the target domain - could generate intelligent agents directly from prompt, saving hours of manual knowledge engineering and behavior development. Unfortunately, such an approach ignores essential steps in the process, such as validation, debug, and integration with a target simulation.
We introduce a novel architecture for generating intelligent agents for CGFs that situates custom-tuned LLMs into an agentic framework capable of composing, refining, evaluating, and integrating new behaviors into a specified simulation environment from prompt. We discuss the process involved iterating on the design of this architecture, the limitations discovered in targeting specific behavior representations. As simulations can vary drastically in how they represent and control CGF behaviors, this architecture includes an approach for automatic discovery of actions and state-retrieval functions, allowing the generated agents to operate seamlessly within various simulation environments.