Scaling Reinforcement Learning Through Feudal Multi-agent Hierarchy
Information
Date
September 2022
Summary
Militaries conduct wargames for training, planning, and research purposes. Artificial intelligence (AI) can improve military wargaming by reducing costs, speeding up the decision-making process, and offering new insights. Previous researchers explored using reinforcement learning (RL) for wargaming based on the successful use of RL for other human competitive games. While previous research has demonstrated that an RL agent can generate combat behavior, those experiments have been limited to small-scale wargames. This thesis investigates the feasibility and acceptability of scaling hierarchical reinforcement learning (HRL) to support integrating AI into large military wargames. Additionally, this thesis also investigates potential complications that arise when replacing the opposing force with an intelligent agent by exploring the ways in which an intelligent agent can cause a wargame to fail. The resources required to train a feudal multi-agent hierarchy (FMH) and a standard RL agent and their effectiveness are compared in increasingly complicated wargames. While FMH fails to demonstrate the performance required for large wargames, it offers insight for future HRL research. Finally, the Department of Defense (DoD) verification, validation, and accreditation process is proposed as a method to ensure that any future AI application applied to wargames are suitable.