Incorporation of Military Doctrines and Objectives into an AI Agent Via Natural Language and Reward in Reinforcement Learning
Michael Möbius, Daniel Kallfass, Matthias Flock, Thomas Doll, Dietmar Kunde · 2023
This paper emphasizes the integration of sound tactical behavior in the generation of realistic military simulations, which includes the definition of combat tactics, doctrine, rules of engagement, and concepts of operations. Recent advances in reinforcement learning (RL) enable RL agents to generate a wide range of tactical actions. A multi-agent ground combat scenario is used in this paper to demonstrate how a machine learning (ML) application generates strategies and issues commands while following a given objective. Natural language is used to issue doctrines and objectives to improve communication between the human advisor and the ML agent. This allows us to embed objectives and existing doctrines into the reasoning of an artificial intelligence (AI). The research demonstrates the successful integration of natural language to enable an agent to achieve different objectives. This groundwork will enhance RL agents' ability in the future to uphold the doctrines and rules of military operations.