High-Level Human Intention Learning for Cooperative Decision-Making

Autonomous agents are increasingly popular in various practical domains, assisting humans in performing complex tasks. However, coordinating between agents and humans can be challenging, particularly when communication is limited or non-existent. This paper proposes a method for cooperative decision-making by enabling autonomous agents to infer high-level human intention through their behavioral data. Human is modeled as a sub-optimal reinforcement learning agent. A statistical learning method is developed for implicit probabilistic reasoning of human intentions by accounting for complex and unpredictable human behavior. The proposed method computes the exact likelihood and posterior of human intentions. The method is fully recursive and accounts for human priorities, making it applicable to various domains. The agents’ decision-making is achieved using a combination of the active learning approach and quantified human intentions, which enables effective coordination of tasks and prevents duplication of efforts. The proposed method allows agents to adapt their strategy in real time based on partial knowledge of human intentions. Our numerical experiments demonstrate the efficacy of our proposed method in intention inference and task coordination.