Towards Spontaneous Cooperation in Multi-Agent Reinforcement Learning using Explicit Goal Recognition

Kirin Godhwani, Adam S R Parker, Matthew E. Taylor, William Yeoh, Reuth Mirsky

Keywords: Theory of Mind, Goal Recognition, Multi-Agent Reinforcement Learning, Ad-Hoc Teamwork

TL;DR: We propose a Theory of Mind-inspired approach that enables artificial agents to spontaneously cooperate by inferring and supporting teammates’ hidden goals without communication.

Abstract: Spontaneous cooperation --— the ability to assist others without explicit instruction or coordination --— is a hallmark of intelligent social behavior observed in humans and other animals. However, most Multi-Agent Reinforcement Learning (MARL) approaches lack mechanisms for intuitive, goal-directed helping due to limited modeling of other agents' internal states. In this paper, we explore a Theory of Mind (ToM)-inspired approach to address this gap, enabling artificial agents to infer and support the hidden goals of their teammates. Building on the Hidden Goal Markov Decision Process (HGMDP) framework, we introduce a baseline evaluation in a simplified collaborative domain in which an assistant agent must infer whether a leader agent is hungry or thirsty and deliver the appropriate item without direct communication. This preliminary system demonstrates how basic goal inference can enable spontaneous, context-sensitive cooperation. These findings lay the groundwork for future development of MARL agents capable of adaptive, intuitive assistance in more complex environments.

Submission Number: 6