Exploring Theory of Mind for Agent Coordination in Constrained Operating Environments

Abstract This paper explores the application of the Theory of Mind (ToM) for achieving coordination between agents operating in constrained environments where the observation and communication capability of the agents is limited by the operating conditions. We use ToM for intention/action recognition and belief estimation in peer agents. We run extensive simulations with varied ToM capability settings in agents for two complex tasks: cooperative resource delivery and cooperative goal observation. The results indicate that in most cases, agents with higher-level ToM for intention and belief inferencing solved the tasks more effectively than the agents with lower levels of ToM. Higher-level ToM produced good results in time-constrained settings and promoted inter-agent cooperation for joint action planning even for low-observability conditions. However, for tasks with longer horizons, higher-level ToM was only effective when significant computation time was available to the agents. With smaller computational budgets, instead of modelling agents with ToM capability, using a look-ahead solution with a longer planning horizon rather than a shorter planning horizon with ToM leads to more effective solutions. These results are important as they are the first experiments demonstrating the effect of ToM for constrained environments, and they show the significance of agent modelling in developing a shared understanding of the environment and demonstrate that it is necessary to consider the trade-off between agent modelling vs. planning look-ahead.