Computational Theory of Mind for Human-Agent Coordination.

Abstract

In everyday life, people often depend on their theory of mind, i.e., their ability to reason about unobservable mental content of others to understand, explain, and predict their behaviour. Many agent-based models have been designed to develop computational theory of mind and analyze its effectiveness in various tasks and settings. However, most existing models are not generic (e.g., only applied in a given setting), not feasible (e.g., require too much information to be processed), or not human-inspired (e.g., do not capture the behavioral heuristics of humans). This hinders their applicability in many settings. Accordingly, we propose a new computational theory of mind, which captures the human decision heuristics of reasoning by abstracting individual beliefs about others. We specifically study computational affinity and show how it can be used in tandem with theory of mind reasoning when designing agent models for human-agent negotiation. We perform two-agent simulations to analyze the role of affinity in getting to agreements when there is a bound on the time to be spent for negotiating. Our results suggest that modeling affinity can ease the negotiation process by decreasing the number of rounds needed for an agreement as well as yield a higher benefit for agents with theory of mind reasoning.