Proceedings of the Annual Meeting of the Cognitive Science Society

A major challenge for research in Artificial Intelligence (AI)is to develop systems that can infer humans’ goals and beliefswhen observing their behavior alone (i.e., systems that haveTheory of Mind, ToM). In this research we use a theoretically-grounded, pre-existent cognitive model to demonstrate the de-velopment of ToM from observation of other agents’ behavior.The cognitive model relies on Instance-Based Learning The-ory (IBLT) of experiential decision making, that distinguishesit from previous models that are hand-crafted for particular set-tings, complex, or unable to explain a cognitive developmentof ToM. An IBL model was designed to be an observer ofagents’ navigation in gridworld environments and was queriedafterwards to predict the actions of new agents in new (notexperienced before) gridworlds. The IBL observer can inferand predict potential behaviors from just a few samples ofagents’ past behavior of random and goal-directed reinforce-ment learning agents. Furthermore the IBL observer is able toinfer the agent’s false belief and pass a classic ToM test com-monly used in humans. We discuss the advantages of usingIBLT to develop models of ToM, and the potential to predicthuman ToM.