October 28, 2019

Modeling Theory of Mind for Competition, Cooperation and Communication

Tao Gao, University of California, Los Angeles

Video not available for this presentation.

Theory of mind (ToM) refers to the attribution of an agent’s motion to its mental states, including belief, desire and intention. Modeling ToM is built upon two principles. First, the “rationality principle” (utility theory), assuming that an agent takes actions to maximize its utility. Second, the Bayes’ theorem, solving ToM by maximizing the posterior of mental states conditioning on the observed actions. A model of ToM is a model of social commonsense that can explain a wide range of human interactions. I will start from a zero-sum chasing game, in which a human-controlled prey detects and avoids a computer-controlled predator. Both human and modeling results show that perceived chasing is severely disrupted when the predator’s actions violate the rationality principle, enabling the predator to stalk the prey stealthily. ToM becomes more prominent in cooperative tasks. Multi-agent ToM is challenging due to its recursive nature: I infer your inference of my inference of you. Here I advocate an “Imagined We”(IW) approach that avoids this recursion trap. “We” is defined as a super-agent that can rationality and centrally control all agents as body parts to maximize the joint utility. However, this “We” agent does not exist in reality. Instead, each agent actively imagines what We believes, and follows “what We wants me to do” voluntarily. IW predicts spontaneous “role assignment” and “goal commitment”. Furthermore, IW also serves as a cognitive infrastructure of human communication. I will show that the combination of cooperative logic, utility theory, and Bayes’ theorem strongly constrain the interpretation of even highly ambiguous signals, enabling humans to communicate so much by expressing so little.