Human social interaction depends on the ability to understand other people's actions in terms of the mental states that produce behavior. Much like visual perception, action understanding proceeds unconsciously and effortlessly but is the result of sophisticated computations designed to solve a highly under-constrained inverse problem. While vision is a kind of "inverse graphics", action understanding is a kind of "inverse planning". The goal is to recover the goals and beliefs that lead an agent to act in some observed way. The core assumption is the principle of rationality: a rational agent tends to choose actions that satisfy its goals most efficiently given its model of the environment. Observing an agents actions, we can then work backwards to infer its goals or its environment model (or perhaps both). Evidence from behavioral experiments suggests that action understanding in adults and even preverbal infants is qualitatively consistent with this "inverse planning" view. Our aim here is to develop a mathematically precise of this account, to assess its quantitative predictive power for human judgments about the goals of actions, and if possible to distinguish it from simpler heuristic approaches. Our models take a Bayesian approach to inverse probabilistic planning in partially observable Markov decision problems (POMDPs). This inverse planning framework includes many specific models that differ in their representations of agents mental states and actions -- the priors needed to solve inverse-planning problems. Our experiments were designed both to test the general framework and to probe the nature of peoples representations of agents goal structures. One set of experiments examined subjects' online and retrospective goal inferences from incomplete trajectories, while another set of experiments investigated subjects' predictions of agents' future actions, given observations of previous actions. We identify a class of inverse planning model that correlate very highly with people's judgments, and that crucially allow goals to have complex temporal dynamics and componential structure. |