Human social interaction depends on our ability to understand and predict other people's actions in terms of the psychological states that produce
behavior: chiefly, beliefs and desires. Much like visual perception, action understanding proceeds unconsciously and effortlessly but is the result of sophisticated computations. These computations effectively solve an ill-posed inverse problem, working backwards from sparse data to rich representations of the underlying hidden causes. I will discuss some of the key phenomena of action understanding in human adults and infants, and then present a computational approach to modeling these inferences.
While vision is often said to be a kind of "inverse graphics", inverting the physical processes that form images from scenes in order to infer the scene structure most likely to have generated an observed image, action understanding is a kind of "inverse planning" or "inverse reinforcement learning". Observing the actions of an agent, we can work backwards to infer the agent's goals or its environment model (or perhaps both), by inverting a model of the agent's planning processes. This inversion typically hinges on some version of the "principle of rationality", the assumption that a rational agent tends to choose actions that satisfy its goals or desires most efficiently, or that maximize its expected reward, given its model or beliefs about the environment. We will formalize inferences based on the principle of rationality using the framework of Markov decision processes (MDPs) introduced in the preceding lectures, and show how human action understanding can be modeled an inversion of an appropriately defined goal-based MDPs. These models yield new insights into how people represent the structure of other agents' goals. They also make suprisingly accurate quantitative predictions about people's goal judgments, which can be contrasted with simpler accounts that do not involve solving a full inverse-planning problem.
Copyright. All Rights Reserved.