How Action Understanding can be Rational, Bayesian and Tractable

An important aspect of human sociality is our ability to understand the actions of others as being goal-directed. Recently, the now classic rational approach to explaining this ability has been given a formal incarnation in the Bayesian Inverse Planning (BIP) model of Baker, Saxe, and Tenenbaum (2009). The BIP model enjoys considerable empirical support when tested on ‘toy domains’. Yet, like many Bayesian models of cognition, it faces the charge of computational intractability: i.e., the computations that the model postulates may be too resource demanding for the model to be scalable to domains of realworld complexity. In this paper, we investigate ways in which the BIP model can possibly parry the charge. We will show that there are specific conditions under which the computations postulated by the model are tractable, despite the model being rational and Bayesian.