FriSem

Julio Martinez (FYP), Student in the Department of Psychology, Stanford University

Title: Measuring and Modeling Physical Intrinsic Motivation

Abstract: People are interactive agents driven to explore the physical world and seek out situations with interesting physical dynamics. In this work we seek to formalize what the functional form of intrinsic motivation is that generates interestingness to ask whether this behavior is an information seeking behavior useful for improving physical prediction. To do this we measure interestingness judgements in humans under a variety of physical 3D simulated scenarios. We model intrinsic reward using simple scene features as well as features that depend on physical world model predictions. We then compare both scene features and world model based features to human judgements. We find that the single best predictors of intrinsic motivation are the information seeking features. This suggests that information seeking features are an important component of explaining physical intrinsic motivation. Secondly, that simple scene features are scene specific and do not generalize across a variety of different scenarios. Furthermore, all features are far below noise ceiling, which motivates the need to improve information seeking features as well as the formulation of additional scene features. Finally, we find that the experimental design probing interestingness results in measuring judgements that are generated for information seeking purposes as well as dynamics involving lots of activity such as high number of collisions suggesting that physical intrinsic motivation may be the consequence of both, and in many cases more of one than the other.