Modeling purposeful adaptive behavior with the principle of maximum causal entropy

Abstract

Predicting human behavior from a small amount of training examples is a challenging machine learning problem. In this thesis, we introduce the principle of maximum causal entropy, a general technique for applying information theory to decision-theoretic, game-theoretic, and control settings where relevant information is sequentially revealed over time. This approach guarantees decision-theoretic performance by matching purposeful measures of behavior (Abbeel & Ng, 2004), and/or enforces game-theoretic rationality constraints (Aumann, 1974), while otherwise being as uncertain as possible, which minimizes worst-case predictive log-loss (Grunwald & Dawid, 2003). We derive probabilistic models for decision, control, and multi-player game settings using this approach. We then develop corresponding algorithms for efficient inference that include relaxations of the Bellman equation (Bellman, 1957), and simple learning algorithms based on convex optimization. We apply the models and algorithms to a number of behavior prediction tasks. Specifically, we present empirical evaluations of the approach in the domains of vehicle route preference modeling using over 100,000 miles of collected taxi driving data, pedestrian motion modeling from weeks of indoor movement data, and robust prediction of game play in stochastic multi-player games.