Unsupervised inverse reinforcement learning with noisy data

Abstract

In this paper we propose an approach for unsupervised Inverse Reinforcement Learning (IRL) with noisy data using a hidden variable Markov Decision Processes (hMDP) representation. hMDP accounts for observation uncertainty by using a hidden state variable. We develop a nonparametric Bayesian IRL technique for hMDP based on Dirichlet Processes mixture model. We provide an efficient Markov Chain Monte Carlo based sampling algorithm whereby one can automatically cluster noisy data into different behaviors, and estimate the underlying reward parameters per cluster. We demonstrate our approach for unsupervised learning, and prediction and classification of agent behaviors in a simulated surveillance scenario.