Frontiers in Inverse Reinforcement Learning

Abstract

How do we teach machines to do something that we can perform reasonably well, but cannot easily express as a utility maximization problem? Can machines learn underlying utility of a domain from many human demonstrations? The goal of the field of Inverse Reinforcement Learning (IRL) is to infer the crux of the goal of a domain from expert (human) demonstrations. This thesis categorically surveys the current IRL literature with a formal introduciton and motivation for the problem. We discuss the central challenges of the domain and expound upon how different algorithms deal with the challenges. We propose an reformulation of the IRL problem by including ranked set of trajectories of different levels of expert capability and discuss how that might lead towards a new set of algorithms in the field, motivated by some very recently developed approaches. We conclude with discussing some broad advances in the research area and possibilities for further extension.