Evaluation of inverse reinforcement learning

Abstract

Inverse Reinforcement Learning (IRL) is a technique that is concerned with learning the intrinsic reward function of an expert by observing them perform a task. Many different methods exist from linear programming to deep learning for efficiently computing the reward values. However, evaluation of the performance of these methods is often minimal. Typically there is an evaluation of if a goal is reached. This is not enough to establish that the reward function of an expert is successfully captured. There is no standard for evaluating the outcome of IRL. This thesis proposes using a method for measuring the accuracy of any IRL method. This measure requires the construction of two graphs. The first is a graph that is constructed from the observation trajectories that are being used with IRL. The second is a graph created by an agent implementing the results of the IRL in the environment. These can then be compared using graph edit distance, which will give a discrete measure for evaluating the accuracy of any IRL method.