Deep Inverse Reinforcement Learning via Adversarial One-Class Classification

Abstract

Traditional inverse reinforcement learning (IRL) methods require a loop to find the optimal policy for each reward update (called an inner loop), resulting in very time-consuming reward estimation. In contrast, classification-based IRL methods, which have been studied recently, do not require an inner loop and estimate rewards quickly, although it is difficult to prepare an appropriate baseline corresponding to the expert trajectory. In this study, we introduced adversarial one-class classification into the classification-based IRL framework, and consequently developed a novel IRL method that requires only expert trajectories. We experimentally verified that the developed method can achieve the same performance as existing methods.

Tasks

Reproductions