Problem decomposition for behavioural cloning

Abstract

In behavioural cloning of the human operator’s skill, a controller is usually induced directly as a classifier from system’s states into actions. Experience shows that this often results in brittle controllers. In this paper we explore a decomposition of the cloning problem into two learning problems: the learning of operator’s control trajectories and the learning of the system’s dynamics separately. We analyse advantages of such indirect controllers. We give characterization of the learner’s error that is plausible explanation of why this decomposition approach has empirically proved to be usually superior to direct cloning.