Cooperative Inverse Reinforcement Learning - Cooperation and learning in an asymmetric information setting with a suboptimal teacher

Abstract

There exists many different scenarios where an artificial intelligence (AI) may have to learn from a human. One such scenario is when they both have to cooperate but only the human knows what the goal is. This is the study of cooperative inverse reinforcement learning (CIRL). The purpose of this report is to analyze CIRL when the human is not behaving fully optimally and may make mistakes. The effect of different behaviours by the human is investigated and two frameworks are developed, one for when there is a finite set of possible goals and one for the general case where the set of possible goals is infinite. Two benchmark problems are designed to compare the learning performance. The experiments show that the AI learns, but also that the humans behaviour has a large affect on learning. Also highlighted by the experiments, is the difficulty of differentiating between the actual goal and other possible goals that are similar in some aspects.