Imitation learning from compressed observation domain

Abstract

In general, when a novice attempts to imitate an expert’s behavior, the novice can not fully grasp the expert’s internal state. For example, when we learn the movement of a professional soccer player, we performe imitation learning while watching the video of the player, and we do not see the internal state such as the player’s joint or psychological state. In this case, the state information of the expert is compressed and transmitted to the novice, and the novice observes the compressed state information and performs imitation learning from this information. Therefore, the novice needs the ability of imitation learning from the expert state information projected onto the compressed observation domain. In the case of a person, this work can be done naturally through empathy with others, but it is a very challenging task for machines because machines require a lot of cross-domain pairs between the original state domain and the compressed observation domain. In this thesis, we propose an algorithm that can perform imitation learning from the compressed observation domain with only a very small amount of expert state sample, without the need of the cross-domain pairs between the state domain and the compressed observation domain.