École Nationale Supérieure de
Techniques Avancées

Abstract

The purpose of this paper is to describe the implementation of a 3-part architecture described in the paper World Models allowing the learning of a simplified environment on classical reinforcement learning environments. This is made possible by the combined training of an encoder summarizing the perceived information of the environment and a predictor giving some insights on the next observation, given a state and an action. It is thus possible to train a small controller based on this summarized world model to achieve better performances than those obtained with classical algorithms. Moreover, it is also possible to train the controller without interactions with the real environment but inside the dream generated by the predictor before using the controller in the real environment