Pablo Campos

In the Reinforcement Learning (RL) framework, two main elements are given by the policies, which define some agent’s way of behaving, and the rewards, which define the goals in a RL problem. Besides, Generative Adversarial Networks (GANs) [Goodfellow et al., 2014] are a class of generative models that have been regarded as the most interesting idea in the last years in the Machine Learning (ML) field, that are capable of produce realistic data in a wide range of domains. Nevertheless, the use of GANs in the context of RL remains a little explored research area. In this project we explore the idea of applying GANs in the context of RL in order to generate policies and rewards and also the idea of accelerating new learning by using the learned generative models. Using a simple RL environment, we show that it is completely possible to train GANs in order to be able to produce realistic policies and rewards that can be used independently to eventually accelerate new learning.