Efficient Reinforcement Learning via Self-supervised learning and Model-based methods

Abstract

In recent years, Reinforcement Learning systems have led to remarkable accomplishments. Applications of Reinforcement Learning are vast and range from optimizing the energy consumption of data-centers and more efficient chip design to intelligent autonomous robots. However, Reinforcement Learning algorithms still have their limitations, most notably their lack of data efficiency. Different solutions were proposed that aim to mitigate this problem and in principle, there are two main approaches: Self-supervised learning and Model-based methods. This paper discusses the applications and theoretical foundations of Self-supervised learning and Model-based methods in the Reinforcement Learning context and investigates their individual strengths and weaknesses as well as their commonalities and intersections. The combination of both methods might lead to even better results and therefore, this paper will conclude by proposing potential ways to unify them.