Behavioral Suite Analysis of Self-Supervised Learning in Atari

Somjit Nath, Rishav Rishav, Gopeshh Subbaraj, Derek Nowrouzezahrai, Samira Ebrahimi Kahou

Keywords: reinforcement learning, atari, representation learning, self-supervised learning

Abstract: Self-supervised learning (SSL) methods have shown promise in improving representation learning for reinforcement learning (RL), yet their effectiveness across diverse environments remains unclear. In this paper, we systematically evaluate various SSL methods on a broad set of Atari games, categorized according to the Behavioral Suite (bsuite) framework (Osband et al., 2020). By analyzing performance across these distinct categories, we identify which SSL approaches yield the most significant improvements in different types of Atari environments. Our empirical results, supported by comprehensive performance plots, provide insights into the strengths and limitations of SSL methods in deep RL, guiding future research directions in representation learning for reinforcement learning.

Submission Number: 15