Random Projection Based Representations for Learning Policies in Deterministic Atari Games

Tudor Berariu, Florin Gogianu, Adina Magda Florea

Abstract: Recent advances in sample efficient reinforcement learning algorithms in quasi-deterministic environments highlight the requirement for computationally inexpensive visual representations. Here we investigate non-parametric dimensionality reduction techniques based on random linear transformations and we provide empirical evidence on the importance of high-variance projections using sparse random matrices in the context of episodic controllers learning deterministic policies. We also propose a novel Maximum-Variance Random Projection and improve on the performance of the original Model-Free Episodic Control results with respect to both sample efficiency and final average score.