Understanding representation learning for deep reinforcement learning

Abstract

Representation learning is essential to practical success of reinforcement learning. Through a state representation, an agent can describe its environment to efficiently explore the state space, generalize to new states and perform credit assignment from delayed feedback. These representations may be state abstractions, hand-engineered or fixed features or implied by a neural network. In this thesis, we investigate several desirable theoretical properties of state representations and, using this categorization, design novel principled RL algorithms aiming at learning these state representations at scale through deep learning. First, we consider state abstractions induced by behavioral metrics and their generalization properties. We show that supporting the continuity of the value function is central to generalization in reinforcement learning. Together with this formalization, we provide an empirical evaluation comparing various metrics and demonstrating the importance of the choice of a neighborhood in RL algorithms. Then, we draw on statistical learning theory to characterize what it means for arbitrary state features to generalize in RL. We introduce a new notion called effective dimension of a representation that drives the generalization to unseen states and demonstrate its usefulness for value-based deep reinforcement learning in Atari games. The third contribution of this dissertation is a scalable algorithm to learn a state representation from a very large number of auxiliary tasks through deep learning. It is a stochastic gradient descent method to learn the principal components of a target matrix by means of a neural network from a handful of entries. Finally, the last part presents our findings on how the state representation in reinforcement learning influences the quality of an agent’s predictions but is also shaped by these predictions. We provide a formal mathematical model for studying this phenomenon and show how these theoretical results can be leveraged to improve the quality of the learning process.