Introduction

Abstract

The goal of reinforcement learning [149] is to enable autonomous agents to learn effective control policies for challenging tasks. Rather than relying on directions from a human expert, a reinforcement learning agent uses its experience interacting with the world to infer a strategy for solving the given problem. Unlike supervised learning methods [96], reinforcement learning methods do not need access to examples of correct or incorrect behavior. Instead, the agent needs only a reward signal to quantify the immediate effects of its actions and it can learn a control policy to maximize the reward it accrues in the long term.