Grokking Deep Reinforcement Learning

front matter

foreword

So, here’s the thing about reinforcement learning. It is difficult to learn and difficult to teach, for a number of reasons. First, it’s quite a technical topic. There is a great deal of math and theory behind it. Conveying the right amount of background without drowning in it is a challenge in and of itself.

Second, reinforcement learning encourages a conceptual error. RL is both a way of thinking about decision-making problems and a set of tools for solving those problem. By “a way of thinking,” I mean that RL provides a framework for making decisions: it discusses states and reinforcement signals, among other details. When I say “a set of tools,” I mean that when we discuss RL, we find ourselves using terms like