You're reading from Deep Reinforcement Learning with Python Master classic RL, deep RL, distributional RL, inverse RL, and more with OpenAI Gym and TensorFlow

Product type Paperback

Published in Sep 2020

Publisher Packt

ISBN-13 9781839210686

Length 760 pages

Edition 2nd Edition

Languages

Python

Tools

Deep Reinforcement Learning

Concepts

Deep Reinforcement Learning

Author (1):

Sudharsan Ravichandiran

View More author details

Table of Contents (22) Chapters

Preface

1. Fundamentals of Reinforcement Learning

2. A Guide to the Gym Toolkit FREE CHAPTER

3. The Bellman Equation and Dynamic Programming

4. Monte Carlo Methods

5. Understanding Temporal Difference Learning

6. Case Study – The MAB Problem

7. Deep Learning Foundations

8. A Primer on TensorFlow

9. Deep Q Network and Its Variants

10. Policy Gradient Method

11. Actor-Critic Methods – A2C and A3C

12. Learning DDPG, TD3, and SAC

13. TRPO, PPO, and ACKTR Methods

14. Distributional Reinforcement Learning

15. Imitation Learning and Inverse RL

16. Deep Reinforcement Learning with Stable Baselines

17. Reinforcement Learning Frontiers

18. Other Books You May Enjoy

19. Index

Appendix 1 – Reinforcement Learning Algorithms

1. Appendix 2 – Assessments

The double DQN

We have learned that in DQN, the target value is computed as:

One of the problems with a DQN is that it tends to overestimate the Q value of the next state-action pair in the target:

This overestimation is due to the presence of the max operator. Let's see how this overestimation happens with an example. Suppose we are in a state and we have three actions a₁, a₂, and a₃. Assume a₃ is the optimal action in the state . When we estimate the Q values of all the actions in state , the estimated Q value will have some noise and differ from the actual value. Say, due to the noise, action a₂ will get a higher Q value than the optimal action a₃.

We know that the target value is computed as:

Now, if we select the best action as the one that has the maximum value then we will end up selecting the action a₂ instead of optimal action a₃, as shown here: