Reinforcement Learning Algorithms with Python: Learn, understand, and develop smart algorithms for addressing AI challenges

Reinforcement Learning Algorithms with Python

Learn, understand, and develop smart algorithms for addressing AI challenges

Andrea Lonza

BIRMINGHAM - MUMBAI

Reinforcement Learning Algorithms with Python

Copyright © 2019 Packt Publishing

All rights reserved. No part of this book may be reproduced, stored in a retrieval system, or transmitted in any form or by any means, without the prior written permission of the publisher, except in the case of brief quotations embedded in critical articles or reviews.

Every effort has been made in the preparation of this book to ensure the accuracy of the information presented. However, the information contained in this book is sold without warranty, either express or implied. Neither the author, nor Packt Publishing or its dealers and distributors, will be held liable for any damages caused or alleged to have been caused directly or indirectly by this book.

Packt Publishing has endeavored to provide trademark information about all of the companies and products mentioned in this book by the appropriate use of capitals. However, Packt Publishing cannot guarantee the accuracy of this information.

Commissioning Editor: Pravin Dhandre

Acquisition Editor: Winston Christopher

Content Development Editor: Roshan Kumar

Senior Editor: Jack Cummings

Technical Editor: Joseph Sunil

Copy Editor: Safis Editing

Project Coordinator: Kirti Pisat

Proofreader: Safis Editing

Indexer: Rekha Nair

Production Designer: Nilesh Mohite

First published: October 2019

Production reference: 1181019

Published by Packt Publishing Ltd.

Livery Place

35 Livery Street

Birmingham

B3 2PB, UK.

ISBN 978-1-78913-111-6

www.packt.com

Thanks to you, Mom and Dad, for giving me that light called life and for always being present for me.  Fede, you're a furious mad. You've always inspired me to do more. Thanks brother.

Packt.com

Subscribe to our online digital library for full access to over 7,000 books and videos, as well as industry leading tools to help you plan your personal development and advance your career. For more information, please visit our website.

Why subscribe?

Spend less time learning and more time coding with practical eBooks and Videos from over 4,000 industry professionals

Improve your learning with Skill Plans built especially for you

Get a free eBook or video every month

Fully searchable for easy access to vital information

Copy and paste, print, and bookmark content

Did you know that Packt offers eBook versions of every book published, with PDF and ePub files available? You can upgrade to the eBook version at www.packt.com and as a print book customer, you are entitled to a discount on the eBook copy. Get in touch with us at [email protected] for more details.

At www.packt.com, you can also read a collection of free technical articles, sign up for a range of free newsletters, and receive exclusive discounts and offers on Packt books and eBooks. 

Contributors

About the author

Andrea Lonza is a deep learning engineer with a great passion for artificial intelligence and a desire to create machines that act intelligently. He has acquired expert knowledge in reinforcement learning, natural language processing, and computer vision through academic and industrial machine learning projects. He has also participated in several Kaggle competitions, achieving high results. He is always looking for compelling challenges and loves to prove himself.

About the reviewer

Greg Walters has been involved with computers and computer programming since 1972. He is extremely well-versed in Visual Basic, Visual Basic .NET, Python and SQL using MySQL, SQLite, Microsoft SQL Server, Oracle, C++, Delphi, Modula-2, Pascal, C, 80x86 Assembler, COBOL, and Fortran. He is a programming trainer and has trained numerous people on many pieces of computer software, including MySQL, Open Database Connectivity, Quattro Pro, Corel Draw!, Paradox, Microsoft Word, Excel, DOS, Windows 3.11, Windows for Workgroups, Windows 95, Windows NT, Windows 2000, Windows XP, and Linux. He is retired and, in his spare time, is a musician and loves to cook, but he is also open to working as a freelancer on various projects.

Packt is searching for authors like you

If you're interested in becoming an author for Packt, please visit authors.packtpub.com and apply today. We have worked with thousands of developers and tech professionals, just like you, to help them share their insight with the global tech community. You can make a general application, apply for a specific hot topic that we are recruiting an author for, or submit your own idea.

Table of Contents

Title Page

Copyright and Credits

Reinforcement Learning Algorithms with Python

Dedication

About Packt

Why subscribe?

Contributors

About the author

About the reviewer

Packt is searching for authors like you

Preface

Who this book is for

What this book covers

To get the most out of this book

Download the example code files

Download the color images

Conventions used

Get in touch

Reviews

Section 1: Algorithms and Environments

The Landscape of Reinforcement Learning

An introduction to RL

Comparing RL and supervised learning

History of RL

Deep RL

Elements of RL

Policy

The value function

Reward

Model

Applications of RL

Games 

Robotics and Industry 4.0

Machine learning

Economics and finance

Healthcare

Intelligent transportation systems

Energy optimization and smart grid

Summary

Questions

Further reading

Implementing RL Cycle and OpenAI Gym

Setting up the environment

Installing OpenAI Gym 

Installing Roboschool 

OpenAI Gym and RL cycles

Developing an RL cycle

Getting used to spaces

Development of ML models using TensorFlow

Tensor

Constant

Placeholder

Variable

Creating a graph

Simple linear regression example

Introducing TensorBoard

Types of RL environments

Why different environments?

Open source environments

Summary

Questions

Further reading

Solving Problems with Dynamic Programming

MDP

Policy

Return

Value functions

Bellman equation

Categorizing RL algorithms

Model-free algorithms

Value-based algorithms

Policy gradient algorithms

Actor-Critic algorithms

Hybrid algorithms

Model-based RL

Algorithm diversity

Dynamic programming

Policy evaluation and policy improvement

Policy iteration

Policy iteration applied to FrozenLake

Value iteration

Value iteration applied to FrozenLake

Summary

Questions

Further reading

Section 2: Model-Free RL Algorithms

Q-Learning and SARSA Applications

Learning without a model

User experience

Policy evaluation

The exploration problem

Why explore?

How to explore

TD learning

TD update

Policy improvement

Comparing Monte Carlo and TD

SARSA

The algorithm

Applying SARSA to Taxi-v2

Q-learning

Theory

The algorithm

Applying Q-learning to Taxi-v2

Comparing SARSA and Q-learning

Summary

Questions

Deep Q-Network

Deep neural networks and Q-learning

Function approximation 

Q-learning with neural networks

Deep Q-learning instabilities

DQN

The solution

Replay memory

The target network

The DQN algorithm

The loss function

Pseudocode

Model architecture

DQN applied to Pong

Atari games 

Preprocessing

DQN implementation

DNNs

The experienced buffer

The computational graph and training loop

Results

DQN variations

Double DQN

DDQN implementation

Results

Dueling DQN

Dueling DQN implementation

Results

N-step DQN

Implementation

Results

Summary

Questions

Further reading

Learning Stochastic and PG Optimization

Policy gradient methods

The gradient of the policy

Policy gradient theorem

Computing the gradient

The policy

On-policy PG

Understanding the REINFORCE algorithm

Implementing REINFORCE

Landing a spacecraft using REINFORCE

Analyzing the results

REINFORCE with baseline

Implementing REINFORCE with baseline

Learning the AC algorithm

Using a critic to help an actor to learn

The n-step AC model

The AC implementation

Landing a spacecraft using AC 

Advanced AC, and tips and tricks

Summary

Questions

Further reading

TRPO and PPO Implementation

Roboschool

Control a continuous system

Natural policy gradient

Intuition behind NPG

A bit of math

FIM and KL divergence

Natural gradient complications

Trust region policy optimization

The TRPO algorithm

Implementation of the TRPO algorithm

Application of TRPO

Proximal Policy Optimization

A quick overview

The PPO algorithm

Implementation of PPO

PPO application

Summary

Questions

Further reading

DDPG and TD3 Applications

Combining policy gradient optimization with Q-learning

Deterministic policy gradient

Deep deterministic policy gradient

The DDPG algorithm

DDPG implementation

Appling DDPG to BipedalWalker-v2

Twin delayed deep deterministic policy gradient (TD3)

Addressing overestimation bias

Implementation of TD3

Addressing variance reduction

Delayed policy updates

Target regularization

Applying TD3 to BipedalWalker

Summary

Questions

Further reading

Section 3: Beyond Model-Free Algorithms and Improvements

Model-Based RL

Model-based methods

A broad perspective on model-based learning

A known model

Unknown model

Advantages and disadvantages

Combining model-based with model-free learning

A useful combination

Building a model from images

ME-TRPO applied to an inverted pendulum

Understanding ME-TRPO

Implementing ME-TRPO

Experimenting with RoboSchool

Results on RoboSchoolInvertedPendulum

Summary

Questions

Further reading

Imitation Learning with the DAgger Algorithm

Technical requirements

Installation of Flappy Bird

The imitation approach

The driving assistant example

Comparing IL and RL

The role of the expert in imitation learning

The IL structure

Comparing active with passive imitation

Playing Flappy Bird

How to use the environment

Understanding the dataset aggregation algorithm

The DAgger algorithm

Implementation of DAgger

Loading the expert inference model

Creating the learner's computational graph

Creating a DAgger loop

Analyzing the results on Flappy Bird

IRL

Summary

Questions

Further reading

Understanding Black-Box Optimization Algorithms

Beyond RL

A brief recap of RL

The alternative

EAs

The core of EAs

Genetic algorithms

Evolution strategies

CMA-ES

ES versus RL

Scalable evolution strategies

The core

Parallelizing ES

Other tricks

Pseudocode

Scalable implementation

The main function

Workers

Applying scalable ES to LunarLander

Summary

Questions

Further reading

Developing the ESBAS Algorithm

Exploration versus exploitation

Multi-armed bandit

Approaches to exploration

The ∈-greedy strategy

The UCB algorithm

UCB1

Exploration complexity

Epochal stochastic bandit algorithm selection

Unboxing algorithm selection

Under the hood of ESBAS

Implementation

Solving Acrobot

Results

Summary

Questions

Further reading

Practical Implementation for Resolving RL Challenges

Best practices of deep RL

Choosing the appropriate algorithm

From zero to one

Challenges in deep RL

Stability and reproducibility

Efficiency

Generalization

Advanced techniques

Unsupervised RL

Intrinsic reward

Transfer learning

Types of transfer learning

1-task learning

Multi-task learning

RL in the real world

Facing real-world challenges

Bridging the gap between simulation and the real world

Creating your own environment

Future of RL and its impact on society

Summary

Questions

Further reading

Assessments

Other Books You May Enjoy

Leave a review - let other readers know what you think

Preface

Reinforcement learning (RL) is a popular and promising branch of artificial intelligence that involves making smarter models and agents that can automatically determine ideal behavior based on changing requirements. Reinforcement Learning Algorithms with Python will help you master RL algorithms and understand their implementation as you build self-learning agents.

Starting with an introduction to the tools, libraries, and setup needed to work in the RL environment, this book covers the building blocks of RL and delves into value-based methods such as the application of Q-learning and SARSA algorithms. You'll learn how to use a combination of Q-learning and neural networks to solve complex problems. Furthermore, you'll study policy gradient methods, TRPO, and PPO, to improve performance and stability, before moving on to the DDPG and TD3 deterministic algorithms. This book also covers how imitation learning techniques work and how Dagger can teach an agent to fly. You'll discover evolutionary strategies and black-box optimization techniques. Finally, you'll get to grips with exploration approaches such as UCB and UCB1 and develop a meta-algorithm called ESBAS.

By the end of the book, you'll have worked with key RL algorithms to overcome challenges in real-world applications, and you'll be part of the RL research community.

Who this book is for

If you are an AI researcher, deep learning user, or anyone who wants to learn RL from scratch, this book is for you. You'll also find this RL book useful if you want to learn about the advancements in the field. Working knowledge of Python is necessary.

What this book covers

Chapter 1, The Landscape of Reinforcement Learning, gives you an insight into RL. It describes the problems that RL is good at solving and the applications where RL algorithms are already adopted. It also introduces the tools, the libraries, and the setup needed for the completion of the projects in the following chapters.

Chapter 2, Implementing RL Cycle and OpenAI Gym, describes the main cycle of the RL algorithms, the toolkit used to develop the algorithms, and the different types of environments. You will be able to develop a random agent using the OpenAI Gym interface to play CartPole using random actions. You will also learn how to use the OpenAI Gym interface to run other environments.

Chapter 3, Solving Problems with Dynamic Programming, introduces to you the core ideas, terminology, and approaches of RL. You will learn about the main blocks of RL and develop a general idea about how RL algorithms can be created to solve a problem. You will also learn the differences between model-based and model-free algorithms and the categorization of reinforcement learning algorithms. Dynamic programming will be used to solve the game FrozenLake.

Chapter 4, Q-Learning and SARSA Applications, talks about value-based methods, in particular Q-learning and SARSA, two algorithms that differ from dynamic programming and scale well on large problems. To become confident with these algorithms, you will apply them to the FrozenLake game and study the differences from dynamic programming.

Chapter 5, Deep Q-Networks, describes how neural networks and convolutional neural networks (CNNs) in particular are applied to Q-learning. You'll learn why the combination of Q-learning and neural networks produces incredible results and how its use can open the door to a much larger variety of problems. Furthermore, you'll apply the DQN to an Atari game using the OpenAI Gym interface.

Chapter 6, Learning Stochastic and PG Optimization, introduces a new family of model-free algorithms: policy gradient methods. You will learn the differences between policy gradient and value-based methods, and you'll learn about their strengths and weaknesses. Then you will implement the REINFORCE and Actor-Critic algorithms to solve a new game called LunarLander.

Chapter 7, TRPO and PPO Implementation, proposes a modification of policy gradient methods using new mechanisms to control the improvement of the policy. These mechanisms are used to improve the stability and convergence of the policy gradient algorithms. In particular you'll learn and implement two main policy gradient methods that use these techniques, namely TRPO and PPO. You will implement them on RoboSchool, an environment with a continuous action space.

Chapter 8, DDPG and TD3 Applications, introduces a new category of algorithms called deterministic policy algorithms that combine both policy gradient and Q-learning. You will learn about the underlying concepts and implement DDPG and TD3, two deep deterministic algorithms, on a new environment.

Chapter 9, Model-Based RL, illustrates RL algorithms that learn the model of the environment to plan future actions, or, to learn a policy. You will be taught how they work, their strengths, and why they are preferred in many situations. To master them, you will implement a model-based algorithm on Roboschool.

Chapter 10, Imitation Learning with the DAgger Algorithm, explains how imitation learning works and how it can be applied and adapted to a problem. You will learn about the most well-known imitation learning algorithm, DAgger. To become confident with it, you will implement it to speed up the learning process of an agent on FlappyBird.

Chapter 11, Understanding Black-Box Optimization Algorithms, explores evolutionary algorithms, a class of black-box optimization algorithms that don't rely on backpropagation. These algorithms are gaining interest because of their fast training and easy parallelization across hundreds or thousands of cores. This chapter provides a theoretical and practical background of these algorithms by focusing particularly on the Evolution Strategy algorithm, a type of evolutionary algorithm.

Chapter 12, Developing ESBAS Algorithm, introduces the important exploration-exploitation dilemma, which is specific to RL. The dilemma is demonstrated using the multi-armed bandit problem and is solved using approaches such as UCB and UCB1. Then, you will learn about the problem of algorithm selection and develop a meta-algorithm called ESBAS. This algorithm uses UCB1 to select the most appropriate RL algorithm for each situation.

Chapter 13, Practical Implementations to Resolve RL Challenges, takes a look at the major challenges in this field and explains some practices and methods to overcome them. You will also learn about some of the challenges of applying RL to real-world problems, future developments of deep RL, and their social impact in the world.

To get the most out of this book

Working knowledge of Python is necessary. Knowledge of RL and the various tools used for it will also be beneficial.

Download the example code files

You can download the example code files for this book from your account at www.packt.com. If you purchased this book elsewhere, you can visit www.packtpub.com/support and register to have the files emailed directly to you.

You can download the code files by following these steps:

Log in or register at www.packt.com.

Select the Support tab.

Click on Code Downloads.

Enter the name of the book in the Search box and follow the onscreen instructions.

Once the file is downloaded, please make sure that you unzip or extract the folder using the latest version of:

WinRAR/7-Zip for Windows

Zipeg/iZip/UnRarX for Mac

7-Zip/PeaZip for Linux

The code bundle for the book is also hosted on GitHub at https://github.com/PacktPublishing/Reinforcement-Learning-Algorithms-with-Python. In case there's an update to the code, it will be updated on the existing GitHub repository.

We also have other code bundles from our rich catalog of books and videos available at https://github.com/PacktPublishing/. Check them out!

Download the color images

We also provide a PDF file that has color images of the screenshots/diagrams used in this book. You can download it here: http://www.packtpub.com/sites/default/files/downloads/9781789131116_ColorImages.pdf.

Conventions used

There are a number of text conventions used throughout this book.

CodeInText: Indicates code words in text, database table names, folder names, filenames, file extensions, pathnames, dummy URLs, user input, and Twitter handles. Here is an example: In this book, we use Python 3.7, but all versions above 3.5 should work. We also assume that you've already installed numpy and matplotlib.

A block of code is set as follows:

import gym

# create the environment

env = gym.make(CartPole-v1)

# reset the environment before starting

env.reset()

# loop 10 times

for i in range(10):

    # take a random action

    env.step(env.action_space.sample())

    # render the game

  env.render()

# close the environment

env.close()

Any command-line input or output is written as follows:

$ git clone https://github.com/pybox2d/pybox2d

$ cd pybox2d

$ pip install -e .

Bold: Indicates a new term, an important word, or words that you see onscreen. For example, words in menus or dialog boxes appear in the text like this. Here is an example: "In reinforcement learning (RL), the algorithm is called the agent, and it learns from the data provided by an environment."

Warnings or important notes appear like this.

Tips and tricks appear like this.

Get in touch

Feedback from our readers is always welcome.

General feedback: If you have questions about any aspect of this book, mention the book title in the subject of your message and email us at [email protected].

Errata: Although we have taken every care to ensure the accuracy of our content, mistakes do happen. If you have found a mistake in this book, we would be grateful if you would report this to us. Please visit www.packtpub.com/support/errata, selecting your book, clicking on the Errata Submission Form link, and entering the details.

Piracy: If you come across any illegal copies of our works in any form on the Internet, we would be grateful if you would provide us with the location address or website name. Please contact us at [email protected] with a link to the material.

If you are interested in becoming an author: If there is a topic that you have expertise in and you are interested in either writing or contributing to a book, please visit authors.packtpub.com.

Reviews

Please leave a review. Once you have read and used this book, why not leave a review on the site that you purchased it from? Potential readers can then see and use your unbiased opinion to make purchase decisions, we at Packt can understand what you think about our products, and our authors can see your feedback on their book. Thank you!

For more information about Packt, please visit packt.com.

Section 1: Algorithms and Environments

This section is an introduction to reinforcement learning. It includes building the theoretical foundation and setting up the environment that is needed in the upcoming chapters.

This section includes the following chapters:

Chapter 1, The Landscape of Reinforcement Learning

Chapter 2,Implementing RL Cycle and OpenAI Gym

Chapter 3, Solving Problems with Dynamic Programming

The Landscape of Reinforcement Learning

Humans and animals learn through a process of trial and error. This process is based on our reward mechanisms that provide a response to our behaviors. The goal of this process is to, through multiple repetitions, incentivize the repetition of actions which trigger positive responses, and disincentivize the repetition of actions which trigger negative ones. Through the trial and error mechanism, we learn to interact with the people and world around us, and pursue complex, meaningful goals, rather than immediate gratification. 

Learning through interaction and experience is essential. Imagine having to learn to play football by only looking at other people playing it. If you took to the field to play a football match based on this learning experience, you would probably perform incredibly poorly.

This was demonstrated throughout the mid-20th century, notably by Richard Held and Alan Hein's 1963 study on two kittens, both of whom were raised on a carousel. One kitten was able to move freely (actively), whilst the other was restrained and moved following the active kitten (passively). Upon both kittens being introduced to light, only the kitten who was able to move actively developed a functioning depth perception and motor skills, whilst the passive kitten did not. This was notably demonstrated by the absence of the passive kitten's blink-reflex towards incoming objects. What this, rather crude experiment demonstrated is that regardless of visual deprivation, physical interaction with the environment is necessary in order for animals to learn. 

Inspired by how animals and humans learn, reinforcement learning (RL) is built around the idea of trial and error from active interactions with the environment. In particular, with RL, an agent learns incrementally as it interacts with the world. In this way, it's possible to train a computer to learn and behave in a rudimentary, yet similar way to how humans do.

This book is all about reinforcement learning. The intent of the book is to give you the best possible understanding of this field with a hands-on approach. In the first chapters, you'll start by learning the most fundamental concepts of reinforcement learning. As you grasp these concepts, we'll start developing our first reinforcement learning algorithms. Then, as the book progress, you'll create more powerful and complex algorithms to solve more interesting and compelling problems. You'll see that reinforcement learning is very broad and that there exist many algorithms that tackle a variety of problems in different ways. Nevertheless, we'll do our best to provide you with a simple but complete description of all the ideas, alongside a clear and practical implementation of the algorithms.

To start with, in this chapter, you'll familiarize yourself with the fundamental concepts of RL, the distinctions between different approaches, and the key concepts of policy, value function, reward, and model of the environment. You'll also learn about the history and applications of RL.

The following topics will be covered in this chapter:

An introduction to RL

Elements of RL

Applications of RL

An introduction to RL

RL is an area of machine learning that deals with sequential decision-making, aimed at reaching a desired goal. An RL problem is constituted by a decision-maker called an Agent and the physical or virtual world in which the agent interacts, is known as the Environment. The agent interacts with the environment in the form of Action which results in an effect. As a result, the environment will feedback to the agent a new State and Reward. These two signals are the consequences of the action taken by the agent. In particular, the reward is a value indicating how good or bad the action was, and the state is the current representation of the agent and the environment. This cycle is shown in the following diagram:

In this diagram the agent is represented by PacMan that based on the current state of the environment, choose which action to take. Its behavior will influence the environment, like its position and that of the enemies, that will be returned by the environment in the form of a new state and the reward. This cycle is repeated until the game ends.

The ultimate goal of the agent is to maximize the total reward accumulated during

its lifetime. Let's simplify the notation: if   is the action at time   and   is the reward at time  , then the agent will take actions 

, to maximize the sum of all rewards 

.

To maximize the cumulative reward, the agent has to learn the best behavior in every situation. To do so, the agent has to optimize for a long-term horizon while taking care of every single action. In environments with many discrete or continuous states and actions, learning is difficult because the agent should be accountable for each situation. To make the problem harder, RL can have very sparse and delayed rewards, making the learning process more arduous.

To give an example of an RL problem while explaining the complexity of a sparse reward, consider the well-known story of two siblings, Hansel and Gretel. Their parents led them into the forest to abandon them, but Hansel, who knew of their intentions, had taken a slice of bread with him when they left the house and managed to leave a trail of breadcrumbs that would lead him and his sister home. In the RL framework, the agents are Hansel and Gretel, and the environment is the forest. A reward of +1 is obtained for every crumb of bread reached and a reward of +10 is acquired when they reach home. In this case, the denser the trail of bread, the easier it will be for the siblings to find their way home. This is because to go from one piece of bread to another, they have to explore a smaller area. Unfortunately, sparse rewards are far more common than dense rewards in the real world.

An important characteristic of RL is that it can deal with environments that are dynamic, uncertain, and non-deterministic. These qualities are essential for the adoption of RL in the real world. The following points are examples of how real-world problems can be reframed in RL settings:

Self-driving cars are a popular, yet difficult, concept to approach with RL. This is because of the many aspects to be taken into consideration while driving on the road (such as pedestrians, other cars, bikes, and traffic lights) and the highly uncertain environment. In this case, the self-driving car is the agent that can act on the steering wheel, accelerator, and brakes. The environment is the world around it. Obviously, the agent cannot be aware of the whole world around it, as it can only capture limited information via its sensors (for example, the camera, radar, and GPS). The goal of the self-driving car is to reach the destination in the minimum amount of time while following the rules of the road and without damaging anything. Consequently, the agent can receive a negative reward if a negative event occurs and a positive reward can be received in proportion to the driving time when the agent reaches its destination.

In the game of chess, the goal is to checkmate the opponent's piece. In an RL framework, the player is the agent and the environment is the current state of the board. The agent is allowed to move the game pieces according to their own way of moving. As a result of an action, the environment returns a positive or negative reward corresponding to a win or a loss for the agent. In all other situations, the reward is 0 and the next state is the state of