Reinforcement Learning From Scratch: Understanding Current Approaches - with Examples in Java and Greenfoot

ISBN-13: 9783031090295 / Angielski / Twarda / 2022 / 184 str.

Uwe Lorenz

Opis

In ancient games such as chess or go, the most brilliant players can improve by studying the strategies produced by a machine. Robotic systems practice their own movements. In arcade games, agents capable of learning reach superhuman levels within a few hours. How do these spectacular reinforcement learning algorithms work? With easy-to-understand explanations and clear examples in Java and Greenfoot, you can acquire the principles of reinforcement learning and apply them in your own intelligent agents. Greenfoot (M.Kölling, King's College London) and the hamster model (D. Bohles, University of Oldenburg) are simple but also powerful didactic tools that were developed to convey basic programming concepts. The result is an accessible introduction into machine learning that  concentrates on reinforcement learning. Taking the reader through the steps of developing intelligent agents, from the very basics to advanced aspects, touching on a variety of machine learning algorithms along the way, one is allowed to play along, experiment, and add their own ideas and experiments.  

Spis treści

Foreword of Michael Koelling (King’s College London)

​Preface

Introduction

Chapter 1: Reinforcement learning as subfield of machine learning

1.1 Machine Learning as automated processing of feedback from the environment

1.2 Machine Learning methods

1.3 Reinforcement Learning with Java

Bibliography

Chapter 2: Basic concepts of reinforcement learning

2.1 Agents

2.2 Control of the agent

2.3 Evaluation of states and actions (Q-function, Bellman equation)

Bibliography

Chapter 3: Optimal decision-making in a known environment

3.1 Value Iteration

3.1.1 Target-oriented state evaluation (“backward induction”)

3.1.2 Policy-based state valuation (reward prediction)

3.2 Iterative policy search

3.2.1 Direct policy improvement

3.2.2 Mutual improvement of policy and value-function

3.3 Optimal policy in a board game scenario

3.4 Summary

Bibliography

Chapter 4: decision making and learning in an unknown environment

4.1 Exploration vs. exploitation

4.2 Retroactive processing of experience ("model-free reinforcement learning")

4.2.1 Goal-oriented learning ("value-based")

4.2.2 Policy search

4.2.3 Combined methods (Actor-Critic)

4.3 Exploration with predictive simulations ("Model-Based Reinforcement Learning")

4.3.1 Dyna-Q

4.3.2 Monte-Carlo rollout

4.3.3 Artificial curiosity

4.3.4 Monte Carlo Tree Search (MCTS).

4.3.5 Remarks on the Concept of Intelligence

4.4 Systematic of learning methods

Bibliography

Chapter 5: Artificial Neural Networks as estimators for state values and the action selection

5.1 Artificial neural networks

5.1.1     Pattern recognition with the perceptron

5.1.2     The adaptability of artificial neural networks

5.1.3     Backpropagation Learning

5.1.4     Regression with multilayer perceptrons

5.2 State evaluation with generalizing approximations

5.3         Neural estimators for action selection

5.3.1     Policy gradient with neural networks

5.3.2     Proximal Policy Optimization

5.3.3     Evolutionary strategy with a neural policy

Bibliography

Chapter 6: Guiding ideas in Artificial Intelligence over time

6.1 Changing basic ideas

6.2 On the relationship between humans and Artificial Intelligence

Bibliography

 

 

O autorze

After studying computer science and philosophy with a focus on artificial intelligence and machine learning at the Humboldt University Berlin and for a few years as a project engineer, Uwe Lorenz currently works as a high school teacher for computer science and mathematics and at the Free University of Berlin in the Computing Education Research Group, - since his first contact with computers at the end of the 1980s he couldn't let go of the topic of artificial intelligence.

Blurb

In ancient games such as chess or go, the most brilliant players can improve by studying the strategies produced by a machine. Robotic systems practice their own movements. In arcade games, agents capable of learning reach superhuman levels within a few hours. How do these spectacular reinforcement learning algorithms work? 

With easy-to-understand explanations and clear examples in Java and Greenfoot, you can acquire the principles of reinforcement learning and apply them in your own intelligent agents. Greenfoot (M.Kölling, King's College London) and the hamster model (D. Bohles, University of Oldenburg) are simple but also powerful didactic tools that were developed to convey basic programming concepts. 

The result is an accessible introduction into machine learning that  concentrates on reinforcement learning. Taking the reader through the steps of developing intelligent agents, from the very basics to advanced aspects, touching on a variety of machine learning algorithms along the way, one is allowed to play along, experiment, and add their own ideas and experiments.  

This book is a translation of an original German edition. The translation was done with the help of artificial intelligence (machine translation by the service DeepL.com). A subsequent human revision was done primarily in terms of content, so that the book will read stylistically differently from a conventional translation.