Deep Learning and the Game of Go 1st Edition by Kevin Ferguson, Max Pumperla 1638354014 9781638354017

Part 1. Foundations

1 Toward deep learning: a machine-learning introduction

1.1. What is machine learning?

1.1.1. How does machine learning relate to AI?

1.1.2. What you can and can’t do with machine learning

1.2. Machine learning by example

1.2.1. Using machine learning in software applications

1.2.2. Supervised learning

1.2.3. Unsupervised learning

1.2.4. Reinforcement learning

1.3. Deep learning

1.4. What you’ll learn in this book

1.5. Summary

2 Go as a machine-learning problem

2.1. Why games?

2.2. A lightning introduction to the game of Go

2.2.1. Understanding the board

2.2.2. Placing and capturing stones

2.2.3. Ending the game and counting

2.2.4. Understanding ko

2.3. Handicaps

2.4. Where to learn more

2.5. What can we teach a machine?

2.5.1. Selecting moves in the opening

2.5.2. Searching game states

2.5.3. Reducing the number of moves to consider

2.5.4. Evaluating game states

2.6. How to measure your Go AI’s strength

2.6.1. Traditional Go ranks

2.6.2. Benchmarking your Go AI

2.7. Summary

3 Implementing your first Go bot

3.1. Representing a game of Go in Python

3.1.1. Implementing the Go board

3.1.2. Tracking connected groups of stones in Go: strings

3.1.3. Placing and capturing stones on a Go board

3.2. Capturing game state and checking for illegal moves

3.2.1. Self-capture

3.2.2. Ko

3.3. Ending a game

3.4. Creating your first bot: the weakest Go AI imaginable

3.5. Speeding up game play with Zobrist hashing

3.6. Playing against your bot

3.7. Summary

Part 2. Machine learning and game AI

4 Playing games with tree search

4.1. Classifying games

4.2. Anticipating your opponent with minimax search

4.3. Solving tic-tac-toe: a minimax example

4.4. Reducing search space with pruning

4.4.1. Reducing search depth with position evaluation

4.4.2. Reducing search width with alpha-beta pruning

4.5. Evaluating game states with Monte Carlo tree search

4.5.1. Implementing Monte Carlo tree search in Python

4.5.2. How to select which branch to explore

4.5.3. Applying Monte Carlo tree search to Go

4.6. Summary

5 Getting started with neural networks

5.1. A simple use case: classifying handwritten digits

5.1.1. The MNIST data set of handwritten digits

5.1.2. MNIST data preprocessing

5.2. The basics of neural networks

5.2.1. Logistic regression as simple artificial neural network

5.2.2. Networks with more than one output dimension

5.3. Feed-forward networks

5.4. How good are our predictions? Loss functions and optimization

5.4.1. What is a loss function?

5.4.2. Mean squared error

5.4.3. Finding minima in loss functions

5.4.4. Gradient descent to find minima

5.4.5. Stochastic gradient descent for loss functions

5.4.6. Propagating gradients back through your network

5.5. Training a neural network step-by-step in Python

5.5.1. Neural network layers in Python

5.5.2. Activation layers in neural networks

5.5.3. Dense layers in Python as building blocks for feed-forward networks

5.5.4. Sequential neural networks with Python

5.5.5. Applying your network handwritten digit classification

5.6. Summary

6 Designing a neural network for Go data

6.1. Encoding a Go game position for neural networks

6.2. Generating tree-search games as network training data

6.3. Using the Keras deep-learning library

6.3.1. Understanding Keras design principles

6.3.2. Installing the Keras deep-learning library

6.3.3. Running a familiar first example with Keras

6.3.4. Go move prediction with feed-forward neural networks in Keras

6.4. Analyzing space with convolutional networks

6.4.1. What convolutions do intuitively

6.4.2. Building convolutional neural networks with Keras

6.4.3. Reducing space with pooling layers

6.5. Predicting Go move probabilities

6.5.1. Using the softmax activation function in the last layer

6.5.2. Cross-entropy loss for classification problems

6.6. Building deeper networks with dropout and rectified linear units

6.6.1. Dropping neurons for regularization

6.6.2. The rectified linear unit activation function

6.7. Putting it all together for a stronger Go move-prediction network

6.8. Summary

7 Learning from data: a deep-learning bot

7.1. Importing Go game records

7.1.1. The SGF file format

7.1.2. Downloading and replaying Go game records from KGS

7.2. Preparing Go data for deep learning

7.2.1. Replaying a Go game from an SGF record

7.2.2. Building a Go data processor

7.2.3. Building a Go data generator to load data efficiently

7.2.4. Parallel Go data processing and generators

7.3. Training a deep-learning model on human game-play data

7.4. Building more-realistic Go data encoders

7.5. Training efficiently with adaptive gradients

7.5.1. Decay and momentum in SGD

7.5.2. Optimizing neural networks with Adagrad

7.5.3. Refining adaptive gradients with Adadelta

7.6. Running your own experiments and evaluating performance

7.6.1. A guideline to testing architectures and hyperparameters

7.6.2. Evaluating performance metrics for training and test data

7.7. Summary

8 Deploying bots in the wild

8.1. Creating a move-prediction agent from a deep neural network

8.2. Serving your Go bot to a web frontend

8.2.1. An end-to-end Go bot example

8.3. Training and deploying a Go bot in the cloud

8.4. Talking to other bots: the Go Text Protocol

8.5. Competing against other bots locally

8.5.1. When a bot should pass or resign

8.5.2. Let your bot play against other Go programs

8.6. Deploying a Go bot to an online Go server

8.6.1. Registering a bot at the Online Go Server

8.7. Summary

9 Learning by practice: reinforcement learning

9.1. The reinforcement-learning cycle

9.2. What goes into experience?

9.3. Building an agent that can learn

9.3.1. Sampling from a probability distribution

9.3.2. Clipping a probability distribution

9.3.3. Initializing an agent

9.3.4. Loading and saving your agent from disk

9.3.5. Implementing move selection

9.4. Self-play: how a computer program practices

9.4.1. Representing experience data

9.4.2. Simulating games

9.5. Summary

10 Reinforcement learning with policy gradients

10.1. How random games can identify good decisions

10.2. Modifying neural network policies with gradient descent

10.3. Tips for training with self-play

10.3.1. Evaluating your progress

10.3.2. Measuring small differences in strength

10.3.3. Tuning a stochastic gradient descent (SGD) optimizer

10.4. Summary

11 Reinforcement learning with value methods

11.1. Playing games with Q-learning

11.2. Q-learning with Keras

11.2.1. Building two-input networks in Keras

11.2.2. Implementing the ϵ-greedy policy with Keras

11.2.3. Training an action-value function

11.3. Summary

12 Reinforcement learning with actor-critic methods

12.1. Advantage tells you which decisions are important

12.1.1. What is advantage?

12.1.2. Calculating advantage during self-play

12.2. Designing a neural network for actor-critic learning

12.3. Playing games with an actor-critic agent

12.4. Training an actor-critic agent from experience data

12.5. Summary

Part 3. Greater than the sum of its parts

13 AlphaGo: Bringing it all together

13.1. Training deep neural networks for AlphaGo

13.1.1. Network architectures in AlphaGo

13.1.2. The AlphaGo board encoder

13.1.3. Training AlphaGo-style policy networks

13.2. Bootstrapping self-play from policy networks

13.3. Deriving a value network from self-play data

13.4. Better search with policy and value networks

13.4.1. Using neural networks to improve Monte Carlo rollouts

13.4.2. Tree search with a combined value function

13.4.3. Implementing AlphaGo’s search algorithm

13.5. Practical considerations for training your own AlphaGo

13.6. Summary

14 AlphaGo Zero: Integrating tree search with reinforcement learning

14.1. Building a neural network for tree search

14.2. Guiding tree search with a neural network

14.2.1. Walking down the tree

14.2.2. Expanding the tree

14.2.3. Selecting a move

14.3. Training

14.4. Improving exploration with Dirichlet noise

14.5. Modern techniques for deeper neural networks

14.5.1. Batch normalization

14.5.2. Residual networks

14.6. Exploring additional resources

14.7. Wrapping up

14.8. Summary

Appendix A. Mathematical foundations

Vectors, matrices, and beyond: a linear algebra primer

Vectors: one-dimensional data

Matrices: two-dimensional data

Rank 3 tensors

Rank 4 tensors

Calculus in five minutes: derivatives and finding maxima

Appendix B. The backpropagation algorithm

A bit of notation

The backpropagation algorithm for feed-forward networks

Backpropagation for sequential neural networks

Backpropagation for neural networks in general

Computational challenges with backpropagation

Appendix C. Go programs and servers

Go programs

GNU Go

Pachi

Go servers

OGS

IGS

Tygem

Appendix D. Training and deploying bots by using Amazon Web Services

Model training on AWS

Hosting a bot on AWS over HTTP

Appendix E. Submitting a bot to the Online Go Server

Registering and activating your bot at OGS

Testing your OGS bot locally

Deploying your OGS bot on AWS

Index

List of Figures

List of Tables

List of Listings