Packt+ | Advance your knowledge in tech

You're reading from Java Deep Learning Projects Implement 10 real-world deep learning applications using Deeplearning4j and open source APIs

Product type Paperback

Published in Jun 2018

Publisher Packt

ISBN-13 9781788997454

Length 436 pages

Edition 1st Edition

Languages

Java

Tools

Deeplearning4j

Concepts

Deep Learning

Table of Contents (17) Chapters

Title Page

Packt Upsell

Contributors

Preface

1. Getting Started with Deep Learning FREE CHAPTER

2. Cancer Types Prediction Using Recurrent Type Networks

3. Multi-Label Image Classification Using Convolutional Neural Networks

4. Sentiment Analysis Using Word2Vec and LSTM Network

5. Transfer Learning for Image Classification

6. Real-Time Object Detection using YOLO, JavaCV, and DL4J

7. Stock Price Prediction Using LSTM Network

8. Distributed Deep Learning – Video Classification Using Convolutional LSTM Networks

9. Playing GridWorld Game Using Deep Reinforcement Learning

10. Developing Movie Recommendation Systems Using Factorization Machines

11. Discussion, Current Trends, and Outlook

1. Other Books You May Enjoy

Leave a review - let other readers know what you think

Index

Playing the GridWorld game

For this project, I haven't used any visualization to demonstrate the states and actions. Rather it is a text-based game, as I alluded to earlier. Then you can run the GridWorld.java class (containing the main method) using following invocation:

DeepQNetwork RLNet = newDeepQNetwork(conf, 100000, .99f, 1d, 1024, 500, 1024, InputLength, 4);

In this invocation, here's the parameter description outlined:

conf: This is the MultiLayerConfiguration used to create the DQN
100000: This is the replay memory capacity
.99f: The discount
1d: This is the epsilon
1024: The batch size
500: This is the update frequency; second 1,024 is the replay start size
InputLength: This is the input length of size x size x 2 + 1= 33 (considering size=4)
4: This is the number of possible actions that can be performed by the agent.

We initialize epsilon (ϵ-greedy action selection) to 1, which will decrease by a small amount on every episode. This way, it will eventually reach 0.1 and saturate. Based on...