Book Image

Deep Reinforcement Learning Hands-On

By : Maxim Lapan
Book Image

Deep Reinforcement Learning Hands-On

By: Maxim Lapan

Overview of this book

Deep Reinforcement Learning Hands-On is a comprehensive guide to the very latest DL tools and their limitations. You will evaluate methods including Cross-entropy and policy gradients, before applying them to real-world environments. Take on both the Atari set of virtual games and family favorites such as Connect4. The book provides an introduction to the basics of RL, giving you the know-how to code intelligent learning agents to take on a formidable array of practical tasks. Discover how to implement Q-learning on 'grid world' environments, teach your agent to buy and trade stocks, and find out how natural language models are driving the boom in chatbots.

Related Content you might be interested in

Current Title:

Small book image
Deep Reinforcement Learning Hands-On
Maxim Lapan
Jun, 2018 | 546 pages
Small book image
Reinforcement Learning Algorithms with Python
Andrea Lonza
Oct, 2019 | 366 pages
Small book image
TensorFlow Reinforcement Learning Quick Start Guide
Kaushik Balakrishnan
Mar, 2019 | 184 pages
Small book image
Hands-On Reinforcement Learning for Games
Micheal Lanham
Jan, 2020 | 432 pages
Table of Contents (23 chapters)
Deep Reinforcement Learning Hands-On
Deep Reinforcement Learning Hands-On
Contributors
Contributors
Preface
Preface
Other Books You May Enjoy
Other Books You May Enjoy
Free Chapter
1
What is Reinforcement Learning?
What is Reinforcement Learning?
Learning – supervised, unsupervised, and reinforcement
RL formalisms and relations
Markov decision processes
Summary
2
OpenAI Gym
OpenAI Gym
The anatomy of the agent
Hardware and software requirements
OpenAI Gym API
The random CartPole agent
The extra Gym functionality – wrappers and monitors
Summary
3
Deep Learning with PyTorch
Deep Learning with PyTorch
Tensors
Gradients
NN building blocks
Custom layers
Final glue – loss functions and optimizers
Monitoring with TensorBoard
Example – GAN on Atari images
Summary
4
The Cross-Entropy Method
The Cross-Entropy Method
Taxonomy of RL methods
Practical cross-entropy
Cross-entropy on CartPole
Cross-entropy on FrozenLake
Theoretical background of the cross-entropy method
Summary
5
Tabular Learning and the Bellman Equation
Tabular Learning and the Bellman Equation
Value, state, and optimality
The Bellman equation of optimality
Value of action
The value iteration method
Value iteration in practice
Q-learning for FrozenLake
Summary
6
Deep Q-Networks
Deep Q-Networks
Real-life value iteration
Tabular Q-learning
Deep Q-learning
DQN on Pong
Summary
7
DQN Extensions
DQN Extensions
The PyTorch Agent Net library
Basic DQN
N-step DQN
Double DQN
Noisy networks
Prioritized replay buffer
Dueling DQN
Categorical DQN
Combining everything
Summary
References
8
Stocks Trading Using RL
Stocks Trading Using RL
Trading
Data
Problem statements and key decisions
The trading environment
Models
Training code
Results
Things to try
Summary
9
Policy Gradients – An Alternative
Policy Gradients – An Alternative
Values and policy
The REINFORCE method
REINFORCE issues
PG on CartPole
PG on Pong
Summary
10
The Actor-Critic Method
The Actor-Critic Method
Variance reduction
CartPole variance
Actor-critic
A2C on Pong
A2C on Pong results
Tuning hyperparameters
Summary
11
Asynchronous Advantage Actor-Critic
Asynchronous Advantage Actor-Critic
Correlation and sample efficiency
Adding an extra A to A2C
Multiprocessing in Python
A3C – data parallelism
A3C – gradients parallelism
Summary
12
Chatbots Training with RL
Chatbots Training with RL
Chatbots overview
Deep NLP basics
Training of seq2seq
The chatbot example
Summary
13
Web Navigation
Web Navigation
Web navigation
OpenAI Universe
Simple clicking approach
Human demonstrations
Adding text description
Things to try
Summary
14
Continuous Action Space
Continuous Action Space
Why a continuous space?
Action space
Environments
The Actor-Critic (A2C) method
Deterministic policy gradients
Distributional policy gradients
Things to try
Summary
15
Trust Regions – TRPO, PPO, and ACKTR
Trust Regions – TRPO, PPO, and ACKTR
Introduction
Roboschool
A2C baseline
Proximal Policy Optimization
Trust Region Policy Optimization
A2C using ACKTR
Summary
16
Black-Box Optimization in RL
Black-Box Optimization in RL
Black-box methods
Evolution strategies
ES on CartPole
ES on HalfCheetah
Genetic algorithms
GA on CartPole
GA tweaks
GA on Cheetah
Summary
References
17
Beyond Model-Free – Imagination
Beyond Model-Free – Imagination
Model-based versus model-free
Model imperfections
Imagination-augmented agent
I2A on Atari Breakout
Experiment results
Summary
References
18
AlphaGo Zero
AlphaGo Zero
Board games
The AlphaGo Zero method
Connect4 bot
Connect4 results
Summary
References
Book summary
Index
Index

Results

Let’s now take a look at the results.

The feed-forward model

The convergence on Yandex data for one year requires about 10M training steps, which can take a while (GTX 1080Ti trains at a speed of 230-250 steps per second). During training, we have several charts in TensorBoard showing us what’s going on.

The following are two charts, reward_100 and steps_100, with average reward (which is in percentages) and the average length of the episode for the last 100 episodes, respectively:

Figure 3: The reward plot for the feed-forward version

The charts show us two good things:

  1. Our agent was able to figure out when to buy and sell the share to get positive reward (as we need to pay a commission of 0.1% on the open and close of the position, random actions will have -0.2% reward).

  2. Over the training time, the length of the episode increased from seven bars to 25 and still continues to grow slowly, which means that the agent is holding the share longer and longer to increase the final profit.

Unfortunately...

Previous Section
End of Section 8
Next Section