You're reading from Deep Reinforcement Learning Hands-On. - Second Edition

Product typeBook

Published inJan 2020

Reading LevelIntermediate

PublisherPackt

ISBN-139781838826994

Edition2nd Edition

Languages

Python

Tools

Keras TensorFlow

Concepts

Deep Reinforcement Learning

Author (1)

Maxim Lapan

Ways to Speed up RL

In Chapter 8, DQN Extensions, you saw several practical tricks to make the deep Q-network (DQN) method more stable and converge faster. They involved the basic DQN method modifications (like injecting noise into the network or unrolling the Bellman equation) to get a better policy, with less time spent on training. But there is another way: tweaking the implementation details of the method to improve the speed of the training. This is a pure engineering approach, but it's also important in practice.

In this chapter, we will:

Take the Pong environment from Chapter 8 and try to get it solved as fast as possible
In a step-by-step manner, get Pong solved 3.5 times faster using exactly the same commodity hardware
Discuss fancier ways to speed up reinforcement learning (RL) training that could become common in the future

Why speed matters

First, let's talk a bit about why speed is important and why we optimize it at all. It might not be obvious, but enormous hardware performance improvements have happened in the last decade or two. 14 years ago, I was involved with a project that focused on building a supercomputer for computational fluid dynamics (CFD) simulations performed by an aircraft engine design company. The system consisted of 64 servers, occupied three 42-inch racks, and required dedicated cooling and power subsystems. The hardware alone (without cooling) cost almost $1M.

In 2005, this supercomputer occupied fourth place for Russian supercomputers and was the fastest system installed in the industry. Its theoretical performance was 922 GFLOPS (billion floating-point operations per second), but in comparison to the GTX 1080 Ti released 12 years later, all the capabilities of this pile of iron look tiny.

One single GTX 1080 Ti is able to perform 11,340 GFLOPS, which is 12.3 times...

The baseline

In the rest of the chapter, we will take the Atari Pong environment that you are already familiar with and try to speed up its convergence. As a baseline, we will take the same simple DQN that we used in Chapter 8, DQN Extensions, and the hyperparameters will also be the same. To compare the effect of our changes, we will use two characteristics:

The number of frames that we consume from the environment every second (FPS). It indicates how fast we can communicate with the environment during the training. It is very common in RL papers to indicate the number of frames that the agent observed during the training; normal numbers are 25M-50M frames. So, if our FPS=200, it will take days. In such calculations, you need to take into account that RL papers commonly report raw environment frames. But if frame skip is used (and it almost always is), this number needs to be divided by the frame skip factor, which is commonly equal to 4. In our measurements, we calculate...

The computation graph in PyTorch

Our first examples won't be around speeding up the baseline, but will show one common, and not always obvious, situation that can cost you performance. In Chapter 3, Deep Learning with PyTorch, we discussed the way PyTorch calculates gradients: it builds the graph of all operations that you perform on tensors, and when you call the backward() method of the final loss, all gradients in the model parameters are automatically calculated.

This works well, but RL code is normally much more complex than traditional supervised learning models, so the RL model that we are currently training is also being applied to get the actions that the agent needs to perform in the environment. The target network discussed in Chapter 6 makes it even more tricky. So, in DQN, a neural network (NN) is normally used in three different situations:

When we want to calculate Q-values predicted by the network to get the loss in respect to reference Q-values approximated...

Several environments

The first idea that we usually apply to speed up deep learning training is larger batch size. It's applicable to the domain of deep RL, but you need to be careful here. In the normal supervised learning case, the simple rule "a large batch is better" is usually true: you just increase your batch as your GPU memory allows, and a larger batch normally means more samples will be processed in a unit of time thanks to enormous GPU parallelism.

The RL case is slightly different. During the training, two things happen simultaneously:

Your network is trained to get better predictions on the current data
Your agent explores the environment

As the agent explores the environment and learns about the outcome of its actions, the training data changes. In a shooter example, your agent can run randomly for a time while being shot by monsters and have only a miserable "death is everywhere" experience in the training buffer. But after...

Play and train in separate processes

At a high level, our training contains a repetition of the following steps:

Ask the current network to choose actions and execute them in our array of environments
Put observations into the replay buffer
Randomly sample the training batch from the replay buffer
Train on this batch

The purpose of the first two steps is to populate the replay buffer with samples from the environment (which are observation, action, reward, and next observation). The last two steps are for training our network.

The following is an illustration of the preceding steps that will make potential parallelism slightly more obvious. On the left, the training flow is shown. The training steps use environments, the replay buffer, and our NN. The solid lines show data and code flow.

Dotted lines represent usage of the NN for training and inference.

Figure 9.6: A sequential diagram of the training process

As you can see, the top two steps...

Tweaking wrappers

The final step in our sequence of experiments will be tweaking wrappers applied to the environment. This is very easy to overlook, as wrappers are normally written once, or just borrowed from other code, applied to the environment, and left to sit there. But you should be aware of their importance in terms of the speed and convergence of your method. For example, the normal DeepMind-style stack of wrappers applied to an Atari game looks like this:

NoopResetEnv: applies a random amount of NOOP operations to the game reset. In some Atari games, this is needed to remove weird initial observations.
MaxAndSkipEnv: applies max to N observations (four by default) and returns this as an observation for the step. This solves the "flickering" problem in some Atari games, when the game draws different portions of the screen on even and odd frames (a normal practice among 2600 developers to increase the complexity of the game's sprites).
EpisodicLifeEnv...

Benchmark summary

I summarized our experiments in the following table. The percentages show the changes versus the baseline version.

...

Going hardcore: CuLE

During the writing of this chapter, NVIDIA researchers published the paper and code for their latest experiments with porting the Atari emulator on GPU: Steven Dalton, Iuri Frosio, GPU-Accelerated Atari Emulation for Reinforcement Learning, 2019, arXiv:1907.08467. The code of their Atari port is called CuLE (CUDA Learning Environment) and is available on GitHub: https://github.com/NVlabs/cule.

According to their paper, by keeping both the Atari emulator and NN on the GPU, they were able to get Pong solved within one to two minutes and reach FPS of 50k (on the advantage actor-critic (A2C) method, which will be the subject of the next part of the book).

Unfortunately, at the time of writing, the code wasn't stable enough. I failed to make it work on my hardware, but I hope that when you read this, the situation will have already changed. In any case, this project shows a somewhat extreme, but very efficient, way to increase RL methods' performance...

Summary

In this chapter, you saw several ways to improve the performance of the RL method using a pure engineering approach, which was in contrast to the "algorithmic" or "theoretical" approach covered in Chapter 8, DQN Extensions. From my perspective, both approaches complement each other, and a good RL practitioner needs to both know the latest tricks that researchers have found and be aware of the implementation details.

In the next chapter, we will apply our DQN knowledge to stocks trading.

References

https://www.aicrowd.com/challenges/neurips-2019-learn-to-move-walk-around

The rest of the chapter is locked

You have been reading a chapter from

Deep Reinforcement Learning Hands-On. - Second Edition

Published in: Jan 2020Publisher: PacktISBN-13: 9781838826994

A free Packt account unlocks extra newsletters, articles, discounted offers, and much more. Start advancing your knowledge today.

undefined

Unlock this book and the full library FREE for 7 days

Get unlimited access to 7000+ expert-authored eBooks and videos courses covering every tech area you can think of

Start free trial

Renews at $15.99/month. Cancel anytime

Author (1)

Maxim Lapan

Maxim has been working as a software developer for more than 20 years and was involved in various areas: distributed scientific computing, distributed systems and big data processing. Since 2014 he is actively using machine and deep learning to solve practical industrial tasks, such as NLP problems, RL for web crawling and web pages analysis. He has been living in Germany with his family.
Read more about Maxim Lapan

Other recommended products

Related to this chapter

TensorFlow Reinforcement Learning Quick Start Guide

This book is an essential guide for anyone interested in Reinforcement Learning. The book provides an actionable reference for Reinforcement Learning algorithms and their applications using TensorFlow and Python. It will help readers leverage the power of algorithms such as Deep Q-Network (DQN), Deep Deterministic Policy Gradients (DDPG), and Proximal Policy Optimization (PPO) to solve challenging control and decision-making problems.

BookMar 2019184 pages

Deep Reinforcement Learning with Python

Deep Reinforcement Learning with Python - Second Edition will help you learn reinforcement learning algorithms, techniques and architectures – including deep reinforcement learning – from scratch. This new edition is an extensive update of the original, reflecting the state-of-the-art latest thinking in reinforcement learning.

BookSep 2020760 pages

Hands-On Intelligent Agents with OpenAI Gym

Walks through the hands-on process of building intelligent agents from the basics and all the way up to solving complex problems including playing Atari games and driving a car autonomously in the CARLA simulator. Discusses various learning environments and how to transform real-world problems into learning environments and solve using the agents.

BookJul 2018254 pages

Hands-On Q-Learning with Python

Q-learning is the reinforcement learning approach behind Deep-Q-Learning and is a values-based learning algorithm in RL. This book will help you get comfortable with developing the effective agents for Q learning and also make you learn to effectively develop and deploy Deep Q networks for complex AI applications.

BookApr 2019212 pages

PyTorch 1.x Reinforcement Learning Cookbook

This book presents practical solutions to the most common reinforcement learning problems. The recipes in this book will help you understand the fundamental concepts to develop popular RL algorithms. You will gain practical experience in the RL domain using the modern offerings of the PyTorch 1.x library.

BookOct 2019340 pages

TensorFlow 2 Reinforcement Learning Cookbook

This cookbook will help you to gain a solid understanding of deep reinforcement learning (RL) algorithms with the help of concise, easy-to-follow implementations from scratch. You'll learn how to implement these algorithms with minimal code and develop AI applications to solve real-world and business problems using RL.

BookJan 2021472 pages

Reinforcement Learning Algorithms with Python

With this book, you will understand the core concepts and techniques of reinforcement learning. You will take a look into each RL algorithm and will develop your own self-learning algorithms and models. You will optimize the algorithms for better precision, use high-speed actions and lower the risk of anomalies in your applications.

BookOct 2019366 pages

Reinforcement Learning with TensorFlow

Reinforcement learning allows you to develop intelligent, self-learning systems. This book shows you how to put the concepts of Reinforcement Learning to train efficient models.You will use popular reinforcement learning algorithms to implement use-cases in image processing and NLP, by combining the power of TensorFlow and OpenAI Gym.

BookApr 2018334 pages

Hands-On Reinforcement Learning for Games

The AI revolution is here and it is embracing games. Game developers are being challenged to enlist cutting edge AI as part of their games. In this book, you will look at the journey of building capable AI using reinforcement learning algorithms and techniques. You will learn to solve complex tasks and build next-generation games using a practical approach.

BookJan 2020432 pages

Mastering Reinforcement Learning with Python

This book focuses on expert-level explanations and implementations of scalable reinforcement learning algorithms and approaches. Starting with the fundamentals, the book covers state-of-the-art methods from bandit problems to meta-reinforcement learning. You’ll also explore practical examples inspired by real-life problems from the industry.

BookDec 2020544 pages

Hands-On Reinforcement Learning with Python

Reinforcement learning is a self-evolving type of machine learning that takes us closer to achieving true artificial intelligence. This easy-to-follow guide explains everything from scratch using rich examples written in Python.

BookJun 2018318 pages

The Reinforcement Learning Workshop

With the help of practical examples and engaging activities, The Reinforcement Learning Workshop takes you through reinforcement learning’s core techniques and frameworks. Following a hands-on approach, it allows you to learn reinforcement learning at your own pace to develop your own intelligent applications with ease.

BookAug 2020822 pages

Personalised recommendations for you

Based on your interests and search pattern

Et al.

Ever wonder why speech recognition systems don't understand the Scottish accent, or what would happen if an astronaut only ate mac 'n' cheese, or other spurious reflections you'd have at a bar? We did, then collated those deliberations into absurd research articles with fake figures and methodologies inspired by even more fictionally absurd studies.

BookAug 2023230 pages5

Generative AI with LangChain

This book is a comprehensive introduction to LLMs and LangChain, demystifying the basic mechanics of LangChain, its functionalities, and the myriad of applications it can be integrated into.

BookDec 2023360 pages4

Generative AI with LangChain

This book is a comprehensive introduction to LLMs and LangChain, demystifying the basic mechanics of LangChain, its functionalities, and the myriad of applications it can be integrated into.

BookDec 2023360 pages5

Generative AI with LangChain

This book is a comprehensive introduction to LLMs and LangChain, demystifying the basic mechanics of LangChain, its functionalities, and the myriad of applications it can be integrated into.

BookDec 2023360 pages1

Generative AI with LangChain

This book is a comprehensive introduction to LLMs and LangChain, demystifying the basic mechanics of LangChain, its functionalities, and the myriad of applications it can be integrated into.

BookDec 2023360 pages5

Mastering Tableau 2023

This book is a comprehensive resource to mastering your Tableau skills and becoming a BI expert. As you progress, you will learn how to build advanced dashboards and improve your storytelling to derive key business insight, as well as make you well-versed with advanced functionalities of Tableau in the business intelligence domain.

BookAug 2023684 pages

Building AI Applications with ChatGPT APIs

This guide covers all ChatGPT API features for effortless creation of robust AI powered apps. With its help, you’ll be able to leverage ChatGPT’s cutting-edge NLP models to take your app development skills to the next level. You’ll also work on ten exciting projects that will give you the practical know-how that you can apply to your existing applications.

BookSep 2023258 pages5

Building AI Applications with ChatGPT APIs

This guide covers all ChatGPT API features for effortless creation of robust AI powered apps. With its help, you’ll be able to leverage ChatGPT’s cutting-edge NLP models to take your app development skills to the next level. You’ll also work on ten exciting projects that will give you the practical know-how that you can apply to your existing applications.

BookSep 2023258 pages2

Data Engineering with AWS

Embark on a journey to master data engineering pipelines on AWS! Our book offers a hands-on experience of AWS services for ingesting, transforming, and consuming data. Whether you're an absolute beginner or someone with basic data engineering experience, this guide is an indispensable resource.

BookOct 2023636 pages5

Modern Data Architecture on AWS

Every organization wants an agile, performant, and cost-effective data platform that meets all their current and future business needs. Purpose-built AWS analytics services and their features play a big part in building such a modern data platform. This book brings to you all the design and architectural patterns that’ll help you achieve this goal.

BookAug 2023420 pages5

Practical Guide to Applied Conformal Prediction in Python

Discover the power of Conformal Prediction with the "Practical Guide to Applied Conformal Prediction in Python." Master the latest techniques to quantify uncertainty in machine learning and computer vision models, and seamlessly apply them to your industry applications.

BookDec 2023240 pages

TinyML Cookbook

With over 70 project-based recipes, the TinyML Cookbook is a practical guide that will help you to get the most out of your microcontrollers. It provides a comprehensive understanding of the theoretical foundations while giving you hands-on experience training ML models for deployment on Arduino Nano 33 BLE Sense, Raspberry Pi Pico, and SparkFun RedBoard Artemis Nano microcontrollers.

BookNov 2023664 pages

Step	FPS	Solution time in minutes
Baseline	159	105
Version without `torch.no_grad()`	157 (-1.5%)	115 (+10%)
Three environments	228 (+43%)	44 (-58%)
Parallel version	402 (+152%)	59 (-43%)
Tweaked wrappers, three environments	330 (+108%)	40 (-62%)
Tweaked wrappers, parallel version	409 (+157%)	31 (-70%)