Designing Autonomous AI A Guide for Machine Teaching (Kence Anderson) (Z-Library)

A nd erson Kence Anderson Designing Autonomous AI A Guide for Machine Teaching

MACHINE LE ARNING ”An easy and compre- hensive introduction to machine teaching, a powerful new par- adigm of useful AI for industrial applications that empowers engi- neers, expert operators, innovation leaders, and business owners to optimize industrial processes by teaching the AI the expertise that took them decades to learn.” —Teresa Escrig Principal Project Manager, Microsoft Designing Autonomous AI US $59.99 CAN $74.99 ISBN: 978-1-098-11075-8 Twitter: @oreillymedia linkedin.com/company/oreilly-media youtube.com/oreillymedia Early rules-based artificial intelligence demonstrated intriguing decision-making capabilities but lacked perception and didn’t learn. AI today, primed with machine learning perception and deep reinforcement learning capabilities, can perform superhuman decision-making for specific tasks. This book shows you how to combine the practicality of early AI with deep learning capabilities and industrial control technologies to make robust decisions in the real world. Using concrete examples, minimal theory, and a proven architectural framework, author Kence Anderson demonstrates how to teach autonomous AI explicit skills and strategies. You’ll learn when and how to use and combine various AI architecture design patterns, as well as how to design advanced AI without needing to manipulate neural networks or machine learning algorithms. Students, process operators, data scientists, machine learning algorithm experts, and engineers who own and manage industrial processes can use the methodology in this book to design autonomous AI. This book examines: • Differences between and limitations of automated, autonomous, and human decision-making • Unique advantages of autonomous AI for real-time decision- making, with use cases • How to design an autonomous AI system from modular components and document your designs Kence Anderson is director of Autonomous AI Adoption at Microsoft. Kence has pioneered uses for autonomous AI in industry and designed over 150 autonomous decision-making AI systems for large enterprises. He now teaches autonomous AI design and consults enterprises on how to build their autonomous systems organizations and practices. A nd erson

Praise for Designing Autonomous AI An easy and comprehensive introduction to machine teaching, a powerful new paradigm of useful AI for industrial applications that empowers engineers, expert operators, innovation leaders, and business owners to optimize industrial processes by teaching the AI the expertise that took them decades to learn. —Teresa Escrig, Principal Project Manager, Microsoft By steering clear of the morass of AI’s mathematical and programming details, Kence Anderson has produced an engaging book that will appeal to a wide swath of readers with an interest in this transformational technology. His thoughtful examples illustrate the key ideas behind AI in ways that will resonate with readers. —Jonathan Schaeffer, Distinguished Professor of Computer Science, University of Alberta The insight that learning systems, by design, must be taught spawned the machine teaching paradigm. This book captures hard-won patterns and practices to make applying it to real-world applications successful. If you’re building an autonomous system, you couldn’t ask for a better guide than Kence Anderson. —Mark Hammond, VP, Autonomous Systems, Microsoft; Cofounder and CEO, Bonsai AI Inc. Machine teaching is a key skill in the next wave of useful, applied AI. This book is a practical, real-world guide, developed from a wealth of experience empowering subject matters in applying AI to business-valuable projects. —Phil Harvey, Autonomous Systems Architect, Bonsai, Microsoft Research

Kence’s brilliance shines through in his use of real-world stories and simple analogies to make the complex approachable. This book is the introduction to a technology that will be an inflection point in the world of manufacturing and industry. —Bryan DeBois, Director of Industrial AI, RoviSys I enthusiastically recommend this book. Kence Anderson, a leading real-world practitioner in emergent autonomous AI, writes about machine teaching with the lucidity and authority of someone who had a hand in its creation (which he did). —Gurdeep Pall, Corporate Vice President, Head of Product Incubations, Microsoft (from the foreword)

Kence Anderson Designing Autonomous AI A Guide for Machine Teaching Boston Farnham Sebastopol TokyoBeijing

978-1-098-11075-8 [LSI] Designing Autonomous AI by Kence Anderson Copyright © 2022 Kence Anderson. All rights reserved. Printed in the United States of America. Published by O’Reilly Media, Inc., 1005 Gravenstein Highway North, Sebastopol, CA 95472. O’Reilly books may be purchased for educational, business, or sales promotional use. Online editions are also available for most titles (http://oreilly.com). For more information, contact our corporate/institutional sales department: 800-998-9938 or corporate@oreilly.com. Acquisitions Editor: Nicole Butterfield Development Editor: Sarah Grey Production Editor: Jonathon Owen Copyeditor: Justin Billing Proofreader: Piper Editorial Consulting, LLC Indexer: Ellen Troutman-Zaig Interior Designer: David Futato Cover Designer: Karen Montgomery Illustrator: Kate Dullea June 2022: First Edition Revision History for the First Edition 2022-06-10: First Release See http://oreilly.com/catalog/errata.csp?isbn=9781098110758 for release details. The O’Reilly logo is a registered trademark of O’Reilly Media, Inc. Designing Autonomous AI, the cover image, and related trade dress are trademarks of O’Reilly Media, Inc. The views expressed in this work are those of the author, and do not represent the publisher’s views. While the publisher and the author have used good faith efforts to ensure that the information and instructions contained in this work are accurate, the publisher and the author disclaim all responsibility for errors or omissions, including without limitation responsibility for damages resulting from the use of or reliance on this work. Use of the information and instructions contained in this work is at your own risk. If any code samples or other technology this work contains or describes is subject to open source licenses or the intellectual property rights of others, it is your responsibility to ensure that your use thereof complies with such licenses and/or rights.

This book is dedicated to my father, Lawrence Anderson (1916–2002), the maestro and master teacher.

Table of Contents Foreword. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiii Preface. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xv Introduction: The Right Brain in the Right Place (Why We Need Autonomous AI). . . . . . . xxv Part I. When Automation Doesn’t Work 1. Sometimes Machines Make Bad Decisions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 Math, Menus, and Manuals: How Machines Make Automated Decisions 5 Control Theory Uses Math to Calculate Decisions 5 Optimization Algorithms Use Menus of Options to Evaluate Decisions 9 Expert Systems Recall Stored Expertise 21 2. The Quest for More Human-Like Decision-Making. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 Augmenting Human Intelligence 28 How Humans Make Decisions and Acquire Skills 29 Humans Act on What They Perceive 30 Humans Build Complex Correlations into Their Intuition with Practice 31 Humans Abstract to Strategy for Complex Tasks 31 There’s a New Kind of AI in Town 36 The Superpowers of Autonomous AI 40 Autonomous AI Makes More Human-Like Decisions 41 Autonomous AI Perceives, Then Acts 41 The Difference Between Perception and Action in AI 42 Autonomous AI Learns and Adapts When Things Change 43 Autonomous AI Can Spot Patterns 43 vii

Autonomous AI Infers from Experience 44 Autonomous AI Improvises and Strategizes 44 Autonomous AI Can Plan for the Long-Term Future 45 Autonomous AI Brings Together the Best of All Decision-Making Technologies 46 When Should You Use Autonomous AI? 46 Autonomous AI Is like a Brilliant, Curious Toddler That Needs to Be Taught 47 Part II. What Is Machine Teaching? 3. How Brains Learn Best: Teaching Humans and AI. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53 Learning Multiple Skills Simultaneously Is Hard for Humans and AI 53 Teaching Skills and Strategies Explicitly 54 Teaching Allows Us to Trust AI 58 The Mindset of a Machine Teacher 60 Teacher More Than Programmer 60 Learner More Than Expert 62 What Is a Brain Design? 62 How Decision-Making Works 63 Acquiring Skill Is like Learning to Navigate by Exploring 68 A Brain Design Is a Mental Map That Guides Exploration with Landmarks 69 4. Building Blocks for Machine Teaching. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73 Case Study: Learning to Walk Is Hard to Evolve, Easier to Teach 76 So, Why Do We Walk? 77 Strategy Versus Evolution 78 Teaching Walking as Three Skills 81 Concepts Capture Knowledge 84 Skills Are Specialized Concepts 84 Brains Are Built from Skills 86 Building Skills 86 Expert Rules Inflate into Skills 87 Perceptive Concepts Discern or Recognize 91 Directive Concepts Decide and Act 96 Selective Concepts Supervise and Assign 97 Brains Are Organized by Functions and Strategies 99 Sequences or Parallel Execution for Functional Skills 100 Hierarchies for Strategies 108 Visual Language of Brain Design 113 viii | Table of Contents

Part III. How Do You Teach a Machine? Understanding the Process 117 Meet with Experts 118 Ask the Right Questions 118 Case Study: Let’s Design a Smart Thermostat 119 5. Teaching Your AI Brain What to Do. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121 Determining Which Actions the Brain Will Take 122 Perception Is Required, but It’s Not All We Need 122 Sequential Decisions 123 Triggering the Action in Your AI Brain 124 Setting the Decision Frequency 125 Handling Delayed Consequences for Brain Actions 125 Actions for Smart Thermostat 127 6. Setting Goals for Your AI Brain. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129 There’s Always a Trade-off 129 Throughput Versus Efficiency 131 Supervisors Have Different Goals Than Crews Do 132 Don’t Prioritize Goals; Balance Them Instead 133 Watch Out for Expert Rules Disguised as Goals 133 Ideal Versus Available 134 Setting Goals 135 Step 1: Identify Scenarios 135 Step 2: Match Goals to Scenarios 136 Step 3: Teach Strategies for Each Scenario 137 Goal Objectives 137 Maximize 137 Minimize 137 Reach, like the Finish Line for a Race 137 Drive, like the Temperature for a Thermostat 138 Avoid, like Dangerous Conditions 138 Standardize, like the Heat in an Oven 139 Smooth, like a Line 139 Expanding Task Algebra to Include Goal Objectives 140 Setting Goals for a Smart Thermostat 141 7. Teaching Skills to Your AI Brain. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143 Teaching Focuses and Guides Practice (Exploration) 144 Skills Can Evolve and Transform 147 Skills Adapt to the Scenario 148 Levels of Teaching Sophistication 148 Table of Contents | ix

The Introductory Teacher Conveys the Facts and Goals 149 The Coach Sequences Skills to Practice 149 The Mentor Teaches Strategy 150 The Maestro Democratizes New Paradigms 151 How Maestros Democratize Technology 153 Levels of Autonomous AI Architecture 154 Machine Learning Adds Perception 155 Monolithic Brains Are Advanced Beginners 156 Concept Networks Are Competent Learners 157 Massive Concept Networks Are Proficient Learners 159 Pursuing Expert Skill Acquisition in Autonomous AI 160 Brains That Come with Hardwired Skills 161 Brains That Define Skills as They Learn 162 Brains That Assemble Themselves 164 Brains with Skills That Coordinate 165 Steps to Architect an AI Brain 166 Step 1: Identify the Skills That You Want to Teach 166 Step 2: Orchestrate How the Skills Work Together 168 Step 3: Select Which Technology Should Perform Each Skill 168 Pitfalls to Avoid When Teaching Skills 168 Pitfall 1: Confusing the solution for the problem 169 Pitfall 2: Losing the forest for the trees 169 Example of Teaching Skills to an AI Brain: Rubber Factory 169 Brain Design for Our Smart Thermostat 171 8. Giving Your AI Brain the Information It Needs to Learn and Decide. . . . . . . . . . . . . . 173 Sensors: The Five Senses for Your AI Brain 174 Variables 174 Proxy Variables 175 Trends 176 Simulators: A Gym for Your Autonomous AI to Practice In 176 Simulating Reality Using Physics and Chemistry 178 Simulating Reality Using Statistics and Events 179 Simulating Reality Using Machine Learning 179 Simulating Reality Using Expert Rules 180 Sensor Variables for Smart Thermostat 180 Part IV. Tools for the Machine Teacher 9. Designing AI Brains That Someone Can Actually Build. . . . . . . . . . . . . . . . . . . . . . . . . 185 Designers and Builders Working Together in Harmony (Mostly) 185 x | Table of Contents

The Autonomous AI Design Fallacy Designs but Won’t Iterate 187 The Autonomous AI Implementation Fallacy Skips Design Altogether 188 Specification for Documenting AI Brain Designs 188 Platform for Machine Teaching 190 Platform for Wiring Multiple Skills Together as Modules 190 What Difference Will You Make with Machine Teaching? 191 Glossary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193 Index. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195 Table of Contents | xi

Foreword It gives me great pleasure to enthusiastically recommend Designing Autonomous AI, written by a leading real-world practitioner in emergent autonomous AI: Kence Anderson. As the Corporate Vice President currently responsible for new product incubations at Microsoft, and as someone who has been closely involved with creation of new prod‐ uct categories like VPNs, real-time communications and now autonomous systems at Microsoft, I have learned that creating new categories is not a straightforward path. Established product categories and their underlying technologies and business mod‐ els are well honed for the real world with well understood problems and solutions. These are things we use in our daily lives. Fundamental research is important to new category creation but lives in the ether of science, generally abstracted from real world considerations. Researchers create novel methods and demonstrate their promise through results in carefully crafted experiments—the amazing state of the art in AI is a recent example of this. Importantly, not every method created by researchers is sufficient or workable in the real world. New categories can be created when one or more of these new methods are developed further, grounded in and applied to the real-world environment. New category creation is a chaotic, primordial process entailing a loop of exploration, learning, and adaptation. This process is very high dimensional and domain depen‐ dent; there is no single formula here. However, at the core of new category creation, inevitably there is always a breakthrough clinching idea that solves the most difficult aspects of the real-world problem. The new category of autonomous AI, which is so significant that it is being called the Fourth Industrial Revolution, is particularly challenging in this regard because the underlying real-world considerations are extremely onerous. Building adequate systems that are safe and performant in the real world is nontrivial. This is a key reason for why we haven’t seen flying cars outside our window yet. xiii

For autonomous AI, one of those breakthrough clinching ideas is machine teaching. In the same way languages and compilers made the once complex and kludgy task of computing easy, efficient, and accessible to millions of programmers, machine teach‐ ing is enabling experts from any discipline (mechanical, electrical, or aeronautical engineering, etc.) to build powerful and usable autonomous systems easily, without requiring a PhD in AI or data science. Machine teaching, without apologies, borrows from well-established techniques from all schools of thoughts of AI: connectionist, symbolist, and experiential, and from proven methods in pedagogy. Machine teaching provides an easy way to codify human expertise in any domain and to apply the power of learning methods auto‐ matically, abstracting away the tyranny of abstruse AI algorithms. Kence Anderson writes about machine teaching with the lucidity and authority of someone who had a hand in its creation (which he did). — Gurdeep Pall (Seattle, WA, 2022) xiv | Foreword

Preface There’s a saying that “those who can’t do, teach; and those who can’t teach, write.” Well, that’s not how this book came to be. I spent four years traveling around the world to steel mills, mines, factories, testing facilities, design studios, machine shops, chemical plants, oil fields, refineries, warehouses, and logistics centers, learning from subject matter experts about the challenges and opportunities of industrial decision- making, then working with them to design useful Artificial Intelligence (AI) that can help them make better decisions. What Is Autonomous AI? Autonomous AI is AI-powered automation that optimizes equipment and processes by sensing and responding in real time. The great feat of modern AI is programming algorithms that can adapt and change their behavior based on feedback. The goal of Designing Autonomous AI is to show you how to put these learning algorithms to work by teaching AI to make successful decisions in real, production environments. I’m not claiming that AI brains can achieve parity with humans or match human capability in any area (I will often refer to a specific instance of autonomous AI as a brain). I’m saying that when an autonomous AI is designed properly and makes full use of existing AI and automation components, it can radically outperform systems that calculate actions from known mathematical relationships, search and select actions using objective criteria, or look up actions from recorded human expertise. Let’s look at each of the attributes a little more deeply. There’s a large gap to bridge between AI research and useful human-like decision- making in industrial situations. So, when AI research demonstrated that AI can learn to perform complex tasks, I took on the task to figure out what kinds of decisions AI can learn to make in the real world. You see, at that time (four years ago), when I started on this journey to probe the capabilities of AI decision-making, the xv

only decisions that learning AI was shown to effectively make were in video games and drastically oversimplified “toy” control problems like you’d find in a first-year university physics or engineering textbook: simplified to the point that a student can calculate what to do next as a word problem. Fast forward to today and I’ve designed over 150 autonomous AI for real applications at large companies. Many of them were built and some of them make valuable, effective decisions that previously only human beings could make. Each one performs a single useful, specific, material task in an enterprise process. After all of those AI projects about everything from controlling bulldozers to ware‐ house scheduling to food manufacturing, here’s my conclusion: real, industrial pro‐ cesses are complex and the decisions to control and optimize them are fuzzy and full of tradeoffs. The human expertise that drives these processes is vast and deep and can’t be replaced by algorithms that search options for solutions or even advanced calculating control systems. Autonomous AI can make a material improvement in control and optimization of these systems and processes, but you have to be willing to wade into how these processes work and learn from subject matter experts to design AI that will produce these kinds of breakthrough results. If you’re looking for diatribes about whether AI is overhyped or whether AI will ever achieve the full capabilities of the human mind, then this isn’t the book for you. Every day I read rants about how AI is either complete marketing hype that has little differentiated value (as if AI is really an element of fiction) or that AI is track‐ ing toward superintelligence and is a serious competitor with the human mind in general cognitive ability (this perspective is more like science fiction). Both of these perspectives cannot be true. My opinion is and my experience shows that AI does have unique decision-making capabilities that differentiate it from other technology, but that it is best used to make specific high-value decisions that complement—not replicate—the human mind. So, if you’re looking for a discourse about how stupid or about how scary intelligent AI is, you’re not likely to find what you’re looking for here in this book. If you’re looking for a path, a plan, and the tools to design AI that can solve even currently unsolvable problems, right now, then you’re in the right place. One of the reasons that some make wild claims about AI capability, while others simultaneously constantly cry “Hype!” about AI feats and techniques is because the discourse is missing requisite nuanced discussion about the capabilities of AI compared to current methods for a particular task. Take natural language processing as the first example. I’m writing this introduction from Spain. I don’t need an AI that can understand and comprehend human language to find out how to say my xvi | Preface

hotel room number in Spanish. The machine translation of 143 (ciento cuarenta y tres) was particularly useful for me on this trip. I used it every day to get into the hotel breakfast area. However, the AI that translated my room number for me is not suitable at all for summarizing paragraphs (another language related task) or writing novels (an even more difficult language related task). The exact same is true for most every autonomous AI I’ve ever designed. If you take it outside of the context of the task that I designed it for, it very well might be “hype.” But if you use it for the task that I designed it for, the task that it practiced mastering over time, that’s no hype. It will outperform existing automation and sometimes attain human expert status. This book is full of examples of such performance achievements, which you will be able to produce in your own AI by the time you’ve finished reading. The first step to successful brain design puts the right brain in the right place. Find situations where machines are making bad decisions that autonomous AI can make more effectively. Who Should Read This Book? Process Experts This book is for the 100 million subject matter experts out there who manage and seek to automate complex equipment and processes. Gartner reported in 2018 that there were approximately 10,000 data scientists in the world. Let’s approximate that this means there are on the order of 10,000 AI experts who can design and build autonomous AI from scratch using code. Most of these experts hold PhD degrees in areas related to AI. In contrast, there are on the order of ten million software engineers in the world. Most of these developers specialize in writing software applications but are not specialists in AI. Their area of subject matter expertise is writing software and they can do this across many diverse applications. Then there’s the 100 million or so subject matter experts in the world. These mechanical engineers, chemical engineers, process engineers, controls engineers, supply chain analysts, logistics analysts, and many others design and manage complex equipment and processes and they know this equipment and these processes inside and out. These populations are visualized in Figure P-1. Preface | xvii

Figure P-1. Diagram of primary intended audience for this book. While AI experts, data scientists, and software engineers can certainly use this book to design autonomous AI, I wrote this book for 100 million subject matter experts who want to make their systems and processes more autonomous. I didn’t need a PhD in AI to devise this framework for designing autonomous AI, and you don’t need one to use it. Data Scientists and Software Engineers The field of data science is booming. Unfortunately, I’ve seen innovation organiza‐ tions pair up process experts and software engineers to great success, but I’ve seen many more organizations expect data scientists to solve process problems as if they’re wizards wielding magic, with limited influence from and access to process experts. This book can help data scientists learn how to take a bird’s-eye view of systems and processes to effectively integrate process expertise. The result will be better, more deployable autonomous AI. xviii | Preface