foundations of computational agents
Artificial Intelligence: Foundations of Computational Agents is a book about the science of artificial intelligence (AI). AI is the study of intelligent computational agents. Our book is structured as a textbook but it is designed to be accessible to a wide audience.
We wrote this book because we are excited about the emergence of AI as an integrated science. As with any science being developed, AI has a coherent, formal theory and a rambunctious experimental wing. Here we balance theory and experiment and show how to link them together intimately. We develop the science of AI together with its engineering applications. We believe the adage, “There is nothing so practical as a good theory.” The spirit of our approach is captured by the dictum, “Everything should be made as simple as possible, but not simpler.” We must build the science on solid foundations; we present the foundations, but only sketch, and give some examples of, the complexity required to build useful intelligent systems. Although the resulting systems will be complex, the foundations and the building blocks should be simple.
This third edition results from extensive revision throughout the text. We have added three new chapters:
Neural Networks and Deep Learning
The Social Impact of Artificial Intelligence
There is also a social impact section for each chapter, covering either beneficial applications or harmful impacts of AI, and often both. With the rise in the use of AI in society, it is imperative that all AI researchers and practitioners understand the possible social impact of their work.
We have restructured the material based on feedback from instructors who have used the book in classes. We have brought it up to date to reflect the current state of the art, made parts that were difficult for students more straightforward, added more intuitive explanations, and coordinated the pseudocode algorithms with open-source Python implementations at AIPython (aipython.org). We have resisted the temptation to cover every recent advance.
AI research is expanding so rapidly now that the volume of potential new text material is vast. However, research teaches us not only what works but also what does not work so well, allowing us to be highly selective. We have included more material on techniques that have proven successful. However, research also has trends and fashions. We have removed techniques that have been shown to be less promising, but we distinguish them from the techniques for problems that are merely out of fashion. We include some currently unfashionable material if the problems attacked still remain and the techniques have the potential to form the basis for future research and development. We have further developed the concept of a single design space for intelligent agents, showing how many bewilderingly diverse techniques can be seen in a simple, uniform framework. This allows us to emphasize the principles underlying the foundations of computational agents, making those ideas more accessible to students.
The book can be used as an introductory text on artificial intelligence for advanced undergraduate or graduate students in computer science or related disciplines such as computer engineering, philosophy, cognitive science, or psychology. It will appeal more to the technically minded; parts are technically challenging, focusing on learning by doing: designing, building, and implementing systems. Any curious scientifically oriented reader will benefit from studying the book. Previous experience with computational systems is desirable, but prior study of the foundations upon which we build, including logic, probability, calculus, and control theory, is not necessary, because we develop the concepts as required.
The serious student will gain valuable skills at several levels ranging from expertise in the specification and design of intelligent agents to skills for implementing, testing, and improving real software systems for several challenging application domains. The thrill of participating in the emergence of a new science of intelligent agents is one of the attractions of this approach. The practical skills of dealing with a world of ubiquitous, intelligent, embedded agents are now in great demand in the marketplace.
The focus is on an intelligent agent acting in an environment. We start with simple agents acting in simple, static environments and gradually increase the power of the agents to cope with more challenging worlds. We explore ten dimensions of complexity that allow us to introduce, gradually and with modularity, what makes building intelligent agents challenging. We have tried to structure the book so that the reader can understand each of the dimensions separately and we make this concrete by repeatedly illustrating the ideas with four different agent tasks: a delivery robot, a diagnostic assistant, a tutoring agent, and a trading agent.
The agent we want the student to envision is a hierarchically designed agent that acts intelligently in a stochastic environment that it can only partially observe – one that reasons online about individuals and relationships among them, has complex preferences, learns while acting, takes into account other agents, and acts appropriately given its own computational limitations. Of course, we cannot start with such an agent; it is still a research question to build such agents. So we introduce the simplest agents and then show how to add each of these complexities in a modular way.
We have made a number of design choices which distinguish this book from competing books, including our earlier book.
We have tried to give a coherent framework in which to understand AI. We have chosen not to present disconnected topics that do not fit together. For example, we do not present disconnected logical and probabilistic views of AI, but we have presented a multidimensional design space in which students can understand the big picture, in which probabilistic and logical reasoning coexist.
We decided that it is better to clearly explain the foundations upon which more sophisticated techniques can be built, rather than present all these more sophisticated techniques. This means that a larger gap may exist between what is covered in this book and the frontier of science. But it also means that the student will have a better foundation to understand current and future research.
One of the more difficult decisions we made was how to linearize the design space. We have chosen a relations-late approach. This approach probably reflects better the research over the past few decades where there has been much progress in reasoning and learning for feature-based representations. The problems of relational learning and reasoning have not gone away, only become more urgent.
This should help students and instructors avoid being overwhelmed by the amount of technical details to be mastered before having a comprehensive view of AI.
We provide open-source Python implementations of most of the algorithms at http://www.aipython.org. These are as close to pseudo-code as we can make them while they still run. We have chosen clarity over efficiency, with good asymptotic complexity. They are not a replacement for a well-engineered library, because they are sometimes a few orders of magnitude slower. However, by keeping everything as simple as possible, the student can see how the algorithms work. The code provides the basic algorithms, and variants can be used as exercises. One choice we have made that might not match every instructor’s preference is to minimize the number of external Python libraries. We only use matplotlib, as we cannot get away without a plotting library if we want to present information.
We have chosen not to present an encyclopedic view of AI. Not every major idea that has been investigated is presented here. We have chosen some basic ideas upon which other, more sophisticated, techniques are based and have tried to explain the basic ideas in detail, sketching how these can be expanded. Once a student has understood the principles here they can go into more specialized topics such as vision, natural language understanding, or robotics.
Figure shows the topics covered in the book. The solid lines depict prerequisites. Often the prerequisite structure does not include all sub-topics. Given the medium of a book, we have had to linearize the topics. However, the book is designed so the topics are teachable in any order satisfying the prerequisite structure.
The references given at the end of each chapter are not meant to be comprehensive; we have referenced works that we have directly used and works that we think provide good overviews of the literature, by referencing both classic works and more recent surveys. We hope that no researchers feel slighted by their omission, and we are happy to have feedback where someone feels that an idea has been misattributed. Remember that this book is not a survey of AI research.
Thanks to Saleema Amershi, Yoshua Bengio, Giuseppe Carenini, Jeff Clune, Mel Comisarow, Cristina Conati, Nando de Freitas, Rina Dechter, Bahare Fatemi, Randy Goebel, Robert Goldman, Jesse Hoey, Robert Holte, Christopher Mole, Kevin Murphy, Sriraam Natarajan, Alex Poole, Francesca Rossi, Justice Sefas, Ben Shneiderman, Peter van Beek, Geoffrey Woollard, and the anonymous reviewers for valuable feedback on this third edition.
Thanks to our editors, Lauren Cowles, Maggie Jeffers, Julie Lancashire, and Stefanie Seaton, our content manager, Rachel Norridge, and the staff at Cambridge University Press for all their support, encouragement, and help.
All remaining mistakes are ours.
We invite you to join us in an intellectual adventure: building a science of intelligent agents.