What do you think?
Rate this book
Artificial Intelligence: A Modern Approach
For one or two-semester, undergraduate or graduate-level courses in Artificial Intelligence. The long-anticipated revision of this best-selling text offers the most comprehensive, up-to-date introduction to the theory and practice of artificial intelligence. *NEW-Nontechnical learning material-Accompanies each part of the book. *NEW-The Internet as a sample application for intelligent systems-Added in several places including logical agents, planning, and natural language. *NEW-Increased coverage of material - Includes expanded coverage of: default reasoning and truth maintenance systems, including multi-agent/distributed AI and game theory; probabilistic approaches to learning including EM; more detailed descriptions of probabilistic inference algorithms. *NEW-Updated and expanded exercises-75% of the exercises are revised, with 100 new exercises. *NEW-On-line Java software. *Makes it easy for students to do projects on the web using intelligent agents. *A unified, agent-based approach to AI-Organizes the material around the task of building intelligent agents. *Comprehensive, up-to-date coverage-Includes a unified view of the field organized around the rational decision making pa
1080 pages, Hardcover
First published December 13, 1994
About the author
Stuart Russell
28 books223 followersLibrarian Note: There is more than one author in the GoodReads database with this name. See this thread for more information.
Ratings & Reviews
Friends & Following
Create a free account to discover what your friends think of this book!
Community Reviews
Displaying 1 - 30 of 167 reviews
January 26, 2015
This monumental work, which completely dominates the AI textbook market, has been compared with classics like Watson's Molecular Biology of the Cell, and eminently succeeds in its goal of providing a clear, single-volume summary of the whole field of Artificial Intelligence. As pointed out on the book's home page, it is used in over 1200 universities in over 100 countries, and is the 25th most cited publication on Citeseer and the 2nd most cited publication of this century. The occasional suggestion you may hear that it "has passed its sell-by" or "gives a decent picture of Good Old-Fashioned AI" can unhesitatingly be written off as envious carping from academics who wish they'd got something even a tenth as impressive on their CVs.
What was that? Ah, yes, as a matter of fact it does cite one of my papers. How did you guess?
What was that? Ah, yes, as a matter of fact it does cite one of my papers. How did you guess?
November 28, 2008
Heh, I opened this up to find the ISBN and found dried blood all over the pages, suggesting I read this during my cocaine-intensive period back in 1999-2000. That's fitting, since cocaine and the study of artificial intelligence seem to enjoy several similarities -- incredible expense as a barrier to entry, exciting short-term effects (see: euphoria, A* search) but letdowns upon prolonged use (see: addiction, combinatorial explosions), and they've both ruined plenty of fine careers in computer science. We used this book for CS4600, but I only got halfway though that semester and remember little of it (see: careers in computer science, aforementioned negative effects of cocaine on). I went back and read most of this in 2003, and found solid coverage of most everything useful I'm aware of from AI.
October 5, 2013
Wants a book that explains broad and deep AI yet in laymen term (nearly)? This is IT. Of all the AI books I have read, this one is arguably the most accessible to undergrads (CS, EE background) It assumes only minimal mathematical formalities and pretty much the maths things are self-contained. The authors did a great job of keeping the contents up-to-date with the latest happenings in AI, while keeping the readers sane. Overall, thumbs up!
May 1, 2015
Holy balls this book has a lot of pages. I also don't know why these things always have to have separate ``international'' editions.
It starts off strongly for a few hundred pages, but then for no reason at all devotes several chapters to high school-level probability and statistics, before devolving into essentially pointless mathematical show-boating for another few hundred pages. Then it finishes off with an interesting but not really relevant and highly unrigorous (not to mention typo-ridden) overview of Google's various products (mostly PageRank and Google Translate).
There's a few more chapters after that, but I think it's best to pretend they don't exist. Chapter 26 (Philosophical Foundations), in particular, was a fucking embarrassment, giving more unnecessary to idiots like John Searle and Ray Kurzweil, and wasting paper on absurd hand-wringing over off-the-wall science-fiction scenarios. AI is too legitimate and interesting a field to justify that sort of crap in a university textbook.
In spite of all that, though, it's still a very good book, and a good overview of the field. I particularly liked that each chapter had an extensive section with historical and biographical notes at the end. If nothing else, it at least demonstrates that if the AI winter was ever a real thing (at least in terms of research activity and progress), it's far behind us now.
It starts off strongly for a few hundred pages, but then for no reason at all devotes several chapters to high school-level probability and statistics, before devolving into essentially pointless mathematical show-boating for another few hundred pages. Then it finishes off with an interesting but not really relevant and highly unrigorous (not to mention typo-ridden) overview of Google's various products (mostly PageRank and Google Translate).
There's a few more chapters after that, but I think it's best to pretend they don't exist. Chapter 26 (Philosophical Foundations), in particular, was a fucking embarrassment, giving more unnecessary to idiots like John Searle and Ray Kurzweil, and wasting paper on absurd hand-wringing over off-the-wall science-fiction scenarios. AI is too legitimate and interesting a field to justify that sort of crap in a university textbook.
In spite of all that, though, it's still a very good book, and a good overview of the field. I particularly liked that each chapter had an extensive section with historical and biographical notes at the end. If nothing else, it at least demonstrates that if the AI winter was ever a real thing (at least in terms of research activity and progress), it's far behind us now.
April 8, 2013
5 stars because there is, quite simply, no substitute.
Artificial Intelligence is, in the context of the infant science of computing, a very old and very broad subdiscipline, the "Turing test" having arisen, not only at the same time, but from the same person as many of the foundations of computing itself. Those of us students of a certain age will recall terms like "symbolic" vs. "connectionist" vs. "probabilistic," as well as "scruffies" and "neats." Key figures, events, and schools of thought span multiple institutions on multiple continents. In short, a major challenge facing anyone wishing to survey Artificial Intelligence is simply coming up with a unifying theme.
The major accomplishment, in my opinion, of AIMA, then, is that: Russell and Norvig take the hodge-podge of AI research, manage to fit it sensibly into a narrative structure centered on the notion of different kinds of "agents" (not to be confused with that portion of AI research that explicitly refers to its constructs as "agents!") and, having dug the pond and filled it with water, skip a stone across the surface. It's up to the reader whether to follow the arcs of the stone from major subject to major subject, foregoing depth, or whether to pick a particular contact point and concentrate on the eddies propagating from it. For the latter purpose, the extensive bibliography is indispensable.
With all of this said, I have to acknowledge that Russell and Norvig are not entirely impartial AI practitioners. Norvig, in particular, is well-known by now as a staunch Bayesian probabilist who, as Director of Search Quality or Machine Learning or whatever Google has decided to call it today, has made Google the Bayesian powerhouse that it is. (Less known is Norvig's previous stint at high-tech startup Junglee, which was acquired by Amazon. So to some extent Peter Norvig powers both Google and Amazon.) So one can probably claim, not without justification, that AIMA emphasizes Bayesian probability over other approaches.
Finally, as good as AIMA is, it is still a survey. Even with respect to Bayesian probability, the treatment is introductory, as I discovered with some shock upon reading Probability Theory: The Logic of Science. That's OK, though: it's the best introduction I've ever seen.
So read it once for the survey, keep it on your shelf for the bibliography, and refer back to it whenever you find yourself thinking "hey, didn't I read about that somewhere before?"
Artificial Intelligence is, in the context of the infant science of computing, a very old and very broad subdiscipline, the "Turing test" having arisen, not only at the same time, but from the same person as many of the foundations of computing itself. Those of us students of a certain age will recall terms like "symbolic" vs. "connectionist" vs. "probabilistic," as well as "scruffies" and "neats." Key figures, events, and schools of thought span multiple institutions on multiple continents. In short, a major challenge facing anyone wishing to survey Artificial Intelligence is simply coming up with a unifying theme.
The major accomplishment, in my opinion, of AIMA, then, is that: Russell and Norvig take the hodge-podge of AI research, manage to fit it sensibly into a narrative structure centered on the notion of different kinds of "agents" (not to be confused with that portion of AI research that explicitly refers to its constructs as "agents!") and, having dug the pond and filled it with water, skip a stone across the surface. It's up to the reader whether to follow the arcs of the stone from major subject to major subject, foregoing depth, or whether to pick a particular contact point and concentrate on the eddies propagating from it. For the latter purpose, the extensive bibliography is indispensable.
With all of this said, I have to acknowledge that Russell and Norvig are not entirely impartial AI practitioners. Norvig, in particular, is well-known by now as a staunch Bayesian probabilist who, as Director of Search Quality or Machine Learning or whatever Google has decided to call it today, has made Google the Bayesian powerhouse that it is. (Less known is Norvig's previous stint at high-tech startup Junglee, which was acquired by Amazon. So to some extent Peter Norvig powers both Google and Amazon.) So one can probably claim, not without justification, that AIMA emphasizes Bayesian probability over other approaches.
Finally, as good as AIMA is, it is still a survey. Even with respect to Bayesian probability, the treatment is introductory, as I discovered with some shock upon reading Probability Theory: The Logic of Science. That's OK, though: it's the best introduction I've ever seen.
So read it once for the survey, keep it on your shelf for the bibliography, and refer back to it whenever you find yourself thinking "hey, didn't I read about that somewhere before?"
June 7, 2018
The Bible on computational decision-making. I use this term as this book is not just about the AI/machine learning we consistently hear about, it’s much more. This textbook tends to perfection, with no stone left unturned. Looking forward to the next edition, which, at the accelerating rate of innovation, looks overdue (the following sentence surely feels outdated: “Current Go programs play at the master level on a reduced 9 × 9 board, but are still at advanced amateur level on a full board”). There are 2 aspects I particularly enjoyed, (1) the historical sections at the end of each chapter; the introduction also gave a fascinating history of AI and its relationship to other fields (neurology, logics, cybernetics…). (2) I also liked all the gaming aspects, such as the Wumpus World which I didn't know before. I truly wish I had discovered that book when it was first published in 1995, sigh
August 6, 2008
For a textbook, this is amazingly accessible and interesting. if you have any interest in the topic, this is the book to read. It's $100 or more, but it's very popular for AI classes, so any good college library should have a copy.
January 10, 2024
3.5 stars.
For me it was a really good book to supplement my AI course at university. I only read the beginning, end and certain topics in the book, so not all of it. I would say the book is good for learning if you have an additional course still. As a beginner with no information about theoretical computer science, I would say it's a bit difficult. I mean, you can do it, but you need to set aside a whole year (or longer if you need to read the math background - yes, you need that) to do it ... and then you have a really good foundation, but it might be a bit outdated after all, because research in AI is very fast at the moment.
I've decided to finish this book and won't be reading the remaining chapters as I need (and want) to read more important books for my career. I would say if you are interested in this topic and want to do something in this field - go for it! You won't find a better book for the basics. But if you're just interested in it and don't really need to work with it - I'd say you should read other books on the subject.
Fun Fact: Also, I wrote my AI exam today and think this is a good time to finish the book!
For me it was a really good book to supplement my AI course at university. I only read the beginning, end and certain topics in the book, so not all of it. I would say the book is good for learning if you have an additional course still. As a beginner with no information about theoretical computer science, I would say it's a bit difficult. I mean, you can do it, but you need to set aside a whole year (or longer if you need to read the math background - yes, you need that) to do it ... and then you have a really good foundation, but it might be a bit outdated after all, because research in AI is very fast at the moment.
I've decided to finish this book and won't be reading the remaining chapters as I need (and want) to read more important books for my career. I would say if you are interested in this topic and want to do something in this field - go for it! You won't find a better book for the basics. But if you're just interested in it and don't really need to work with it - I'd say you should read other books on the subject.
Fun Fact: Also, I wrote my AI exam today and think this is a good time to finish the book!
December 25, 2014
يعتبر هذا الكتاب أهم مرجع للدارس في مجال الذكاء الاصطناعي. الكتاب يعتبر مقدمة لمواضيع كبيرة جدا و متشعبة، فعيتبر بداية التخصص في الذكاء الاصطناعي. يتناول الكتاب مواضيع في تعلم الآلة وخوارزميات البحث وحل المشكلات. بعض أجزاء الكتاب تعبر من المراجع النادرة وخصوصا في فيما يتعلق بموضع الـ reinforcement learning.
مؤلفي الكتاب بيتر نورفق و روسيل من الرواد في مجال الذكاء الاصطناعي.
مؤلفي الكتاب بيتر نورفق و روسيل من الرواد في مجال الذكاء الاصطناعي.
May 16, 2015
A fantastic textbook that's not only a great introduction to AI but also serves as a survey course in technical writing. I only read about 75% of it but definitely plan on revisiting it. Re-reading some earlier chapters taught me how much I missed on a first read (or forgot).
AIMA doesn't presume a ton of background beyond some programming experience, exposure to mathematical notation, and a basic understanding of computational complexity/algorithmic efficiency.
The first 10 chapters or so are the best and the second half of the book can be a bit of a trudge as it devolves into mathematical masturbation. A lot of the chapters are better served by other resources – I highly recommend the CS188 lectures from UC Berkeley for supplementation. Unfortunately, some chapters are straight up bad (the chapter on Philosophical Foundations comes to mind), but these tend to be few and far between.
Despite that, there is no more comprehensive book on AI. Read this, re-read this, and treat it with care – you will reap the rewards for a long time to come.
AIMA doesn't presume a ton of background beyond some programming experience, exposure to mathematical notation, and a basic understanding of computational complexity/algorithmic efficiency.
The first 10 chapters or so are the best and the second half of the book can be a bit of a trudge as it devolves into mathematical masturbation. A lot of the chapters are better served by other resources – I highly recommend the CS188 lectures from UC Berkeley for supplementation. Unfortunately, some chapters are straight up bad (the chapter on Philosophical Foundations comes to mind), but these tend to be few and far between.
Despite that, there is no more comprehensive book on AI. Read this, re-read this, and treat it with care – you will reap the rewards for a long time to come.
March 18, 2023
Best textbook on artificial intelligence. Period. Stuart Russell's clear and comprehensive approach to AI has made this book a must-read for anyone interested in the field.
March 4, 2022
I have read only two chapters out of total seven (which was the plan) and the 3rd edition is already out of date regarding most areas of research (especially deep learning). It is a great book nonetheless.
September 2, 2013
OK so I did not read this cover to cover, but I did look closely at much of what you might call the foundational chapters, just to see 1. is there such a thing as AI, or are we just hoping there will be and 2. what can I learn as a philosopher from AI, whether it exists or not. Goal 2 was much more important as I teach a logic of induction class and of course one major pillar of AI would be developing machines that can perform judgments under uncertainty and apply rational heuristics as well as humans do (which is not very well at all by the way). I found out that I already knew most of this, from studies of Bayesian reasoning (which is very tricky by the way and should not be blindly implemented like this without a clear view of the limitations), and the study of acyclic causal graphs (which is standard academy reading for philosophers). These graphs also admit of howlers and counterexamples as anyone knows. I am more interested in the idea of developing "stupid machines" that function more like neural networks and less like probability maximizers. The human brain is fundamentally (in my view anyway) a stupid-machine, full of crazy workarounds and faulty logic. The correct solution or path is virtually never the one evolution comes up with, it just grinds it out with massive armies of neurons and interconnections and lots of trial and error, but nothing one would call a computation, as in Turing machines. Elegant algorithms for computer vision have, I believe, nothing to do with the way the brain constructs the visual image. One philosopher's take.
December 17, 2007
A comprehensive course in modern AI topics. While the book is dense with information, the authors provide clear explanations that will be easily picked up by the careful reader. An excellent companion to an undergraduate course in artificial intelligence.
March 7, 2011
It was written more like a text book for undergrads with extensive coverage of many topics. However, I was looking for more in-depth information on knowledge representation. But, it was too superficial for my need. May be, in 3rd edition it encompassed the latest ideas in this area.
December 23, 2020
کتاب کمی گنگ و قدیمی هست و معمولا کتب جدید هوش مصنوعی هم درک بهتری به موضوع میدن و هم میدان معرفی بزرگتر و سادگی بالاتری را شامل می شوند.
در این کتاب بیشتر به حل مسائلی چون مسیریابی و الگوریتم های اینچنینی پرداخته شده
در این کتاب بیشتر به حل مسائلی چون مسیریابی و الگوریتم های اینچنینی پرداخته شده
Read
February 7, 2023Read the 2nd edition for an AI course back in 2005 or so. It's interesting to look back at the table of contents and see how much has changed since then about what gets considered "AI".
July 12, 2020
Considering my previous knowledge of Artificial Intelligence, this book was shite.
Constant repetition of previously learnt algorithms and those hideous chapters on logic made me want to puke. The humor was good and the writer came off as friendly, but Jesus was this a waste of time.
Constant repetition of previously learnt algorithms and those hideous chapters on logic made me want to puke. The humor was good and the writer came off as friendly, but Jesus was this a waste of time.
September 22, 2019
Read the some of the parts relevant to my AI course. The standard book in the field of traditional AI techniques, great historical information/case studies in each chapter, but a lot of modern research in AI (deep learning) is based on very different principles, which are mentioned in, but not the focus of this book. Still would call it mandatory reading for anyone who wants to work in AI.
January 5, 2012
This is THE book to read on anything to do with modern artificial intelligence. I regard this as my personal bible and would recommend it to anyone who is involved in technical artificial intelligence.
July 16, 2016
من المراجع النادرة في الكلية اللي لقيت عقلي قادر على استيعاب كلامها :D
أسلوب بسيط وممتع كمان، للي مهتم بالمجال أو مش مهتم بس مضطر يقرأ فيه
في الحالتين لطيف
المشكلة الوحيدة طوله الرهيب، أتمنى ألحق أخلص أكبر قدر ممكن قبل الامتحان
أسلوب بسيط وممتع كمان، للي مهتم بالمجال أو مش مهتم بس مضطر يقرأ فيه
في الحالتين لطيف
المشكلة الوحيدة طوله الرهيب، أتمنى ألحق أخلص أكبر قدر ممكن قبل الامتحان
January 14, 2019
Great overview over such a big and complex field such as Artificial Intelligence
September 20, 2018
TL;DR: Um excelente livro para quem quer estudar fundamentos de IA, recomendo.
O livro é bastante teórico. Li de forma despretensiosa, sem me preocupar com os exercícios, por exemplo. Meu objetivo, desde o início, foi ganhar uma base para continuar os estudos com o curso de Machine Learning de Stanford no Coursera [1] ou com um livro prático de IA aplicada [2].
Várias definições são detalhadamente explicadas no livro, diversos conceitos como:
- conhecimento (representação do conhecimento, incerteza)
- inteligência e aprendizado (como forma de melhorar o desempenho)
* aprendizagem indutiva (supervisionada ou não supervisionada)
* aprendizagem por reforço (baseado em sucesso ou fracasso, recompensa ou penalidade)
* redes neurais artificiais (uma das formas mais populares e eficazes de aprendizagem)
- ambiente
* determinístico (c/ probabilidade, próximo estado = estado anterior + ação)
* estocástico (não determinístico, s/ probabilidade)
* totalmente, parcialmente ou não observável
* episódico ou sequencial
* estático ou dinâmico
* discreto ou contínuo
- agentes (racionais, reativos, baseados em modelo, em objetivo, em utilidade)
- algoritmos (de busca, de caminho mais curto, genéticos, etc)
- comunicação
* processamento de linguagem natural
* classificação, busca e extração de informação
* tradução (que lida com sintaxe, gramática, etc)
* reconhecimento de voz
* processamento de imagem
- robótica (agentes físicos e seus desafios ligados a movimentação, equilíbrio, etc)
- e muitos outros tópicos
Os últimos capítulos do livros são menos técnicos, mais filosóficos, bem interessantes também. Nesse ponto ele fala sobre o impacto da IA na vida das pessoas, por exemplo, sobre a possibilidade de criarmos uma inteligência superior à nossa e inteligente o suficiente para criar de fato uma "ultrainteligência", o que foi chamado de "explosão de inteligência" ou de "singularidade tecnológica". Nessa parte, existem considerações tanto sobre os riscos desses desdobramentos, quanto sobre os benefícios (que já temos hoje, inclusive). Ele descreve que até hoje os programas de IA criaram mais empregos do que eliminaram, criaram empregos mais interessantes e melhores remunerados. Coloca-se também que a sociedade moderna se tornou dependente de computadores em geral e algumas áreas são simplesmente inviáveis apenas com o trabalho humano.
No geral fiquei bastante satisfeito com a leitura do livro, em vários momentos coisas que vi no documentário AlphaGo [3] passaram a fazer mais sentido pra mim :-D
[1] - https://www.coursera.org/learn/machin...
[2] - https://www.goodreads.com/book/show/8...
[3] - https://www.netflix.com/title/80190844
O livro é bastante teórico. Li de forma despretensiosa, sem me preocupar com os exercícios, por exemplo. Meu objetivo, desde o início, foi ganhar uma base para continuar os estudos com o curso de Machine Learning de Stanford no Coursera [1] ou com um livro prático de IA aplicada [2].
Várias definições são detalhadamente explicadas no livro, diversos conceitos como:
- conhecimento (representação do conhecimento, incerteza)
- inteligência e aprendizado (como forma de melhorar o desempenho)
* aprendizagem indutiva (supervisionada ou não supervisionada)
* aprendizagem por reforço (baseado em sucesso ou fracasso, recompensa ou penalidade)
* redes neurais artificiais (uma das formas mais populares e eficazes de aprendizagem)
- ambiente
* determinístico (c/ probabilidade, próximo estado = estado anterior + ação)
* estocástico (não determinístico, s/ probabilidade)
* totalmente, parcialmente ou não observável
* episódico ou sequencial
* estático ou dinâmico
* discreto ou contínuo
- agentes (racionais, reativos, baseados em modelo, em objetivo, em utilidade)
- algoritmos (de busca, de caminho mais curto, genéticos, etc)
- comunicação
* processamento de linguagem natural
* classificação, busca e extração de informação
* tradução (que lida com sintaxe, gramática, etc)
* reconhecimento de voz
* processamento de imagem
- robótica (agentes físicos e seus desafios ligados a movimentação, equilíbrio, etc)
- e muitos outros tópicos
Os últimos capítulos do livros são menos técnicos, mais filosóficos, bem interessantes também. Nesse ponto ele fala sobre o impacto da IA na vida das pessoas, por exemplo, sobre a possibilidade de criarmos uma inteligência superior à nossa e inteligente o suficiente para criar de fato uma "ultrainteligência", o que foi chamado de "explosão de inteligência" ou de "singularidade tecnológica". Nessa parte, existem considerações tanto sobre os riscos desses desdobramentos, quanto sobre os benefícios (que já temos hoje, inclusive). Ele descreve que até hoje os programas de IA criaram mais empregos do que eliminaram, criaram empregos mais interessantes e melhores remunerados. Coloca-se também que a sociedade moderna se tornou dependente de computadores em geral e algumas áreas são simplesmente inviáveis apenas com o trabalho humano.
No geral fiquei bastante satisfeito com a leitura do livro, em vários momentos coisas que vi no documentário AlphaGo [3] passaram a fazer mais sentido pra mim :-D
[1] - https://www.coursera.org/learn/machin...
[2] - https://www.goodreads.com/book/show/8...
[3] - https://www.netflix.com/title/80190844
August 23, 2020
I had to read this as part of my Artificial Intelligence course in Georgia Tech's Online Master's in Computer Science program, and as an AI textbook it was excellent. It provides detailed, and easy to follow, algorithms ranging from minimax and alpha-beta to Bayes Nets, Hidden Markov Models, A*, Neural Nets, and plenty more. I did not read every page of this book, but I can attest that I would not have done nearly as well in my course without it and if I need to look up an AI algo, I'll turn here first to read what Russell and Norvig have to say first and then check other resources.
January 23, 2020
Mind-blowing text diving deep into topics with extremely clear explanations. Straddles the line better than any other textbook I think I've ever read. Check out my implementations here: https://github.com/WarrenGreen/AI-Norvig
March 2, 2018
not bad for an intro: simple math, minimal jargon and pretty organized
June 6, 2022
The ego the professional bureaucrats can have: this is ”a modern approach”. Hurray to Russell! After all those centuries of Universities using the Arabic Medieval Approach to Artificial Intelligence!
May 19, 2023
so proud that i acc read it from cover to cover (literally the only book i read from my masters reading list, and im still recovering from the trauma)
January 11, 2024
500-Artificial Intelligence-Stuart Russell- AI-1994
Barack
2023/12/31
"Artificial Intelligence", was first published in 1994. It is a college textbook on artificial intelligence. It has been called "the world's most popular artificial intelligence textbook" and is considered a standard text in the field of artificial intelligence. As of 2023, it is used in more than 1,500 universities worldwide and has more than 59,000 citations on Google Scholar. This book is intended for undergraduate readers, but can also be used for graduate-level research. The programs in the book are presented in pseudocode and can be used online Implemented in Java, Python, Lisp, JavaScript, and Scala.
Stuart Russell was born in Portsmouth, England in 1962. He studied at the University of Oxford (BA) and Stanford University (PhD). He is a British computer scientist known for his contributions to artificial intelligence (AI). He is a professor of computer science at the University of California, Berkeley, and an adjunct professor of neurosurgery at the University of California, San Francisco, from 2008 to 2011. He is the Smith-Zadeh Chair in Engineering at the University of California, Berkeley. He founded and directed the Center for Human-Compatible Artificial Intelligence (CHAI) at UC Berkeley.
Table of Contents
I Artificial Intelligence
1 Introduction
2 Intelligent Agents
II Problem-solving
3 Solving Problems by Searching
4 Search in Complex Environments
5 Constraint Satisfaction Problems
6 Adversarial Search and Games
III Knowledge, reasoning, and planning
7 Logical Agents
8 First-Order Logic
9 Inference in First-Order Logic
10 Knowledge Representation
11 Automated Planning
IV Uncertain knowledge and reasoning
12 Quantifying Uncertainty
13 Probabilistic Reasoning
14 Probabilistic Reasoning over Time
15 Making Simple Decisions
16 Making Complex Decisions
17 Multiagent Decision Making
18 Probabilistic Programming
V Machine Learning
19 Learning from Examples
20 Knowledge in Learning
21 Learning Probabilistic Models
22 Deep Learning
23 Reinforcement Learning
VI Communicating, perceiving, and acting
24 Natural Language Processing
25 Deep Learning for Natural Language Processing
26 Robotics
27 Computer Vision
VII Conclusions
28 Philosophy, Ethics, and Safety of AI
29 The Future of AI
When we think about how to make artificial intelligence, we should first clarify what “artificial intelligence” we are talking about. In the book, the author proposes two main dimensions to define artificial intelligence. The first dimension focuses on human behavior. From this perspective, artificial intelligence can be divided into two subcategories: one is intelligence that imitates human behavior, and the other is intelligence that can generate rationality. The second dimension is defined in terms of expression. According to this dimension, two subcategories can also be obtained, namely, intelligent thinking processes and intelligent behavior patterns. Combining these two dimensions, we can derive four different types of artificial intelligence: The first is intelligence that behaves like humans, that is, intelligence that passes the Turing test or the complete Turing test, which is designed to determine whether a machine can Through language or behavior, humans cannot tell the difference between them and real people; the second type is similar to humans in the thinking process, and this type of intelligence exhibits a human-like cognitive model; the third type is rule-based rationality Thinking, this type of intelligent agent thinks through logical rules; the last type is the display of rational behavior, this type of intelligent agent makes logical and rational responses in the physical environment based on external conditions. These four criteria not only help us understand different types of artificial intelligence but may also provide a reference for us to evaluate whether a product is an artificial intelligence. For example, AlphaGo in 2016 and ChatGPT in 2023 seem to focus more on rational thinking processes rather than directly imitating human behavior or thinking processes.
When discussing Intelligence, we further thought about what an agent is. The authors' definition of agency contains three key elements. First, the agent is able to receive environmental signals through sensors. Second, agents are able to influence the environment through actuators. Finally, there is a process of information processing between the agent receiving the signal and reacting. Taking humans as an example, we receive external information through our senses such as eyes and ears, and then respond through body parts such as hands and feet. The robot is similar. It may receive external information through a camera and then decide its movement direction. Even a static device like a computer, which cannot physically move, can respond to input signals by printing files or displaying images. The key here is that different agents may produce different outputs for the same input. In fact, the same is true for human beings. Under the same environment, different people have different reactions. This is the difference between the so-called "wise men" and "fools". It is the “intelligence” component of the agent that determines this difference. In other words, how an agent processes and interprets signals determines the performance and effect of its intelligence.
We first consider the simplest deterministic problem, such as the "Eight Queens Problem". The agent's processing flow can be roughly divided into four steps. The first is the formatting of the goal (Goal Formation). This step is similar to what we do when solving word problems, which is to abstract the problem into a clear mathematical problem, no matter how complex the daily life scenario it describes. The second step is Problem Formation. In this step, it is crucial to define the "status" of the problem and the possible "actions". The "status" may refer to the current node and its relationship with other nodes, and the "action" refers to deciding which node to go to next. The third step is Search. This usually involves an exhaustive or partially exhaustive approach, comparing the results of different actions to find the optimal solution. This is why this solution is mainly aimed at deterministic problems since we are searching almost the entire state space. The essence of a solution is a sequence of actions to achieve a goal. The fourth step is execution. In a computer program, this might mean displaying the final execution path or outputting the solution on the screen. Relatively speaking, implementation is usually a relatively straightforward and simple process. This four-step strategy is mainly suitable for deterministic problems, and these problems often have a clear optimal solution. However, in non-deterministic environments, the actions an agent should take may depend on further signals received from the environment. In this case, relying solely on the four steps above may not lead to the best decision.
Having considered in depth the steps for solving problems in deterministic environments, the next challenge we face is how to solve more complex problems. The complexity of a problem often stems from the complexity of the environment, which often means we need to think in an environment of uncertainty. In this environment, we find it difficult or unwilling to exhaust all possibilities at a high cost. A commonly used strategy when dealing with this type of problem is the so-called "greedy algorithm". In the greedy algorithm, we only consider the optimal next action every time we make a decision, and continue to find the local optimal solution. The intuitive feeling of this method is that it is a typical short-sighted behavior, and this short-sighted behavior relies heavily on luck. Especially when the environment is complex and full of multiple peaks and troughs, the local optimal solution may only allow us to reach one of the peaks, and this peak may be insignificant compared to the global optimal peak. Sometimes we may even hit a plateau where the outcome seems to be the same no matter what we choose, leaving us stuck. Randomness is something that relies on luck, and the way to improve overall luck is to increase the number of random times. After all, what we care about is only the best one among all the results of exploration. Therefore, faced with such a challenge, a common solution is "random restart", that is, starting the search at multiple different starting points to increase the possibility of finding the global optimal solution.
In the previous discussion, we explored ways to treat the state space as independent, atomic nodes when looking for problem-solving algorithms. For example, in the game of chess, moving the position of each chess piece by one square is a typical atomic operation. However, if we consider the state space in more detail, not treating it as an atomic operation, but abstracting the problem into Constraint Satisfaction Problems (CSP), we open up a new way of thinking. There are three main reasons for choosing to abstract practical problems into CSP: First, many problems in real life can be effectively abstracted into CSP problems. This abstraction makes the solution to the problem more universal, just like when we do mathematical applications When asking a question, the first thing to think about is, what is the essence of this problem? Is it a sequence problem, a plane geometry problem, or an algebra problem? Secondly, research in the past few decades has accumulated a large number of effective methods to solve CSP problems. If we can formulate a unique problem into a CSP problem, it is possible to find mature solutions from the existing tool library without having to invent new methods from scratch or design separate algorithms for each problem. Finally, the CSP method can greatly compress the search space. For example, in the map coloring problem, once the color of a certain area is determined, the colors of other adjacent areas are also determined, thus greatly reducing the search space. Perhaps when we look back in 10 years, we will be relatively confident that ChatGPT, which will cause a sensation in the industry in 2023, will be regarded as an important milestone in the development of artificial intelligence and even the entire technology field. One of the reasons for its popularity may be because of its versatility. Personal computers, the Internet, and mobile Internet, these solutions are extremely versatile. Therefore, when we face endless problems, it is an efficient method to abstractly summarize different problems into limited types and use mature tools in the tool library to solve them. If each problem is analyzed individually, not only is it inefficient, but the effectiveness of the solution may also be limited.
Barack
2023/12/31
"Artificial Intelligence", was first published in 1994. It is a college textbook on artificial intelligence. It has been called "the world's most popular artificial intelligence textbook" and is considered a standard text in the field of artificial intelligence. As of 2023, it is used in more than 1,500 universities worldwide and has more than 59,000 citations on Google Scholar. This book is intended for undergraduate readers, but can also be used for graduate-level research. The programs in the book are presented in pseudocode and can be used online Implemented in Java, Python, Lisp, JavaScript, and Scala.
Stuart Russell was born in Portsmouth, England in 1962. He studied at the University of Oxford (BA) and Stanford University (PhD). He is a British computer scientist known for his contributions to artificial intelligence (AI). He is a professor of computer science at the University of California, Berkeley, and an adjunct professor of neurosurgery at the University of California, San Francisco, from 2008 to 2011. He is the Smith-Zadeh Chair in Engineering at the University of California, Berkeley. He founded and directed the Center for Human-Compatible Artificial Intelligence (CHAI) at UC Berkeley.
Table of Contents
I Artificial Intelligence
1 Introduction
2 Intelligent Agents
II Problem-solving
3 Solving Problems by Searching
4 Search in Complex Environments
5 Constraint Satisfaction Problems
6 Adversarial Search and Games
III Knowledge, reasoning, and planning
7 Logical Agents
8 First-Order Logic
9 Inference in First-Order Logic
10 Knowledge Representation
11 Automated Planning
IV Uncertain knowledge and reasoning
12 Quantifying Uncertainty
13 Probabilistic Reasoning
14 Probabilistic Reasoning over Time
15 Making Simple Decisions
16 Making Complex Decisions
17 Multiagent Decision Making
18 Probabilistic Programming
V Machine Learning
19 Learning from Examples
20 Knowledge in Learning
21 Learning Probabilistic Models
22 Deep Learning
23 Reinforcement Learning
VI Communicating, perceiving, and acting
24 Natural Language Processing
25 Deep Learning for Natural Language Processing
26 Robotics
27 Computer Vision
VII Conclusions
28 Philosophy, Ethics, and Safety of AI
29 The Future of AI
When we think about how to make artificial intelligence, we should first clarify what “artificial intelligence” we are talking about. In the book, the author proposes two main dimensions to define artificial intelligence. The first dimension focuses on human behavior. From this perspective, artificial intelligence can be divided into two subcategories: one is intelligence that imitates human behavior, and the other is intelligence that can generate rationality. The second dimension is defined in terms of expression. According to this dimension, two subcategories can also be obtained, namely, intelligent thinking processes and intelligent behavior patterns. Combining these two dimensions, we can derive four different types of artificial intelligence: The first is intelligence that behaves like humans, that is, intelligence that passes the Turing test or the complete Turing test, which is designed to determine whether a machine can Through language or behavior, humans cannot tell the difference between them and real people; the second type is similar to humans in the thinking process, and this type of intelligence exhibits a human-like cognitive model; the third type is rule-based rationality Thinking, this type of intelligent agent thinks through logical rules; the last type is the display of rational behavior, this type of intelligent agent makes logical and rational responses in the physical environment based on external conditions. These four criteria not only help us understand different types of artificial intelligence but may also provide a reference for us to evaluate whether a product is an artificial intelligence. For example, AlphaGo in 2016 and ChatGPT in 2023 seem to focus more on rational thinking processes rather than directly imitating human behavior or thinking processes.
When discussing Intelligence, we further thought about what an agent is. The authors' definition of agency contains three key elements. First, the agent is able to receive environmental signals through sensors. Second, agents are able to influence the environment through actuators. Finally, there is a process of information processing between the agent receiving the signal and reacting. Taking humans as an example, we receive external information through our senses such as eyes and ears, and then respond through body parts such as hands and feet. The robot is similar. It may receive external information through a camera and then decide its movement direction. Even a static device like a computer, which cannot physically move, can respond to input signals by printing files or displaying images. The key here is that different agents may produce different outputs for the same input. In fact, the same is true for human beings. Under the same environment, different people have different reactions. This is the difference between the so-called "wise men" and "fools". It is the “intelligence” component of the agent that determines this difference. In other words, how an agent processes and interprets signals determines the performance and effect of its intelligence.
We first consider the simplest deterministic problem, such as the "Eight Queens Problem". The agent's processing flow can be roughly divided into four steps. The first is the formatting of the goal (Goal Formation). This step is similar to what we do when solving word problems, which is to abstract the problem into a clear mathematical problem, no matter how complex the daily life scenario it describes. The second step is Problem Formation. In this step, it is crucial to define the "status" of the problem and the possible "actions". The "status" may refer to the current node and its relationship with other nodes, and the "action" refers to deciding which node to go to next. The third step is Search. This usually involves an exhaustive or partially exhaustive approach, comparing the results of different actions to find the optimal solution. This is why this solution is mainly aimed at deterministic problems since we are searching almost the entire state space. The essence of a solution is a sequence of actions to achieve a goal. The fourth step is execution. In a computer program, this might mean displaying the final execution path or outputting the solution on the screen. Relatively speaking, implementation is usually a relatively straightforward and simple process. This four-step strategy is mainly suitable for deterministic problems, and these problems often have a clear optimal solution. However, in non-deterministic environments, the actions an agent should take may depend on further signals received from the environment. In this case, relying solely on the four steps above may not lead to the best decision.
Having considered in depth the steps for solving problems in deterministic environments, the next challenge we face is how to solve more complex problems. The complexity of a problem often stems from the complexity of the environment, which often means we need to think in an environment of uncertainty. In this environment, we find it difficult or unwilling to exhaust all possibilities at a high cost. A commonly used strategy when dealing with this type of problem is the so-called "greedy algorithm". In the greedy algorithm, we only consider the optimal next action every time we make a decision, and continue to find the local optimal solution. The intuitive feeling of this method is that it is a typical short-sighted behavior, and this short-sighted behavior relies heavily on luck. Especially when the environment is complex and full of multiple peaks and troughs, the local optimal solution may only allow us to reach one of the peaks, and this peak may be insignificant compared to the global optimal peak. Sometimes we may even hit a plateau where the outcome seems to be the same no matter what we choose, leaving us stuck. Randomness is something that relies on luck, and the way to improve overall luck is to increase the number of random times. After all, what we care about is only the best one among all the results of exploration. Therefore, faced with such a challenge, a common solution is "random restart", that is, starting the search at multiple different starting points to increase the possibility of finding the global optimal solution.
In the previous discussion, we explored ways to treat the state space as independent, atomic nodes when looking for problem-solving algorithms. For example, in the game of chess, moving the position of each chess piece by one square is a typical atomic operation. However, if we consider the state space in more detail, not treating it as an atomic operation, but abstracting the problem into Constraint Satisfaction Problems (CSP), we open up a new way of thinking. There are three main reasons for choosing to abstract practical problems into CSP: First, many problems in real life can be effectively abstracted into CSP problems. This abstraction makes the solution to the problem more universal, just like when we do mathematical applications When asking a question, the first thing to think about is, what is the essence of this problem? Is it a sequence problem, a plane geometry problem, or an algebra problem? Secondly, research in the past few decades has accumulated a large number of effective methods to solve CSP problems. If we can formulate a unique problem into a CSP problem, it is possible to find mature solutions from the existing tool library without having to invent new methods from scratch or design separate algorithms for each problem. Finally, the CSP method can greatly compress the search space. For example, in the map coloring problem, once the color of a certain area is determined, the colors of other adjacent areas are also determined, thus greatly reducing the search space. Perhaps when we look back in 10 years, we will be relatively confident that ChatGPT, which will cause a sensation in the industry in 2023, will be regarded as an important milestone in the development of artificial intelligence and even the entire technology field. One of the reasons for its popularity may be because of its versatility. Personal computers, the Internet, and mobile Internet, these solutions are extremely versatile. Therefore, when we face endless problems, it is an efficient method to abstractly summarize different problems into limited types and use mature tools in the tool library to solve them. If each problem is analyzed individually, not only is it inefficient, but the effectiveness of the solution may also be limited.
Currently reading
March 23, 2017We call ourself Homo sepiens - man the wise .....
Displaying 1 - 30 of 167 reviews