Artificial Intelligence: A Modern Approach

Citation

Russell, S., & Norvig, P. (2020). Artificial Intelligence: A Modern Approach (4th ed.). Pearson.

Chapter Summary

The fourth edition has twenty-eight chapters in seven parts. These are followed by two appendices: one on mathematical background (complexity analysis and O() notation, linear algebra, probability distributions) and one on languages and algorithms (Backus-Naur form and the pseudocode conventions). A bibliography and an index close the book, and the main text runs to about 1,020 pages. Every chapter closes with a summary and bibliographical and historical notes.

Part I: Artificial Intelligence (Chapters 1–2)

Chapter 1, Introduction, asks what AI is: acting humanly, thinking humanly, thinking rationally, acting rationally, and beneficial machines. It reviews AI's foundations in philosophy, mathematics, economics, neuroscience, psychology, computer engineering, control theory and linguistics. It then gives a history from 1943 to deep learning and ends with the state of the art and the risks and benefits of AI.

Chapter 2, Intelligent Agents, defines agents and environments, rationality and performance measures, and the properties of task environments. It also covers the structure of agents, from simple reflex agents to goal-based, utility-based and learning agents.

Part II: Problem-solving (Chapters 3–6)

Chapter 3, Solving Problems by Searching, covers:

• problem formulation;
• uninformed strategies (breadth-first, uniform-cost, depth-first, iterative deepening, bidirectional);
• informed search with A*, weighted A* and memory-bounded variants;
• the design and learning of heuristics.

Chapter 4, Search in Complex Environments, handles local search and optimization, continuous spaces, nondeterministic actions, partial observability and online search. Chapter 5, Adversarial Search and Games, treats minimax, alpha-beta pruning, heuristic evaluation, Monte Carlo tree search, and stochastic and partially observable games. Chapter 6, Constraint Satisfaction Problems, covers constraint propagation, backtracking, local search and the structure of problems.

Part III: Knowledge, reasoning, and planning (Chapters 7–11)

Chapter 7, Logical Agents, builds propositional logic using the wumpus world, with theorem proving, resolution, model checking and propositional agents. Chapter 8, First-Order Logic, and Chapter 9, Inference in First-Order Logic, extend the logical machinery; Chapter 9 covers unification, forward and backward chaining, and resolution.

Chapter 10, Knowledge Representation, covers ontological engineering, categories, events, mental objects, semantic networks, description logics and default reasoning. Chapter 11, Automated Planning, treats classical planning, planning heuristics, hierarchical planning, planning in nondeterministic domains, and scheduling.

Part IV: Uncertain knowledge and reasoning (Chapters 12–18)

• Chapter 12, Quantifying Uncertainty: probability notation, joint distributions, independence, Bayes' rule and naive Bayes models.
• Chapter 13, Probabilistic Reasoning: Bayesian networks, exact and approximate inference, and causal networks.
• Chapter 14, Probabilistic Reasoning over Time: hidden Markov models, Kalman filters and dynamic Bayesian networks.
• Chapter 15, Probabilistic Programming: relational and open-universe models.
• Chapter 16, Making Simple Decisions: utility theory, decision networks and the value of information.
• Chapter 17, Making Complex Decisions: Markov decision processes, bandit problems and partially observable MDPs.
• Chapter 18, Multiagent Decision Making: non-cooperative and cooperative game theory, and collective decisions.

Part V: Machine Learning (Chapters 19–22)

Chapter 19, Learning from Examples, covers supervised learning, decision trees, model selection, the theory of learning, linear models, nonparametric methods, ensembles and developing machine learning systems. Chapter 20, Learning Probabilistic Models, covers learning with complete data and with hidden variables (EM).

Chapter 21, Deep Learning, covers feedforward, convolutional and recurrent networks, computation graphs, generalization, and unsupervised and transfer learning. Chapter 22, Reinforcement Learning, covers passive and active reinforcement learning, generalization, policy search and inverse reinforcement learning.

Part VI: Communicating, perceiving, and acting (Chapters 23–26)

Two chapters cover language: Chapter 23, Natural Language Processing, and Chapter 24, Deep Learning for Natural Language Processing (word embeddings, sequence-to-sequence models, the transformer, pretraining). Chapter 25, Computer Vision, covers image formation, features, image classification, object detection and the 3D world. Chapter 26, Robotics, covers hardware, perception, planning and control, and human-robot interaction.

Part VII: Conclusions (Chapters 27–28)

Chapter 27, Philosophy, Ethics, and Safety of AI, covers the limits of AI, whether machines can think, and the ethics of AI. Chapter 28, The Future of AI, looks at AI components and architectures.

Key Concepts

Part I–II: Agents and search

• Rational agent: an agent that acts to maximize expected performance given its percepts. This is the organizing idea of the book.
• A* search and heuristics: informed search guided by an estimate of remaining cost.
• Minimax and alpha-beta pruning: decision-making against an opponent.
• Constraint satisfaction: solving by propagating constraints and searching over variable assignments.

Part III–IV: Knowledge and uncertainty

• Propositional and first-order logic: formal languages for representing knowledge and inferring from it.
• Bayesian networks: compact representations of joint probability distributions.
• Markov decision processes: sequential decision-making with uncertain outcomes.

Part V–VII: Learning, perception and ethics

• Supervised learning and deep learning: learning functions from labeled data, including neural networks.
• Reinforcement learning: learning from rewards.
• Transformer architecture: the sequence model used in modern language processing.
• Ethics and safety of AI: limits, risks and responsibilities.

Critical Analysis

Strengths

• A single organizing idea. The rational agent is introduced in Chapter 2 and used to organize the remaining parts. Logic, probability, search and learning therefore appear as tools for the same problem.
• Breadth with chapter-level structure. The seven parts can be sampled independently: search, logic, probabilistic reasoning, learning, and the perception and robotics chapters are separate units. Course schedules exploit this.
  • The Purdue course described under How to Use assigns sections rather than whole chapters, for example 3.1–3.3, 6.1–6.2 and 17.2.1.
  • It also moves between parts out of book order, reading Chapter 22 on reinforcement learning before Chapter 19.
• Updated for the current field. The contents add probabilistic programming, multiagent decision making, deep learning, transformers and pretraining. There is a chapter on the philosophy, ethics and safety of AI, and the first chapter has sections on beneficial machines and the risks and benefits of AI.
• Support materials. The book's companion site (aima.cs.berkeley.edu) offers pseudocode, code, figures, exercises and a courses page.
• Still a standard course text. Both university courses described under How to Use list the 4th edition as their textbook.

Limitations

• Length and breadth. Twenty-eight chapters over about a thousand pages cannot go deep everywhere. Chapters 24, 25 and 26 are single chapters on fields that have their own textbooks.
• Uneven mathematical prerequisites. The appendix reviews complexity, linear algebra and probability in a few pages, while Parts IV and V rely heavily on probability.
• Not a hands-on programming text. Algorithms are written in pseudocode (Appendix B); code is provided separately online.
• Edition and numbering friction. The 3rd edition is still in use: the Purdue course below accepts it if the student maps sections to the 4th. Chapter numbering can differ between editions. The US and Global versions also differ in contents, so check your own table of contents. As a result, a chapter number on a syllabus does not always identify the right material (see the JHU example under How to Use).

Where It Fits in the Course

Russell and Norvig is the broad survey in this machine learning and AI course. It places learning alongside search, logic, planning and decision-making. University courses use it this way: the Johns Hopkins and Purdue courses described below both build their schedules around it, running from agents and search through logic and probability to learning and reinforcement learning.

The other books on this site cover parts of that ground in more depth or from a different angle:

• Mitchell, Artificial Intelligence: A Guide for Thinking Humans: a nontechnical account of what current systems can and cannot do. It is a good companion before or beside the textbook.
• James, Witten, Hastie, Tibshirani and Taylor, An Introduction to Statistical Learning with Applications in Python: a hands-on, minimal-mathematics treatment of machine learning with labs in Python. It is the practical complement to Part V.
• Murphy, Probabilistic Machine Learning: An Introduction: treats machine learning, including deep learning, through probabilistic modelling. It covers the ground of Chapters 19–22 from that perspective.
• Szeliski, Computer Vision: Algorithms and Applications: expands the single vision chapter into a book.

How to Use This Book

The book is intended for readers with programming experience, discrete mathematics and basic probability, with some linear algebra for the learning chapters.

A practical plan:

• Core for a first course: Chapters 1–3, 5–6 or selected parts of 4, 7, 12–13, 19 and 21.
• Logic track: Chapters 7–11.
• Probability and decision track: Chapters 12–18.
• Learning track: Chapters 19–22, supplemented by a hands-on text for coding practice.
• Language, vision and robotics: Chapters 23–26 as survey units, to be followed by specialist books.
• Closing: Chapters 27–28 can be read at any point after Chapter 1.

Use the exercises at the end of each chapter and the code on the companion site. Read the bibliographical and historical notes to follow up primary sources.

Common course assignments

Purdue CS47100, Introduction to Artificial Intelligence (Spring 2026). The course uses the 4th edition (ISBN 9780134610993) and assigns these readings in 4th-edition numbering, in schedule order:

Topic	Readings (4th ed.)
Introduction and agents	Ch. 1; Ch. 2
Search	3.1–3.3; 3.4; 3.5–3.6; 4.1
Games	5.1–5.2; 5.3
Constraint satisfaction	6.1–6.2; 6.3–6.5
Logic	7.2–7.4; 7.5–7.6; 8.2–8.3; 9.1–9.5
Uncertainty and Bayesian networks	12.2–12.6; 13.1–13.2; 13.3–13.4
Markov decision processes	17.1; 17.2.1; 17.2.2
Reinforcement learning	22.1–22.2; 22.3
Learning from examples	19.1–19.2; 19.4
Deep learning	21.1

This is close to the core plan above. It adds first-order logic (Chapters 8–9), MDPs (Chapter 17) and reinforcement learning (Chapter 22). It leaves out knowledge representation and planning (Chapters 10–11), Chapters 14–16 and 18, and all of Parts VI–VII. Deep learning gets only Section 21.1, so a reader following this route should look to a dedicated ML text for more.

Johns Hopkins 601.664, Artificial Intelligence (Spring 2026, Philipp Koehn). The required textbook is the 4th edition (2020). The schedule cites Russell and Norvig chapters for:

• introduction and intelligent agents;
• basic and informed search, plus local search;
• game playing and constraint satisfaction;
• logical agents, first-order logic and inference;
• planning, probabilistic reasoning and Bayesian networks;
• MDPs;
• statistical learning, deep learning and reinforcement learning.

It also includes a philosophy-of-mind session. Some chapter numbers on that schedule do not match the 4th-edition table of contents:

• Knowledge representation is cited as Chapter 12, but in the 4th edition it is Chapter 10; Chapter 12 is Quantifying Uncertainty.
• Philosophy of mind is cited as Chapter 26.1–2, but in the 4th edition philosophy and ethics are Chapter 27; Chapter 26 is Robotics.

Follow the topics on such a schedule and check them against your own table of contents rather than copying the numbers directly.

Edition notes

• Current edition. The 4th edition is Pearson, 2020, ISBN 9780134610993 (US).
• 3rd edition. It remains usable where a course allows it: the Purdue course above accepts it on condition that students map sections themselves. Expect chapter numbers to shift, especially in the later parts.
• US vs Global. The two versions differ in contents. Do not assume a section number from one matches the other.

Focused reading and edition boundary

The exercise below uses the companion problem on this site; read its stated assumptions and data before answering.

Reading route: In the fourth US edition, read Intelligent Agents and Learning from Examples, then probabilistic, deep or reinforcement learning as relevant.

Try before continuing: Describe an agent that uses the classifier, distinguishing its percept, predicted label, action and evaluation criterion. Checkpoint: A classification accuracy is not automatically an agent’s utility or a safety/decision guarantee.

Edition and access: The US and Global contents differ. Public exercises and code complement the commercial book; they do not make the complete book or all instructor solutions free.

Book companion site