Let me compare two textbooks:

(1) "Artificial Intelligence: A Modern Approach" by Stuart J. Russell and Peter Norvig and

(2) "Artificial Intelligence: Structures and Strategies for Complex Problem Solving" by George F. Luger (sixth edition).

I have an impression that the former (1) is biased towards symbolic AI (especially logic-based) and the latter (2) is biased towards statistical methods. Do you think the same? Or, in wider sense, is (1) more rational and (2) more empirical? Do you know other AI books with a strong emphasis on statistical methods?

  • 1
    $\begingroup$ The sixth edition of the F. Luger book was published in 2009, that was before the big bang in deeplearning. Perhaps there are some subjects discussed but not with the bias found since 2012 in which machine learning gets massive support. $\endgroup$ – Manuel Rodriguez Nov 7 '18 at 18:47
  • 1
    $\begingroup$ The same goes for the 3rd edition of Russell/Norvig, which dates to 2009. (Norvig is currently director of research at Google, so I'm guessing any perceived bias likely relates that edition being pre-Alpha Go era, when we didn't necessarily have strong results/validation for statistical.) Possibly there will be a 4th edition? $\endgroup$ – DukeZhou Nov 7 '18 at 22:29

Methodology bias is difficult to avoid, since we can only see the methodologies that have been developed to proof of concept. Time is a continuous horizon of bias breaking in research. ARPAnet, which is now the Internet, was designed to reduce the bias by narrowing the gap between research laboratories, but it does not bridge across time.

Luger's book is the more modern approach of the two, but still dated, which shows the age of Russell and Norvig's work. In a field that changes so quickly, neither are textbooks of choice for artificial intelligence survey courses at the most cutting edge universities today. There are less biased textbooks from that period too, such as Artificial Intelligence, by Patrick Henry Winston, 1992, Addison-Wesley.

MIT's Artificial Intelligence Lab (course 6.034) formerly used Russell and Norvig, as recently as 2012, but the course is now taught by Winston and uses a publicly available updated extension of his textbook. The original Winston textbook's coverage is more comprehensive than the union of the two textbooks mentioned in the question, even before the updates made by MIT faculty and researchers.

Textbooks are most valuable when the development of a new foundation for a field has solidified and research shifts into what Thomas Kuhn termed normal science, where post-graduate level work hammers out the details of the edges of the field. Once that occurs, the foundation can be codified in standard problems and solutions.

The state of AI research is currently following what Kuhn called a paradigm shift, but the foundations of AI are not nearly at rest. Winston, recognizing this, decided to use a web reference for the course and continuously update it, with the help of PhD candidates, laboratory staff, and associated faculty.

For a more recent mathematical treatment specific to machine learning, Foundations of Machine Learning by Mohri, Rostamizadeh, and Talwalkar, 2012, MIT Press is an excellent work for those with sufficient background to follow the mathematics. It covers probability distributions and expectation in a formal way and using the more recent nomenclature and terminology.

For instance, the second chapter presents a probabilistic framework they call the PAC Learning Framework, which relates certainty, allowable approximation, and sample size requirements across learning approaches.

| improve this answer | |

Not the answer you're looking for? Browse other questions tagged or ask your own question.