Customer Reviews

Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference

5 star:		(6)
4 star:		(1)
3 star:		(0)
2 star:		(0)
1 star:		(2)

 

 

Recently I needed to learn the principles of Bayesian networks quickly, so I bought three books: this one by Pearl, "Pattern Recognition and Machine Learning" by Bishop, and "Bayesian Artificial Intelligence" by Korb and Nicholson. Each has a very different audience and different set of strengths.

The Bishop book would probably be a great text for a serious student with a year to spend learning the theory of machine learning. But I found it a bit too concise, with a bias towards an "algebraic" description rather than a "geometric" one (my preference). I wound up spending a lot of time trying to translate equations into mental pictures in order to grasp the concepts. Too much work, so I dropped this after a couple of days.

Next I tackled the Korb and Nicholson book. This one's aimed at the application engineer who wants to get a network up and running quickly, and is not too concerned about how it works. I've been in that position many times in my career, and have always welcomed books like this for giving me a quick start into a new field. But this time I needed to really understand how Bayesian networks worked, and for this the Korb and Nicholson book is not great. In the first 9 pages of chapter 3 they try to explain the belief propagation algorithm, but their hearts just weren't it in--I found their explanation to be unintelligible. (I suspect most readers just skim this to get to the applications.) So after several days of struggling and getting nowhere, I tossed this aside as well.

The Pearl book was the only left; I put it off to last since I was initially somewhat intimidated by it. After all, this is one of the books that kicked off the "Bayesian revolution," so I was fearing a foundational math book consisting of one dry theorem after another. Not so! Although you have to read 174 pages to get through the belief propagation algorithm for trees, this took far less time than reading the first 62 pages of Korb and Nicholson, which cover roughly the same ground. The reason: Pearl is a gifted teacher and writer. His explanations are a series of baby steps, leaving nothing out, never assuming the reader will make "obvious" inferences, and supplying motivation every step of the way. Although he doesn't have a lot of figures in the book, the ones he does have are excellent, and by the time I hit page 175, I had a clear picture in my head of not just how it all worked, but why it worked. In fact, after just two days of reading, I was able to implement the belief propagation algorithm in software in an afternoon (I tested the software with examples from the Korb and Nicholson book, so that book was ultimately useful). Pearl made the subject seem almost obvious. If you are looking for a book to help you get canned Bayesian software up and running for an application quickly, this is not it. But if you want to really understand how these things work, and don't have a lot of time available, I cannot imagine a better book than this.

One of the best references on probability theory and uncertain reasoning, this book is one of my most prized. It's lucid enough to be an excellent textbook for the novice, and thorough enough to be a valuable reference for the experienced. It's a book that will always remind me (lest I forget) of the importance of probabilistic reasoning in AI.

This book is an absolutely essential book for AI programming. I've found no better book for explaining the recent advances in probability theory and its relevance to real-life, practical artificial intelligence development. It's written in a very down-to-earth and highly entertaining style with plenty of examples.

I've been looking for a good introduction to Bayes nets for a long time, and this one is by far the best and most comprehensive.

Probability is increasingly becoming one of the major foundations of effective artificial intelligence, and I strongly recommend this book to anyone with an interest in AI or probability theory.

Pearl's "Probabilistic Reasoning in Intelligent Systems" is elegantly done seminal work on uncertainty, probabilistic reasoning and all things related inference. As the author says, "This book is a culmination of an investigation into the applicability of probabilistic methods to task requiring automated reasoning under uncertainty", it covers topics on all level i.e. basic ideas, technical and substantive discussions and advanced research. However, my impression of book's target audiences is researchers and readers with a advance understanding of these topics.

"Probabilistic Reasoning in Intelligent Systems" provides very comprehensive and detailed discussion on topics like why uncertainty is important, probabilistic reasoning for query answering system, Markov and Bayesian networks etc; It goes beyond the text and into philosophical discussion as well, for instance it talks about what Bayesian rule's mathematical representation actually mean. The topic "Learning structures from data" is a good discussion of belief networks. As an advance text book, it's equipped with theorem proofs, exercises but not very many examples which disappoints. The book covers default logic very well; topics like semantics for default reasoning, casualty modularity and tree structures, evidential reasoning in taxonomic hierarchies, decision analysis, and autonomous propagation as a computational paradigm are some of the well discussed ones. I particularly enjoyed the Bayesian vs. Dempster-Shafer formulism, probabilistic treatment of the Yale shooting problem and dialogue between logicist and probablist, the concluding discussion.

I'd recommend this book as a secondary resource for advance researchers in the field of probability and uncertainty.

Of course, this book is a classic, and the low number of reviews is only because it was published in the 80s.

This book has revolutionized the field of AI, and made Bayesian networks ubiquitous in computer science today (though, BNs were first proposed in 1970 by Suppes or perhaps even earlier).

[ Interestingly, Suppes used BNs to deal with causality. ]

I think part of this book's originality is the use of a mathematical theory (ie, probability theory) into AI. A similar and earlier revolutionary step was taken by John McCarthy in his use of formal logic in AI.

Chapter 5 is actually about what I'd call probabilistic abduction, but the naming of the chapter is a bit misleading.

There are now newer and perhaps better texts on BNs, eg: "Learning BNs" by Neapolitan, the tome by Koller and Friedman (MIT Press), and Darwiche

Extremely well written. Best resource for learning message-passing algorithms, specially for causal relations or directed acyclic graphs. I bought it for 20$, very cheap for such a cool book. It was soft cover though. In one word, Excellent!!!

Pearl's book is a classic in artificial intelligence. Many of his ideas are still being studied today.

I've read the first half of this book twice now (once for a class, once to pass M.S. test). The book is terrible. There are virtually no examples to help you learn how to construct a Bayesian network. I'm serious. Virtually the only example is actually a homework question at the end of a chapter and the question is wrong!!! (there are dependcies in the table given for the joint distribution, but the acyclic graph shows them as independencies).

This book was written in defense of Bayesian Networks as a "Reasonable" graphical model. At the time, perhaps it was needed, but today we accept them as useful and move on unless we are trying to model medical diagnosis. For this reason the book is written with proofs where there should be examples (and perhaps references to an appendix). Please don't prove to me that Bayesian networks are reasonable, show me how they are useful!

To reiterate, you will learn how to create and use Bayesian networks from somewhere else, even if you buy this book.

Oh, and my FAVORITE example is the Prisoner's Paradox. He uses this example to show relationships that should be representable in a graphical model. But the whole point of the paradox is that humans are VERY bad at thinking in this manner. Though Pearl makes general claims as to the similarity between Bayesian Networks and the way humans think (doctors performing medical diagnosis is not a normal human task!), this example shows the opposite. It is called a paradox because it is unintuitive, weakening claims as to the likeness of Bayesian Networks to human thought.

I used this text in addtition to a few others for a course in probablistic reasoning (Bayes nets, etc.) and found that it was very unhelpful. The explanations were very poor and many parts were difficult to read. Also, there weren't very many examples and those that were provided were not very detailed. If you're looking for a text to learn probabilistic reasoning I would suggest trying a different book. Pearl's book could be useful as a 2nd or 3rd reference but not for the primary text. 2 thumbs down.