هوش مصنوعی: راهنمای سیستم های هوشمند

I got this book as part of a short course in AI by Negnevitsky that I attended a while back. The course was, in my opinion, too short for the material covered. The book, however, appeared to be more promising. I’ll start with the good points. First, it is well-written and covers the “essentials” of AI such as expert systems, fuzzy logic, neural networks, genetic algorithms, hybrid intelligent systems and data mining. Second, each chapter is well-organized with sufficient examples, a summary of key points and questions for review at the end. Third, at just over 400 pages and being only around 9.5 x 6.25 inches, it is also quite easy to carry around and read at your convenience. Fourth, the pages are bright white with crisp black text which also makes for easy reading even where lighting is not perfect.

However, I do have a few issues with the book. First, it does not really cover things like Monte-Carlo search, the minimax algorithm (used in chess) or swarm intelligence, to name a few. I found that as I looked for clarifications about certain things, I came across these other topics which weren’t in the book; which brings me to the second issue. The beginning of each chapter is seductive with its easy-going introduction and general overview, especially to the uninitiated, I would imagine. However, the average reader (I have advanced degrees in computer science, by the way) will likely find himself trying to catch his breath after that. There is a little too much content squeezed into too few pages. Even more, Negnevitsky uses a considerable amount of mathematics, charts and diagrams which are not always easy understand. It is assumed, of course, that the reader has a “basic” understanding of math. If “advanced” math is used in say, rocket science, “basic” is just a relative term. If you simply skip over these things or assume they are true without trying hard to really understand them, you will not likely learn as much.

I did not intend to read this book to relive my undergraduate course in AI but it put me through it nonetheless. I was actually hoping for a less technical but sufficiently lucid explication of the different approaches currently used in AI; a “refresher” course, so to speak. Something that would explain the general principles without focusing too much on actual pen and paper calculations (which are unnecessary, even if one works in AI, unless one actually plans to employ a particular approach; in which case they can pursue it further elsewhere). In that respect, I was somewhat disappointed. This book appears to be intended mainly for undergraduates with the “be ready for the exam” mentality.

The problem is, by the end of the book, you begin to wonder just how much you’ve really learned. I would say it unlikely reaches even 50% of all that has been jam-packed into this book. To test this hypothesis, just see how many of the “questions for review”, in total, that you can answer correctly after reading the whole book. Not to mention actually being able to do the kind of calculations the book seems to emphasize. To summarize the second issue, the book kind of pulls the reader away from gaining an important conceptual perspective of AI techniques and how they relate to each other. This is still possible despite the undergraduate and generally technical nature of the book but you will have to be careful to see the forest for the trees. Having both a strong, technical grasp of the techniques *and* a conceptual overview of how they relate to each other as a field is what, I think, the book tries to do but falls short at the expense of one.

The third issue pertains to the *ten* case studies at the end of the book. I’m not really sure that many are necessary, though (something to keep in mind for a possible 3rd edition of the book). While some of them are a refreshingly straightforward read, by the end of the book, you will likely find yourself having to go back to the chapters in which the techniques employed were initially explained to really make sense of them (even more so if you had skipped over the technical parts, which I didn’t). In certain cases, Negnevitsky seems to have forgotten that while this book was “developed from lectures to undergraduates” (see the back cover), his readers are not necessarily attending those lectures afterward to ask for clarifications. For instance, in Case Study 9, he mentions the Gini coefficient and says they were used in Figure 9.46a but it is not explained *how* exactly they were used. If you look up the Gini coefficient in Wikipedia, it doesn’t help much in this context, either. I, for one, was not previously familiar with it. The fourth issue is that I think there is also at least one significant error in the book in Figure 9.22. It says on page 327 that we can improve digit recognition by feeding the network with ‘noisy’ examples and that this is shown in Figure 9.22 (on the next page). However, the figure seems to show that the network trained with noisy examples has a higher percentage of recognition error. How is this an improvement?

Another thing I noticed is that there isn’t really an equal treatment of even the topics covered. Fuzzy logic and neural networks seem to come up more often. This can be condoned to an extent but I really did not see the purpose of bringing up Adaptive Neuro-Fuzzy Inference Systems (ANFIS) as part of an “introductory text for a course in AI” and later referencing it in Case Study 8, which implies that it should be properly understood. Perhaps it deserved better treatment in the context of this book. Genetic algorithms, on the other hand, was nicely explained and later made Case Study 7 relatively easy to understand. Finally, I have to say that the cover art does the book only further injustice.

In summary, I would still recommend purchasing this book because some parts are beautifully explained and this is good for quick reference, especially when memory fails. However, there is still room out there for a less-technical, conceptually-inclined *introduction* to how things work in AI. Such a book may not be on the required reading list of undergraduate courses in AI or advanced courses in philosophy but it would probably be much more accessible to the public and even computer scientists in general.