It's pretty well done, it covers theory and core areas but - maybe it was more the state of the field when it was written - I found it unsatisfyingly un-synthesized, unconnected, and short of detail (but this is subjective). I found the 2nd edition of Russell and Norvig to be a better introduction where it covers the same topic, which it does for everything I can think of, except VC dimension.

The book sorely needs an update, it was written in 1997 and the field has moved fast. A comparison with Mitchell's current course (materials generously available online) shows that about 1/4 of the topics taught have arisen since the book was published; Boosting, Support Vector Machines and Hidden Markov Models to name the best-known. The book also does not cover statistical or data mining methods.

Despite the subjective complaint about lack of depth it does give the theoretical roots and many fundamental techniques decently and readably. For many purposes though it may have been superceded by R&N 2nd ed.

I agree with some of the previous reviews which criticize the book for its lack of depth, but I believe this to be an asset rather than a liability given its target audience (seniors and beginning grad. students). The average college senior typically knows very little about subjects like neural networks, genetic algorithms, or Baysian networks, and this book goes a long way in demystifying these subjects in a very clear, concise, and understandable way. Moreover, the first-year grad. student who is interested in possibly doing research in this field needs more of an overview than to dive deeply into
one of the many branches which themselves have had entire books written about them. This is one of the few if only books where one will find diverse areas of learning (e.g. analytical, reinforcment, Bayesian, neural-network, genetic-algorithmic) all within the same cover.

But more than just an encyclopedic introduction, the author makes a number of connections between the different paradigms. For example, he explains that associated with each paradigm is the notion of an inductive-learning bias, i.e. the underlying assumptions that lend validity to a given learning approach. These end-of-chapter discussions on bias seem very interesting and unique to this book.

Finally, I used this book for part of the reading material for an intro. AI class, and received much positive feedback from the students, although some did find the presentation a bit too abstract for their undergraduate tastes

I teach AI at the graduate level in a major US research University, and I specialize in the area. The book does cover many different areas of Machine Learning. Unfortunately, the treatment is quite superficial. A student would find it extremely difficult to grasp imortant concepts without referring to other material. It may be a good reference, but I would definitely not recommend it as the main textbook. Unfortunately, there seem to be very few books in this area adequate for a senior or graduate level course.