Rating: 
Summary: An inspiring story about an artificially evolved algorithm
Review: This book tells you the story of a computer algorithm that, in the words of the author, "taught itself to play checkers". The second section, which contains the core of the book, is organized as an abridged diary that describes the steps taken to create this program and to test it in rated competition under the name of Blondie24 against human opponents at an online gaming site.Since the book is aimed at a general audience, the first section introduces all the relevant knowledge required to fathom the rest of the book. This section introduces the purpose of artificial intelligence, the demystification of the all-famous computer Deep Blue that defeated chess champion Garry Kasparov, the basics of artificial neural networks and of the game of checkers, and a survey of previous attempts at producing computer programs to play checkers. The last chapter in this preparatory section states the two fundamental questions that the author (David B. Fogel) and his partner in this endeavor (Kumar Chellapilla) set out to answer when they started to build their own checkers program, which can be phrased as: (1) Can a computer program learn on its own the features important to play checkers at the level of a human expert? (2) Can this learning be achieved by just playing games against itself and receiving feedback only after a series of games without even knowing which games were won or lost but only how many? Their approach consisted of using an idea borrowed from "mother nature" that only until recently has started to be embraced by the scientific community in the field of artificial intelligence. This idea is evolution. By combining random mutation with selection over a "population" of checkers-playing artificial neural networks that played against each other they obtained after 250 generations a program that was able to reach the expert-level rating and that even scored a few victories against human players rated at the master-level. Most technical details are left out in order to make the text accessible to a wider audience. However, in the spirit of being a scientific document, there are references to all relevant scientific papers in case you want to do further research. The writing style is both engaging and easy to follow. In addition to the main text of the book, there is a wealth of notes in a special section at the end of the book which the author uses to expand on specific topics that might be of interest to the reader. It is for this separate notes section that you might benefit from using two bookmarks, instead of just one, while you read the book. There is also an interesting section in which the author addresses a series of objections that have been raised against the ideas he discusses in this book. In my opinion, the only weakness of the book is that it spends too much ink in telling you about the moves that Blondie24 (and its siblings) made in some specific games. This will be of interest to you only if you are keen on checkers. If you have an academic training equivalent to a B.S. in Computer Science you will have the additional benefit of ending up with a clear picture of how to reproduce the ideas used to create Blondie24, although no computer code is offered. The idea they use is so simple and yet so powerful that you'll be temped to jump into the bandwagon of evolutionary computation after reading this book.
Rating: 
Summary: Pushing the Envelope of Artificial Intelligence
Review: This is an engaging first-person account, oriented toward the general reader, of how a brilliant young researcher has expanded the envelope of artificial intelligence (AI). The author begins by describing how he has grappled with the fundamental issue of what AI should be. He, like an increasing number of workers who have parted ways with the traditional AI community, insists that intelligent systems exhibit the ability to adapt to changing circumstances in order to achieve goals. Astonishing as it may be to some, most AI systems have no adaptivity whatsoever, and a definition of intelligence as adaptive behavior is indeed a radical departure. The rest of the story relates how the author (with his collaborator, Kumar Chellapilla) has applied his general principles of AI to the specific problem of having the computer teach itself to play checkers. The approach was to obtain a key component of a standard checkers program not by programming it with expert knowledge, but by allowing it to emerge in a process of simulated evolution. In this process, components were evaluated in play against one another, and the better ones survived and generated mutant offspring, while the others were discarded. I omit many details, but the essence is that this simple adaptive system gave rise to Blondie24, a checkers program that plays at the expert level. There are checkers programs that play much better than Blondie24, but none has ever played so well with so little input of human expertise. In sum, the principles are potentially revolutionary, the practical results are extraordinary, and the reading is downright fun.
Rating: 
Summary: only half a book
Review: While containing an introduction to some AI and game programming concepts, unfortunately the authors put out only half an effort and rushed out what is only half a book. The story of the development (and probably the development itself) of a checkers playing program stopped too soon. Blondie24 only got to about the 2200 point or master playing level. If the authors had bothered to let it run for three times as many generations and/or gotten faster hardware, they could have gotten it to play at a much better level. It doesn't even appear that the developers performed any more generations of self-learning after they improved the program's speed. That doesn't make any sense. They should have optimized BEFORE letting it run for 6 months!! Duh! So instead of proving that they could let a program teach itself through AI to be a really really good checkers playing program, they merely accomplished that they could get it to be a pretty good program that is not capable of beating any real grand master human players. Furthermore, the authors did not even bother to reveal what the program actually learned: what were the 17 factors the program considers and how did it assess their weightings after the 800+ generations??? We don't know, because the vital conclusion is mysteriously missing! So what we are left with is really only half a book and only half a story. If the authors had waited a year or two, got another couple thousand generations of the faster program under their belts, used at least a 1 gighertz system instead of a 450 mega hertz system, and revealed what Blondie figured out about checkers... then you would have a whole book. Instead, what they have presented is almost a waste of time. I guess they were so impatient that they just wanted to publish a book... even just half of one rather than doing it right. A much better book is the longer One Jump Ahead, by Johnathon Schaffer.
| 
