||New Reviews| |Software Methodologies| |Popular Science| |AI/Machine Learning| |Programming| |Java| |Linux/Open Source| |XML| |Software Tools| |Other| |Web| |Tutorials| |All By Date| |All By Title| |Resources| |About||
Keywords: Neural networks, artificial intelligence, genetic algorithms
Title: Blondie 24: Playing At The Edge of AI
Author: David Fogel
Publisher: Morgan KaufmannISBN: 1558607838
Verdict: A great read and a good introduction
Blondie24 is the first pop-science book that I've come across that is exclusively devoted to the subject of evolutionary algorithms. Genetic algorithms and neural networks are featured strongly in books like Steven Johnson's Emergence and Mitchell Waldrop's Complexity, but here David Fogel gives them pride of place.
Evolutionary computation is inspired by the process of evolution, and what natures does using natural selection, these algorithms attempt to do in software. The challenge that Fogel and Kumar Chellapilla set themselves was to build a program that could learn to play draughts (or checkers if you're from the other side of the Atlantic). There are numerous computer programs that play the game, and indeed one of these programs, Chinook, is the world champion. However Chinook and other such programs have taken the same route that IBM took with the Big Blue computer that defeated chess world champion Gary Kasparov. In this approach the only learning that takes place is in the people that build the software. Big Blue and Chinook are both programmed with tactics, set-piece positions, traps and winning combinations. What these programs do is scan a database of moves and then pick the move that offers the best chance of winning (or the least chance of losing). Big Blue is as dumb as a housebrick - it learns nothing in the process of playing, everything has to be programmed in by its designers.
Fogel and Chellapilla took a very different approach. They wanted to build a system that learned for itself how to play the game. There was to be no pre-programming of strategies, tactics or any other form of knowledge. The software would know only how to make valid moves, everything else would have to be learned along the way. To make the problem even more difficult, the software would not even be told if a game it had played was a winning game or not - it was simply awarded a set of points for each game of five that it completed.
The player program was implemented as a neural network, which is a type of system very loosely modelled on the kind of things - neurons, synapses, dendrites etc - that our own brains are made of. The reasoning is that we're pretty good at learning, and that this learning uses networks of neurons which somehow store knowledge and understanding.
Each of these player networks could then be set to play against a set of competing networks in a series of tournaments. The networks which scored the highest number of points were then selected to survive and reproduce, while the networks which did less well were killed off. As a parallel to natural selection it's pretty rough and ready, but in terms of a software solution the results are impressive. Without explicit knowledge, the networks became increasingly competent until the winning network was able to compete to a high standard against human experts.
The fact that the object of the exercise was a game rather than some obscure mathematical problem makes the book accessible and entertaining. There's far too much about the game of draughts, but then this is more than balanced by the more interesting material about neural networks, genetic algorithms and the history of evolutionary computation. It is an entertaining story, particularly when Fogel describes the competitions against real players that were played over the internet (hence the name Blondie24, it was a way a better way of finding opponents than signing on as David or Kumar).
As pop-science books go this is one that shouldn't tax the reader too much, and it makes for a good introduction to much current work in the area of Artificial Intelligence. If there's a quibble it isn't with the book, it's with the tacky way that the end product - the network that can play a mean game of draughts - is being sold, but that's a different story.