Richardthughes
Posts: 11178 Joined: Jan. 2006
|
Tom English kills 'frontloading":
http://www.uncommondescent.com/index.php/archives/1481#comment-55271
Quote | Salvador: “Specifically, have you read No Free Lunch which specifically deals with evolutionary algorithms?”
Bill’s treatment of the Blondie24 study of Fogel and Chellapilla is quite weak. Some here may not be familiar with Blondie24. The researchers set up an evolutionary program in which the population was a collection of checkers strategies (static board evaluators implemented as artificial neural networks, to be precise). The strategies in the initial population were chosen randomly, and they were so bad that most humans could beat them.
In each generation, the fitness of each strategy in the population was assessed. But here’s the key point: There was no fitness function. Instead, each strategy played games against randomly selected opponents in the population. For each game, both players were awarded points at the end. The payoff was 3 points for a win, -2 points for a loss, and 0 points for a draw. After all of the games were played, the strategies with fewer points were culled from the population, and those with more points engaged in reproduction-with-variation.
After close to one thousand generations, the researchers took the best strategy from the population and began using it to play against humans on the Internet. The evolved strategy played under the name Blondie24, and its opponents did not know they were playing against a computer. The strategy established an expert rating, playing better than 99% of human players.
I want to make it clear that the abstract environment in which the program evaluated strategies was the game of checkers. Those strategies that did better in competition were more likely to reproduce than those that did worse. There was emphatically not a fitness function. How could there have been? How can you inspect a single strategy and say how well it will do in play? You have to play strategies against one another to know which are better than others.
The form of evolution implemented by the evolutionary program is often referred to as coevolution. The program is independent of checkers. What does Bill Dembski identify as a source of CSI? He identifies the asymmetry of the payoff function — the fact that a win gains a strategy more (3 points) than a loss costs it (2 points). Does he compute the CSI derived from that asymmetry? No. Does he give any formal argument? No.
Coevolution is a bigger problem for you to deal with than Dave’s program. What are you going to do when I present results obtained with a symmetric payoff function?
By the way, Gil Dodgen has said some nice things about David Fogel’s book Blondie24: Playing at the Edge of AI. Check it out at Amazon:
http://www.amazon.com/gp/product/product-description/1558607838
Comment by Tom English — August 22, 2006 @ 10:17 pm
|
-------------- "Richardthughes, you magnificent bastard, I stand in awe of you..." : Arden Chatfield
"You magnificent bastard! " : Louis
"ATBC poster child", "I have to agree with Rich.." : DaveTard
"I bow to your superior skills" : deadman_932
"...it was Richardthughes making me lie in bed.." : Kristine
|