quote:
I read Blondie24 (a good read), and I can't agree with fup in that it achieved "world class ranking" (and I think it was 200 generations). They didn't try the program in tournaments or agains hand-made evaluation functions.
I just knew someone would make me go and check! ;0) Let me go get the book...
Okay, you are correct about the number of generations (the final network took 250). I have no idea where I got 70 from. Whatever, 250 is still not very many at all when discussing evolutionary algorithms (which often run for thousands or tens of thousands of generations)
Fogel's program was ranked in the top 500 players registered with zone.com - out of a total of 120,000 players. I must have been so impressed with that I remembered it as a world class ranking. Doh! B24 achieved a rating of 2045, which is "Expert" level officially. I think this is still incredibly impressive though, given the conditions for the experiment.
Also, I'd like to point out that the purpose of the Blondie24 experiment was not to produce a world class player. It was a test to see if an evaluation function could be evolved using just win-draw-loss information. That's why the search wasn't state-of-the-art and that's why they never entered it into any serious tournaments (The book does mention a couple of online tournaments it entered though. I think it won one of them).
Anyhow, sorry I was over enthusiastic about the results but I'm getting old, and my memory is fading...
(although not as old as some, eh Steve, Geta?
![](wink.gif)
)
ai-junkie.com[edited by - fup on October 10, 2002 8:29:25 PM]