Author: Robert Henry Durrett
Date: 16:22:59 05/25/02
Go up one level in this thread
On May 18, 2002 at 20:52:13, martin fierz wrote:
<snip>
>
>the approach with neural nets can be plugged into a standard alpha-beta search,
>using the neural net as evaluation function. recently, david fogel and kumar
>chellapilla used a genetic algorithm to evolve neural networks to evaluate
>checkers positions. they made a bit much hype about it ("learned to play
>checkers!" (only learned to evaluate - the search part also belongs to playing
>checkers) "defeated a commercial program!" (yeah, one for kids...)), it is a
>weak program, mainly because the neural net is very slow in evaluating. but it
>plays a much better game than a program which only evaluates material and got
>something like a 2000 rating on the MSN zone.
>and on the whole, the approach is very interesting, because today's forms of
>machine learning used by chess programs are rather poor.
>
<snip>
>
>aloha
> martin
Perhaps the problem "It is a weak program, mainly because the neural net is very
slow in evaluating." will go away eventually. As noted by Bob Hyatt on rgcc,
after the computers are a lot better than all humans, the incentive to make
engines even stronger may be replaced with [paraphrasing] other objectives.
Although Bob was referring to "user features," perhaps interest in use of neural
nets, in the manner in which you describe, may be rekindled.
Seem likely?
Bob
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.