Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Hello from Edmonton (and on Temporal Differences)

Author: Will Singleton

Date: 20:38:56 07/31/02

Go up one level in this thread


On July 31, 2002 at 21:33:29, James Swafford wrote:

>On July 31, 2002 at 17:49:05, Jay Scott wrote:
>
>>On July 30, 2002 at 22:43:36, James Swafford wrote:
>>
>>>why isn't
>>>everyone doing it??
>>
>>In my view, it's because top chess programmers are amazingly conservative. Or to
>>look at it more positively, they have a lot of time invested in and knowledge
>>gained about their traditional manual methods, and they do not believe in making
>>big changes. It's hard to argue with success!
>>
>>Over the years I've posted a bunch of machine learning suggestions (few of them
>>original to me) to rec.games.chess.computer and to this forum. Maybe it's my
>>writing style or something, but in every single case the general first reaction
>>was to ignore or dismiss the idea. That happened even when I pushed opening book
>>learning, which was not used in chess programs at the time but has become common
>>since. Arthur Samuels' classic checkers program already used a similar kind of
>>rote learning, so nobody should call it a radical new idea, but despite
>>seemingly obvious advantages it somehow took decades to show up in chess
>>programs.
>>
>>Another problem is that many of the people who've played around with learning
>>algorithms were only playing around. It takes serious knowledge to create a good
>>learning program, and different serious knowledge to create a good playing
>>program, and you have to have both to get really impressive results. Nobody's
>>done it yet.
>>
>>My advice for those who have great new ideas: Implement them yourself and become
>>a smashing success. *That's* convincing. The only problem is that to become a
>>smashing success, you'll also have to implement a lot of great old ideas.
>
>I agree with you.  I have studied TD-Leaf for a while, and I am definitely
>going to pursue creating a strong TD chess player.
>
>One of my biggest concerns was the time to train a complex evaluator.
>I spoke with Rich Sutton about this today, and he convinced me that it's
>doable.
>
>--
>James

James,

From my work with checkers, chess and td algorithms, I would have to disagree.
I would be thrilled to be proven wrong, but there's better ways to do chess than
td learning.  At least for the present.  I suspect it will be a interesting
experiment, though.

I'll offer a proposition.  I'm working on a new program based on standard,
well-known techniques.  I assume from your comments that you are working on a TD
chess player.  Six months from now, let's have a match, open source code.  Loser
pays for winner's trip (incl family, round-trip) to loser's home city for dinner
and drinks.

How about it?

Will




This page took 0.04 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.