Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: TDchess discussion

Author: jonathan Baxter

Date: 14:05:33 02/03/99

Go up one level in this thread


On February 02, 1999 at 18:40:01, Will Singleton wrote:

>
>On February 02, 1999 at 17:23:01, jonathan Baxter wrote:
>
>>On February 01, 1999 at 12:27:00, Will Singleton wrote:
>>
>>>This week TDChess pushed past the magic 2600 mark.  Those who have been watching
>>>will have noted that TDChess improved in each of the last four weeks (while
>>>playing an average of about 450 games per week), starting at 2338 five weeks
>>>ago. Don't know whether this is due to program changes or the effects of TD
>>>lambda, perhaps Jon could enlighten us.
>
>
>>
>>Well, the rating increase is due to 3 factors:
>>
>>1:
>>4 weeks ago I realised I had left a compiler switch on that meant TDChess was
>>playing with a quiescence search that only looks at promotions. That made a big
>>difference.
>>
>>2:
>>Over Christmas, TDChess was playing with some really dumb and simple parameters
>>in its eval, but again about 4 weeks ago I turned on the TD learning for a
>>week's play and it improved dramatically and altered the parameters a lot.
>>
>>3: TD learning slows the program down by a factor of 2 because you have to pass
>>big structures around in the code. So I turned that off after a week (about 3
>>weeks ago) and just left the *opening book* learning on. TDChess has quite a
>>complicated way of learning its opening book, and note that it has absolutely
>>zero stored games so all the openings it plays it has learnt from scratch.
>>
>
>Jon,
>
>Interesting.  So you're saying that TD learning has been on only for about a
>week since TDChess started playing on ICC in December, right?  Is that because
>you've got diminishing returns (eval params vs speed) beyond 500-600 games, or
>rather that you are concentrating on the book learning for now?

Both. After about 1000-2000 games the parameters tend not to change the ordering
of moves much, so there's not a lot of point in learning after that. And I
wanted to look at the book learning. The current evaluation function is dumb but
fast, and I have never learnt a really big brain with that kind of eval. So I am
curious to see what happens. The brain currently has about 17,000 entries.


>
>What's the outlook for TD learning in chess?  Do you anticipate much
>improvement, or is it rather a function of search depth?  I would tend to think
>the latter, given the Crafty Goes Deep line of thinking.

I think you can only get so far with TD learning and then you have to look at
ways to get tactically stronger (which really means get deeper). TD learning is
great for learning the eval, but the eval only predicts positional things, it
can't even predict simple tactics like a potential fork. After a certain point
your program becomes fundamentally limited by its tactical strength, no matter
how strong it is positionally.

I think the next big step in machine learning for chess (and other games) will
be to learn to search better. I think about how to do this a lot, but I simply
don't have the time to try out any of my ideas at the moment. I am pretty sure
it will require doing a very different kind of search to minimax.

>
>Also, I've been wondering what the changes are from KnightCap to TDChess.
>

TDChess is about 5-6 times faster in nodes per second than KnightCap. KnightCap
had a really complicated eval which I have "dumbed down" a lot to get more
speed. I also changed the search a fair bit, including how it does move ordering
(its done on the basis of a static-exchange evaluation now) and how quiescence
works. The sorting of moves is a lot faster due to the fact that I now sort only
"on demand". Also the book learning is very different from the original
KnightCap. There are hundreds of more minor changes too.


Cheers,

Jon

>Will
>
><article snipped, see previous msg)



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.