Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: TDLeaf evaluation learning in EXchess

Author: Oliver Roese

Date: 21:58:14 07/01/00

Go up one level in this thread


On July 01, 2000 at 19:28:19, Dan Homan wrote:

>I keep meaning to retire EXchess and write an engine with some
>more advanced techniques like bitboards or 0x88, but I always
>end up coming back to EXchess to try "one more thing".
>
>The "one more thing" that I have been meaning to try for a while
>is Temporal Difference evaluation learning (e.g. knightcap) because
>it seemed simple to implement and I hate to manually tune my evaluation
>... so much so that I haven't spent more than a few hours on tuning
>the parameters.
>
>Now I can report my first impressions with TD learning.  I sat down
>this week and spent a few hours on a couple of different nights putting
>it into my program and debugging (which took most of the time).  I can say
>that the TD learning code (at least what I've done so far) is not much more
>difficult to implement than tablebase probing.  The knightcap guys really
>lay it all out very well in their published papers.
>
>My code is still messy and there are some issues I still need to work
>on.  With that said, I did try it out on the piece values in my program.
>I started the values at
>
>PAWN = 100  (fixed)
>KING = 10000  (meaningless, so learning should not affect it)
>KNIGHT = 0
>BISHOP = 0
>ROOK = 0
>QUEEN = 0
>
>and played a 100 game, 1 0 lightning match against GNU-chess.  After the
>match, the parameters were at
>
>PAWN = 100
>KING = 10000
>KNIGHT = 296
>BISHOP = 327
>ROOK = 508
>QUEEN = 997
>
>Which are pretty close to my "hand-tuned" values.  I did interfere with
>the match after game 25 to reduce the learning rate from a mean of 100
>points per update to a mean of 10 points per update.  (The actual update
>to the piece scores depends on the details of the given game and could
>be a few times the mean or a modest fraction of the mean).  My decision
>of when to reduce the learning rate probably influenced the final values,
>but I am not certain by how much... after 25 games the values were within
>a 100 points of the ones above but with lots of noise from game to game.
>
>This has been a fun little project and I have lots of issues to examine
>and improve in my particular implementation, but it seems to work and
>rather well at that.  Eventually this stuff will end up in a future
>release of EXchess, but probably not for several months at least.
>
>If any programmers are interested, I can discuss rather generally the
>types of changes I needed to make to my program to get this to work.
>
> - Dan

I am interested.
What type of changes do you mean?

Oliver Roese



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.