Author: Christophe Theron
Date: 16:27:30 08/11/98
Go up one level in this thread
On August 11, 1998 at 12:19:34, Frank Schneider wrote:
>On August 11, 1998 at 01:28:50, Shaun Graham wrote:
>
>> It is not too uncommon here to hear programs being described as a player or a
>>solver. Generally it seems to me at least to mean that some programs solve test
>>positions better than they play chess, and that other programs play chess better
>>than they solve test problems. Since at least to my knowledge programs don't so
>>much plan or have strategys, then in essence they are solving each position anew
>>on every move. Thus i can't figure out how or perhaps why programs are reffered
>>to as being solvers or players. Yet there is evidence of just this for instance
>>chess tiger seems to solve almost everything, but at least to my knowledge it
>>isn't the absolute strongest program. My only conclusion as of yet, is that
>>some programs are more skilled than others at leading a game to positions that
>>they can have a better understanding of??? Does anyone have a reason on how
>>program x, can be a better solver than y but not be as good of a player?
>One reason might be a tradeoff between evaluationfunction and searchspeed
>(speed is more important for solvers, evaluation for players). Another reason
>could be the evaluation itself. Just lower piecevalues relative to pawns and
>your program becomes better at testsuites but worse at real play.
>
>Frank
The real reason in the case of Chess Tiger is just that I am a weak chess player
and also that I have been focusing for years on search efficiency.
So my program is able to see quite deep (it is not the only one), but my
evaluation function is not very accurate. Also I think I am doing too much
pruning, and Tiger sometimes miss deep subtle positional moves.
I tried to work on this in Chess Tiger 11.5. Now my evaluation is better, and I
have changed my search to avoid missing deep positional moves. The result is
that Chess Tiger 11.5 is less good on test suites, but better in real games. It
was my goal.
Some are going to be desappointed by my latest results on LCT-II:
CT11.4 on K5-100, 16Mb hash: LCT elo = 2555 (pos=46%, cmb=79%, fin=65%)
CT11.5 on K5-100, 16Mb hash: LCT elo = 2500 (pos=49%, cmb=70%, fin=52%)
...but don't be mislead by the 55 points drop on LCT. Better in positional, less
good in tactics... But it pays off in real games.
CT11.5 against CT11.4, 60 blitz games: 41.5 - 18.5 (69%) !
Christophe
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.