Author: Uri Blass
Date: 06:36:53 01/02/06
Go up one level in this thread
On January 02, 2006 at 09:01:25, Ross Boyd wrote: >On January 02, 2006 at 04:20:59, Uri Blass wrote: > >>On January 02, 2006 at 03:20:36, Ross Boyd wrote: >> >>> >>>I tested TRACE against Gerbil 2 to compare with Uri's results posted earlier. >>>Surprisingly TRACE eclipsed Movei's performance... but what TRACE lacks in >>>intelligent search/eval she compensates for with exceptional luck. ;-) >>> >>>Hardware was P3-600Mhz... no pondering. All 3,4 and 5 man EGTBs. >>>Time control was 40 moves in 1 minute (repeating). >>>The final result was 24-4-12 (W-L-Draw) >> >>This result does not surprise me because of the following reasons. >> >>1)Trace is probably better than Movei in endgames thanks to using hash for >>pruning in the endgame. > >Possibly. I don't know how much difference it makes. > >>2)I still did not implement special knowledge for these endgames and one of the >>reasons is that normal games is a very long way to test because many of them do >>not get this type of endgames(I will probably test again in these endgames when >>I implement special knowledge for endhames). >> >>3)My hardware is faster than your hardware and I guess that longer time control >>help the weaker program to be closer to 50% > >Maybe not. I ran the test again on my Athlon XP2800+ and the score this time was >24-4-12 so she scored only a half point less (30-10). > >I think the reason TRACE outscored Movei is that TRACE has a good deal of >technical draw knowledge which allows her to swindle games especially against >engines that don't have such knowledge in their eval (like Gerbil). I do not know the reason that it outscored movei but I may try later a match against trace after Dann compile for me new version without the tablebases bug. As far as I remember latest version has no more endgame knowledge than 00_8_295 and I am not sure at this moment if the changes made it better or worse than 295 one change in the evaluation is that mobility is no more based on the number of the opponent moves in previous ply and it make movei slower. On the other hand eval hash made it faster. > >The endgame suite does highlight some pawn eval problems which I hope to fix. >TRACE doesn't know enough about candidate passers and it cost her a few points. Movei at this moment know nothing about candidate pawns and it also know nothing about backward pawns so I am not surprised if trace has more knowledge than it. If I try to implement some knowledge about it then the first test of it is going to be in the endgame test suite. Uri
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.