Not logged inRybka Chess Community Forum
Up Topic Rybka Support & Discussion / Rybka Discussion / Parameters Experiment 38: 64 Elo over R4 default
1 2 3 4 5 6 Previous Next  
- - By patrick delaurentis (**) Date 2010-11-29 23:11
i am wondering if any one knows the best depth you can get for your dollar according to their experiments? i used to be the type that loved to have each side at like 20 ply but am now more interested in looking over my blitz games and trying to improve on them.
i would like to know if there is a depth that is considered to be the prime for your time........a depth best for some one looking for lots of moves and lots of games.
Parent - - By mindbreaker (****) Date 2010-11-30 03:03 Edited 2010-11-30 04:05
Well, if you are analyzing your blitz games, I suggest you just truncate the games.  Just look at the first 25 moves or so to clean up holes in your opening repertoire.  That way you can have the engine look at more games.  The rest of your time can be better spent studying tactics and endgames.  For tactics I recommend Chess Tempo http://chesstempo.com/  For endgame I still think books are the best--there is no substitute for the knowledge in the books and hands on practice of technique.  You can spar with a computer in a variety of won positions and try to finish the machine.  Or if you are feeling brave defend a drawn but difficult game.

Other than analyzing those first 25 moves, you can also look at miniature games in all the variations you play.  They show where the inherent weakness lie.  Most of the complications in lines are measures addressing these weaknesses and attempting to exploit them in the enemy's camp.  Then if book is left early, you have some idea what measures are still needed and what are not, allowing you to be more efficient.
Parent - - By Jouni (**) Date 2010-12-02 13:10
I am sceptical for these miracle setting: like Ingo Bauer after 200 games vs Stockfish and Critter I didn't get ANY improvement at all.

Jouni
Parent - - By InspectorGadget (*****) Date 2010-12-02 13:41
Jouni, I'm gonna try one of these settings against Stockfish and Critter tonight :grin:
Parent - - By mindbreaker (****) Date 2010-12-02 21:31
I think the best was Exp 49 against Stockfish 1.9.  1.9.1 is the same strength.  The matches are only 100 game matches though.
Parent - - By Geomusic (*****) Date 2010-12-12 21:40 Edited 2010-12-12 21:47
is it just me or do the result % vs stockfish look totally random. LOL

Something is not right! More games needed!
Parent - By mindbreaker (****) Date 2010-12-13 01:26 Edited 2010-12-13 01:43
That is why there are several opponents for ratings.  However "random" is a bit strong. Also, I only count five times the Exps did worse than the 42.5% Default permitted at CCRL: http://www.computerchess.org.uk/ccrl/404.live/cgi/engine_details.cgi?print=Details&each_game=1&eng=Stockfish%201.9.1%2064-bit%204CPU#Stockfish_1_9_1_64-bit_4CPU
Parent - By mindbreaker (****) Date 2010-12-13 01:33
Recent results:
Parent - - By mindbreaker (****) Date 2010-12-02 21:46
The best against Critter 0.80 is Exp 43 which managed 79% Probably an anomaly. Sorry about the table, darn thing would not present in the other format.
Parent - - By yanquis1972 (****) Date 2010-12-02 21:52
sorry for the obvious question MB, but you do have to create both a white & a black personality for each, depending on which side the engine is playing/analysing, correct?
Parent - - By mindbreaker (****) Date 2010-12-02 22:22
Not for playing. "white" is self.  For analysis it stays on white.  And it is a mater of debate as to whether that is preferable. You can make a second engine with the values reversed when analyzing from black's perspective.
Parent - By Uly (Gold) Date 2010-12-02 23:08

> You can make a second engine with the values reversed when analyzing from black's perspective.


This should read "when you are black on your game", that is, when using contempt and analyzing your game from the black side, you create this personality with reversed values, and use it even when analyzing from the white side. This will make scores make sense (i.e. a -0.30 remains so whether you analyze from the white or black side).
Parent - By yanquis1972 (****) Date 2010-12-03 00:01
thanks, that's a big help for including any of these in any engine tourneys i may run. i thought i'd have to split one match in two -- one with the 'white' personality given white all the time, & one with the black personality always with black colors. heh.
Parent - - By InspectorGadget (*****) Date 2010-12-03 06:20
Have you tried these settings on non-incremental time controls, let's say 60 min/game or 30 min/game?
Parent - By mindbreaker (****) Date 2010-12-03 20:35
No, I only have one machine.  This one is busy playing 7/24.  I will probably get a new one when the AMD Interlagos comes out.  Annoyingly that is supposed to be delayed till the third quarter 2011.

The big problem with longer time controls is that it takes a very long time to get anything like statistically significant results.  And it still would be coming from me.  That would be a lot of time to devote just to have people question the results.

In fact, if you don't have a team of testers or several machines, the engines are replaced by new iterations before you finish evaluating the old versions.

As for increment or not...I like increment for testing.  I like engines to be able to show their stuff in all phases of the game.  Also it is rare to find a human tournament that does not use both multiple time controls and/or increment.  In fact, use of the multiple time controls is an attempt to mirror time controls in long games.  For machines after all, calculations are much faster, you could argue that they are playing at virtually postal speed in bullet games given the speed that they can process.  And if one was using a ten year old machine the equivalent of my time control would probably be 30 or 60 min per 70 moves.  Why 70 moves? Part of that is the Fritz interface...it only allows minute size chunks on the base time and part is because the average length of a computer game is much longer in moves.  So it makes sense to give them more moves per time control.

I plan to do some serious testing of the timings parameters to come up with a formula for setting the optimum timings (TC Buffer, TC Normal MT, and TC Max MT).  It will not start for a bit.  If this cycle turns out the way it is going then I will try just 3-4 in the next because we are very close to the optimum, I think...or at least where my process ends.  Then I plan to have a tournament of champions.  Default and all the winners of the cycles will play an extra long cycle that will have some newer opponents and will probably use a new book.  That should be as fair as any conditions could be.  All engines will be run at the same time side by side in the same cycle.  It should be interesting and definitive and and potentially destructive to my ego ;)  I do think that Default is underrated by about 15 Elo given the rating of the opponents.  After that, I will look into those timing parameters.
Parent - By mindbreaker (****) Date 2010-12-02 22:15
Sorry about that button thing in the middle...I was playing some bullet on ICC :)
Parent - By mindbreaker (****) Date 2010-12-02 22:11
Try 1'/70+1'/70+0'+1" Perfect2010.ctg 5-man tables and under on flash (USB thumb drive, SD card, or SSD).  And of course 64-bit.
Parent - By Geomusic (*****) Date 2011-02-07 19:30
how many games were played in total between all the versions?
- - By fustuk25 Date 2011-02-03 05:06
what r the parameters of exp 67?
Up Topic Rybka Support & Discussion / Rybka Discussion / Parameters Experiment 38: 64 Elo over R4 default
1 2 3 4 5 6 Previous Next  

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill