Not logged inRybka Chess Community Forum
Up Topic Rybka Support & Discussion / Rybka Discussion / Ribka 2.3.2 vs 2.3.2a
- - By j_futur (**) Date 2007-06-22 12:17
Hi all,

Did you ever have some curiosity to know witch is the really the best engine between the two versions?
Normally, the 2.3.2a is the best but after my test (14 games - 2x30mn on PIV 3.2 2Go RAM and 192Mo hash
for each engine, repeted position for each opening - opening book Fritz 10 original), I am actually surprised with the score (2.3.2 leads 3-0 and all is drawn). I don't know why?

Is it about the PC config (not a quad core or other power PC) ?

FUTUR.
Parent - - By Felix Kling (Gold) Date 2007-06-22 12:18
Simple answer: play more games :)
Parent - - By lkaufman (*****) Date 2007-06-22 12:50
No, play MANY more games! Sometimes in my tests one version has a nice lead after ten thousand games, but after twenty thousand the situation is reversed!
Parent - - By j_futur (**) Date 2007-06-22 14:41
Ok thanks. So normally, the 2.2.3a will win in your opinion?
Parent - - By lkaufman (*****) Date 2007-06-22 14:56
We believe that it should win about 50.3% or so of the games. If you play a thousand games that's 503-497. Naturally luck can easily overcome this tiny lead. So you would need many thousands of games to prove this.
Parent - - By turbojuice1122 (Gold) Date 2007-06-22 18:33
Just so some other people know, assuming that half of the games between equal opponents are draws, you still have 500 decisions out of 1000 games--now create a random walk in one dimension with 500 events--on "average", the distance from the "center" (i.e. complete draw) is Sqrt(500) ~= 22.5.  Thus, throwing back in the 500 draws, one would expect that if two opponents are equally matched, the score after 1000 games will NOT be 500-500, but will, on average, be about 522.5 - 477.5.  Now it's easy to see that even after 1000 games, the effect of one opponent tending to score 50.3% is utterly unnoticeable.  In fact, the effect goes up with Sqrt(# games).  Thus, to get double the effect, we need 4000 games--this gives 6 extra elo points, still a bit below the 22.5 average fluctuation from center.  Thus, we need to double our difference, and double it again--thus means we need to multiply our 4000 games by a factor of 16, meaning that if one opponent tends to score 50.3%, we can expect to have to play 64,000 games to notice that there is a non-statistical difference.
Parent - - By jaeger1975 (**) Date 2007-06-22 18:50
that is too much!!!!

Well, this theme has a lot of controversy, but we should believe what Vas said " a difference around 2 ELO points between them". He is the one who knows both engines in its deep nature and of course he had to perform many tests to state that there is only 2 points ELO difference!

best regards and lets wait for the Rybka 3.0!

DAVID AMIEL
Parent - By lkaufman (*****) Date 2007-06-22 19:49
For this type of change, we can run 100,000 games or more at very short search depth, like 4 ply or so, and get an estimated rating gain. It might not be too accurate, but it's certainly more accurate than running a few hundred games at a normal time control, as others have explained.
Parent - - By j_futur (**) Date 2007-06-23 07:45
Hi Larry, 

Sorry, perhaps I didn't follow something before or I didn't understand but I'd like to know really witch of the two engine (2.3.2 and 2.3.2a) received the following knowledge or (what is the difference between the both):

- more positional knowledge, focused on
  * proper trade/no-trade decisions
  * pawn structure
  * evaluation of passed pawns
- improved search efficiency
- new endgame knowledge
- fully correct handling of zugzwang in all situations
- separate endgame module which attempts to estimate winning chances in king and pawn endgames
- revamped transposition table, or program "memory", which presents the user with a more consistent and stable picture of each position.

Thks!
Parent - By Fulcrum2000 (****) Date 2007-06-23 12:12
2.3.2a = 2.3.2 with some additional zugzwang fixes (and other small bugs). I would advise to use the 2.3.2a version.
Parent - By lkaufman (*****) Date 2007-06-23 16:15
All of that except for the zugzwang handling is in both versions.
Parent - - By er (**) Date 2007-06-22 19:34
Off topic!

Hi Larry,

Do You plan to write a revised edition of
"The Chess Advantage in Black and White"?

Erik
Parent - By lkaufman (*****) Date 2007-06-22 19:47
I have thought about doing so, but I wouldn't say I "plan" on it. Right now I'm pretty busy with Rybka.
Up Topic Rybka Support & Discussion / Rybka Discussion / Ribka 2.3.2 vs 2.3.2a

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill