Author: Arnon Yogev
Date: 03:13:47 01/10/06
Go up one level in this thread
Hi Vas.
>
>Enrico's explanation is pretty good. I want to add three things:
>
>1) 60% speedup (as you get going from 32-bit Rybka to 64-bit Rybka) gives around
>30 rating points. This is more-or-less consistent with previous experiments
>about the topic of engine speed, and is about the most trustworthy piece of >data you will see in any of these discussions.
Ofcourse hardware advcantage provided much help for Rybka here, however my
initial assumption was that it couldn't have been soley due to it.
your next explanation however gave me the answer.
also I have no clue which is usually faster - 64bit version or multi CPU
support. I guess that's depends mainly on efficent tuning and proper
adjustments.
>2) There is no way settings of a top engine can improve the level by more than
>20 rating points or so. Yes, computer chess is hard, but it's not that hard.
This really strucks me.. when I come to think about it the only significant
setting improvments belongs to the CM family(at least it's consistent =) )
>3) There is another phenomenon with settings. Let's say you make 10 different
>settings, all of the same playing strength, and now play some matches against a
>different engine. You might see this:
>
>setting A: wins 40-20
>setting B: equal, ~30-30
>setting C: equal, ~30-30
>setting D-G: equal, ~30-30
>seting H: loses, 20-40
>
>Now, setting A becomes "the thing". It might even be a statistically significant
>result - keep in mind that if you use 95% confidence, then every 20 matches, you
>will get one bogus statistically significant result.
>
I understand now .. this clarifies the statistical gap between the various
results reported really good.
>No matter how honest a programmer tries to be with himself, he will run into
>this effect. Ie. you tinker, tinker, tinker - and then finally get a good
>result. Of course immediately this good result makes perfect sense - it's human
>psychology. I keep a very close log of all of the trustable Rybka results, and
>quite often end up taking a second look at what I thought were sound
>conclusions.
Well I hope those {good} trustable results will help to improve both
Rybka&Hiarcs in the future.
After all this is what healthy competition is all about.
>
>Vas
Arnon.
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.