Friotz9 3-Rybka beta 1

Main

Date: 19 Mar 2006 15:43:38
From: Henri H. Arsenault
Subject: Friotz9 3-Rybka beta 1

Although small neumbers like this are meaningless, this minimatch
yielded some interesting information, as the debate on whetehr Rybka
is really superior goes on. Maybe I'll have them playing a longer
match when I have more time.

OK, Fritz had a 3-1 score against Rybka beta at a 4-2 blitz match, and
it could just as weel been the opposite in a match this short.I had
both engines using 75 mb of hash and using no pondering, and the Nunn
opening book starting at move 5 to cancel out opening book knowledge
and to ensure a variety of opening situations.

But I noticed a few interesting things. One of them is that Rybka only
considered about 10 times fewer nodes than Fritz 9, whereas both
engines looked about 12 moves deep during the games. This means that
Rybka is pruning much more than Ftritz by using its positional
knowledge.

This is interesting because it refutes the idea that computers can
only be strong by using brute-force deep analysis of tactics. This is
also interesting for the future of computers, because although depth
of analysis grows exponentially which means that after 12 moves or so,
and doubling analysis time does not significantly increase the quality
of play, there is no such known limit on positional power.In sum, it
implies that as computers get faster and faster, there may be a bigger
payoff in putting the extra time into positional knowledge instead of
on brute force analysis.

In any case, if it turns out that Rybka is REALLY stronger than the
best programs, you can bet your last dollar that the next generation
of computer programs will have a good combination of positional and
tactical power. Imagine the power of Deep Blue combined with
positional knowledge of the same order.

Henri

Date: 19 Mar 2006 21:50:29
From: Ange1o DePa1ma
Subject: Re: Friotz9 3-Rybka beta 1

Do any of these programs learn from past games? I ran a bunch of games with
Shredder 8 vs. the last free version of Fruit. Shredder went something like
6-3 but then went on to lose by a substantial number of points, something
like 40-55.

adp

Date: 20 Mar 2006 03:20:19
From: RD
Subject: Re: Friotz9 3-Rybka beta 1

On Sun, 19 Mar 2006 21:50:29 -0500, "Ange1o DePa1ma"
<angelodpnospam@nospam.gmail.com > wrote:

>Do any of these programs learn from past games? I ran a bunch of games with
>Shredder 8 vs. the last free version of Fruit. Shredder went something like
>6-3 but then went on to lose by a substantial number of points, something
>like 40-55.

No they don't learn. It's just probability. When both engines are
very strong and difference in strength is less than 100 rating
points, sometimes a program would lose 3-7 and sometimes it will win
against the same engine by 7-3. That's why small number of games
doesn't mean anything. However, after 100s or 1000s of games, it would
be clear which engine is stronger. SSDF rating list for examples have
hundreds/thousands of games.

For now I think Rybka 1.1 is strongest (though with some obvious
weaknesses that need to be improved in future). It's now the job of
shredder, fritz, hiracs etc to improve their engines before future
version like Rybka 2 gets even better and increases the difference in
rating.

Date: 20 Mar 2006 09:37:21
From: davidp.
Subject: Re: Friotz9 3-Rybka beta 1

Hi,

Thanks for your information.
A lot of people talk about Rybka in this forum recently:

http://216.25.93.108/forum/index.php

"Henri H. Arsenault" <arseno@phy.*nospam*ulaval.ca > 撰寫於郵件
news:441d78ba.48328500@news.videotron.ca...
> Although small neumbers like this are meaningless, this minimatch
> yielded some interesting information, as the debate on whetehr Rybka
> is really superior goes on. Maybe I'll have them playing a longer
> match when I have more time.
>
> OK, Fritz had a 3-1 score against Rybka beta at a 4-2 blitz match, and
> it could just as weel been the opposite in a match this short.I had
> both engines using 75 mb of hash and using no pondering, and the Nunn
> opening book starting at move 5 to cancel out opening book knowledge
> and to ensure a variety of opening situations.
>
> But I noticed a few interesting things. One of them is that Rybka only
> considered about 10 times fewer nodes than Fritz 9, whereas both
> engines looked about 12 moves deep during the games. This means that
> Rybka is pruning much more than Ftritz by using its positional
> knowledge.
>
> This is interesting because it refutes the idea that computers can
> only be strong by using brute-force deep analysis of tactics. This is
> also interesting for the future of computers, because although depth
> of analysis grows exponentially which means that after 12 moves or so,
> and doubling analysis time does not significantly increase the quality
> of play, there is no such known limit on positional power.In sum, it
> implies that as computers get faster and faster, there may be a bigger
> payoff in putting the extra time into positional knowledge instead of
> on brute force analysis.
>
> In any case, if it turns out that Rybka is REALLY stronger than the
> best programs, you can bet your last dollar that the next generation
> of computer programs will have a good combination of positional and
> tactical power. Imagine the power of Deep Blue combined with
> positional knowledge of the same order.
>
> Henri

Date: 19 Mar 2006 09:33:45
From: irchans
Subject: Re: Friotz9 3-Rybka beta 1

Hi Henri,

I posted the blitz results at the bottom of this post a couple days
ago. I am also running a longer blitz test right now and I will post
the results in a couple of days.

I found a few sights on the internet that have Rybka ratings (both
blitz and 40moves/40min):

http://www.husvankempen.de/nunn/cegtrating4040all.html
http://computerchess.org.uk/ccrl/4040/
http://www.utzingerk.com/rating_list.htm
http://www.utzingerk.com/rybka10beta_test.htm
http://www.geocities.com/sedatchess/Spartacus_top26.html

Based on those sites Rybka 1.0 32-bit is about 50 points stronger than
Fritz9 and about 80 points stronger than Shredder 8. Rybka 1.1 32 bit
seems to be about 50 points stronger than Rybka 1.0 32-bit. The 64-bit
Rybka versions seem to be about 30-50 points stronger than the 32-bit
versions.

Bottom line: Single cpu Rybka 1.1 64 bit version seems to be about
80-140 points stronger than the next best single cpu chess programs
(Hiarcs 10 and Fruit 2.2) and about 110-190 points stronger than the
standards Fritz 9 and Shredder 9.

Cheers,
Irchans

----- 200 blitz games of Rybka 1.1 32bit vs. Shredder 8 -----

108 wins for Rybka Rybka v1.1.w32
30 wins for Shredder 8 UCI
62 ties.

According to Arena, this corresponds to rating difference of about 144
points! (There is still a lot of uncertainty with only 200 games. I
think the uncertainty is about + or - 25 points.) The large difference
in ratings might have been caused by the tournament conditions:

Blitz: 2 min + 1 second per move,
Rybka 79 MB
Shredder 71 MB
Windows XP, AMD 3200, 512 MB ram total.
Opening book: mainbook.abk provided by Arena.
Ponder Off, Adjudicate games after 250 moves.
I'm not sure if any endgame tablebases were used by Shredder. Rybka
did not use any table bases.