Author: Albert Silver
Date: 12:43:35 12/07/05
Go up one level in this thread
On December 07, 2005 at 15:06:27, Steve Maughan wrote:
I've let the SP profile ("slightly positional") continue its Nunn2 match this
afternoon against Deep Fritz 8, noting that the default "Very Positional" style
won 24-16. Right now, all I can say is the games are quite different, and it is
giving DF a complete thrashing at the moment as it leads 20.5-6.5. It could very
well be one of those style/opponent issues, but that's the score at the moment.
N2-RybkaSP-DF8-10+2 2005
1 Rybka 1.0 Beta 32-bit 2830 +15/-1/=11 75.93 20.5/27
2 Deep Fritz 8 2780 +1/-15/=11 24.07 6.5/27
You're quite right on a number of accounts, such as its pawn race evaluation,
and the really bizarre endgame holes that crop up. I even saw it misevaluate a
KP vs. K endgame that was a dead draw, and saw the eval far above +1. It is in
one of the game I published below. Absolutely incredible results if you consider
this you know. I'm not sure I agree about the shallower depth you allude to. The
node count is clearly unreliable, and is only good to measure relative
performance as opposed to other machines. I consider DF8 to be a fast searcher,
and I don't see Rybka lagging behind in the plies it shows.
In any case, it really boggles the mind that two amateur engines such as this
could simply appear in a new version, and simply top the lists as they are
doing. The last such instance I can recall was Theron's Chess Tiger 12, which
was not widely available then.
Albert
>For what’s worth here are my thoughts and comments about Rybka from an aspiring
>programmer’s perspective.
>
>I hate to admit to this but my first knee-jerk thought when I heard about the
>amazing results was, “is it a clone”. I think it’s a valid question since I
>hadn’t heard of the program or the programmer before the release of the beta
>version. But it’s also a question I’d hate anyone else to ask of my engine -
>yet with Fruit, Crafty and Glaurung’s source in the public domain it’s difficult
>to avoid. Can I suggest that if anyone does want to avoid the ‘c’ word they
>might consider releasing earlier weaker versions of their programs? As soon as
>I downloaded and played with the engine it became apparent that it’s not a
>clone. In case anyone is wondering here are my main reasons for this
>conclusion.
>
>1) Different PVs to any other engine – generally shorter PVs than Fruit’s /
>Glaurung’s or Crafty’s
>
>2) Evaluation of positions has no correlation to Fruit’s / Glaurung’s or
>Crafty’s
>
>3) Cannot convert some basic endgames e.g. KBN v K
>
>4) Mate scores are always shown as +/-320 – most engines show +/-M??
>
>5) It occasionally crashes!! Fruit and Glaurung have never crashed on my
>machine
>
>6) Upon searching for “Vasik” in the CC Search engine there are a five pages of
>posts. Clearly he has been lurking around CCC for some time.
>
>7) Rybka played in CCT6, coming towards the bottom of the pack – i.e. it does
>have a pedigree
>
>So I’m 99.9% sure it’s not a clone.
>
>The next thing that pricked my attention was that the author, Vasik Rajlich, is
>an International Master (he’s even in the ChessBase players database along with
>his photograph). Now I always thought that strong chess players didn’t make
>great chess programmers. This is something that Christophe Therone has
>commented about in the past – they get too hung up on concepts that are
>difficult to program. Now as a relatively weak chess player (~1700 ELO) and
>aspiring chess programmer I’ve always taken some comfort in this thought. Yet
>Rybka seems to disprove this theory; or at least demonstrate an exception. And
>Vasik seems to have taken the typical IM / GM approach of cramming knowledge
>into his program with little care for the speed of the search. This fact has
>certainly given me pause for thought. One thing that Fruit has demonstrated
>(especially versions 1.0 and 1.5) is that a good solid search with a relatively
>basic evaluation can still play great chess. In the last five years I would
>even say that the consensus view has, up to this point, been that search was the
>area of a program that really impacted the strength. I’m not saying that Rybka
>has a weak search, on the contrary it seems to be brilliant at tactics, but for
>me Rybka demonstrates the importance of a great evaluation. I also came across
>a CCC post by Vasik where he stated that if you have a relatively fast engine
>and you start adding a little knowledge you may well see a fall in strength
>since the program’s set of knowledge is not balances and in most positions it’s
>just slower than the ‘dumb’ version. This makes sense and is something that
>I’ve observed with Monarch. I’m certainly inspired to work harder on Monarch’s
>evaluation having seen Rybka play – which must be a good thing!
>
>I think with Fruit and Rybka we have interesting times ahead. Others may
>disagree but I see Fruit and Rybka as a somewhat different paradigms. Fabian
>has a brilliantly clean search while I feel Rybka has a fantastic evaluation.
>Will they converge or diverge in their approach – who knows?
>
>I’ve also pitted Rybka against the usual suspects and it’s performed admirably.
>As others have commented it seems to have sophisticated mobility and space
>control heuristics. It also seems to be relatively happy with odd looking pawn
>configurations. Rybka wins many games due to its pawn race evaluation which
>seems to be the best in the industry. One thing I have noticed is that Rybka
>seems to have a tendency to swap off queens early on in the game. I suspect it
>would do better to be a little stickier with its queen and let its tactical
>prowess shine through.
>
>Anyway that’s my 2 cents – you may or may not agree. Here’s a game I enjoyed
>where Rybka rips into Gandalf 6.
>
>[Event "5 Minutes/Game + 1 Seconds/Move"]
>[Site "Lake Forest, FL"]
>[Date "2005.12.05"]
>[Round "1"]
>[White "Rybka 1.0 Beta 32-bit"]
>[Black "Gandalf 6.0"]
>[Result "1-0"]
>
>1. d4 {book 0s} d5 {book 0s} 2. c4 {book 0s} dxc4 {book 0s}
>3. Nf3 {book 0s} Nf6 {book 0s} 4. e3 {book 0s} e6 {book 0s}
>5. Bxc4 {book 0s} a6 {book 0s} 6. Qe2 {book 0s} c5 {book
>0s} 7. dxc5 {book 0s} Bxc5 {book 0s} 8. O-O {book 0s} b5
>{book 0s} 9. Bd3 {+0.12/11 7s} Nc6 {book 0s} 10. Rd1
>{+0.11/11 10s} Qe7 {-0.16/13 21s} 11. a4 {+0.11/11 7s} b4
>{-0.25/13 11s} 12. Nbd2 {+0.15/11 10s} O-O {-0.19/13 18s}
>13. b3 {+0.22/12 15s} e5 {-0.35/13 14s} 14. Ne4 {+0.38/9
>5s} Be6 {-0.32/12 8s} 15. Nxf6+ {+0.98/12 9s} Qxf6
>{-0.32/12 14s} 16. Qc2 {+0.90/12 6s} e4 {-0.54/12 17s}
>17. Bxe4 {+1.58/11 6s} Qxa1 {-1.05/12 27s} 18. Bb2
>{+1.51/12 8s} Qxd1+ {-1.31/13 23s} 19. Qxd1 {+1.62/13 5s}
>Rfd8 {-1.31/13 19s} 20. Qc2 {+1.95/12 11s} Na5 {-1.21/11
>6s} 21. Bxh7+ {+2.33/11 8s} Kh8 {-1.05/12 6s} 22. Bd3
>{+2.60/11 5s} Bxb3 {-1.50/11 8s} 23. Qxc5 {+3.42/13 6s}
>Rxd3 {-2.46/13 15s} 24. h3 {+3.42/13 10s} Rd5 {-2.36/12 4s}
>25. Qxb4 {+3.84/12 16s} f6 {-3.39/12 13s} 26. Kh2 {+3.84/12
>4s} Bd1 {-2.43/10 8s} 27. e4 {+4.53/11 3s} Nc6 {-4.38/14
>11s} 28. Qb7 {+4.61/12 4s} Rdd8 {-4.67/13 3s} 29. Qxc6
>{+4.74/11 7s} Rdc8 {-4.73/11 3s} 30. Qd5 {+4.75/12 3s} Bxf3
>{-4.89/12 3s} 31. gxf3 {+4.75/15 5s} Re8 {-4.86/11 3s}
>32. e5 {+5.56/13 3s} fxe5 {-5.82/13 5s} 33. Bxe5 {+6.62/14
>3s} Rad8 {-6.30/14 5s} 34. Qe4 {+7.43/14 6s} 1-0
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.