Not logged inRybka Chess Community Forum
Up Topic Rybka Support & Discussion / Rybka Discussion / Rybka @ 28th Dutch CCC
- - By Vasik Rajlich (Silver) [pl] Date 2008-11-18 06:53 Edited 2008-11-18 09:07
The 28th Dutch Computer Chess Championship finished this past weekend and Rybka managed to finally reach again that elusive perfect score, winning with 9/9 ahead of Hiarcs and Sjeng.

Our team was:

Operation: Hans van der Zijden
Hardware: Lukas Cimiotti
Openings: Jeroen Noomen
Engine: Vasik Rajlich, Larry Kaufman

Lukas again ran Rybka on his 40-core cluster, having made a couple of hardware tweaks since Beijing. The Beijing version of the software was used, as the newest version wasn't quite ready. The scaling performance of this Beijing version is already quite close to what we can reasonably expect from Lukas' hardware and only some small changes are still planned. This setup is essentially what we'll be using in the near future.

On the book side, Jeroen mostly stuck to his Beijing strategy. The openings again went really well. In all black games, Rybka emerged from the book a shade better (although probably closer to = than to =/+ in all four cases), while when Rybka had white only the Baron managed to reach anything close to equality.

The games can be found here.

Reports during the tournament can be found here and here.

A brief rundown of the games:

Rd 1, Rybka-Baron 1-0: The 4. Nc4 Petroff from Beijing against the same opponent was repeated. Black was prepared this time and improved with 9. .. Nd7, as b6 is a more logical square for this piece than c6. Things quickly reached the boiling point when on move 12 Rybka lashed out with a pawn sacrifice. The simplest defense would have been 12. .. Nxc4 13. Bxc4 cxd4 14. cxd4 Bxd4 15. Rad1 Qf6, when white has compensation but probably not much more. Instead, black chose the more complex 12. .. d5 13. Nd6 cxd4 14. cxd4 Bxd4 15. Rab1 Qh4, with a big mess which Rybka likes for white and which black wasn't able to hold together.

Rd 2, King-Rybka 0-1: 9. .. a5 is a modest-looking but effective novelty. Normal is 9. .. 0-0 10. b4 leading to standard white manoevers on the queenside, whereas in the game white found it difficult to settle on a plan. Rybka opened the a-file before gradually returning to the inevitable kingside play. In a tough position, white panicked with 34. Qc1 and black's attack crashed through.

Rd 3, Rybka-Pandix Al 2008 1-0: A smooth Spanish manoevering game. After a favorable opening, Rybka systematically improved her pieces and took control of the board. One move 21, she sacrificed a pawn in order to reroute a knight from c4 to d5 and black was quickly suffocated.

Rd 4, Hiarcs-Rybka 0-1: Both books ended after 14. .. Rfd8 in a position which is quite tricky for white to navigate. 15. Rc5 looks odd but is probably best, since the obvious 15. Rfd1 runs into 15. .. Bxf3 16. gxf3 Bxd4! One move later, 16. Rfd1 is still bad, this time due to 16. .. e5! 16. dxe5 Qxa2!. White seems to have two acceptable solutions at his disposal. One is the natural 16. Qc4 Be6 17. Qc1 Qxa2 18. Rxc6 Bd7 19. Rc7 a5 20. d5, which should hold the balance. The other is the weird-looking 16. e5 ?!!? Be6 17. Rxc6 Qxa2 18. Qxa2! Bxa2 19. Ra1 Bd5 20. Rca6 when black has trouble advancing his a-pawn. Optically this last position looks excellent for black, but it seems that the advantage is only symbolic. In the game, white was not able to get to the bottom of these issues and after the incorrect 18. Rc2? was left to suffer.

Rd 5, Rybka-Hermann, 1-0: Another smooth Spanish manoevering game. Rybka got a long-lasting positional advantage from the opening and then systematically improved her position. Finally, on move 34, she cashed in by simplifying into a winning pawn-up knight endgame. Black lost this game without making any obvious mistakes.

Rd 6, Rybka-Ktulu 1-0: A good illustration of Jeroen's strategy since his public book release: an offbeat opening (6. h3 vs Najdorf), analyzed in detail. White sacrificed a knight on move 10 for unclear compensation and by the time our book ended on move 16, black had returned the piece, leaving white with an extra pawn and the bishop pair. This should have been the end of the story, but some unnecessary excitement arose when Rybka simplified into a rook and bishop endgame where she underestimated black's drawing possibilities. This became moot when black himself did not not choose the best defense.

Rd 7, Sjeng-Rybka 0-1: The 2. .. a6 Sicilian from Beijing reappeared, confirming that it wasn't a one-time try. Our previous two games against Sjeng featured a long forcing variation (ICT '08) and an unusual material imbalance (Beijing '08). This game featured both! No human as white would risk 19. b3 when more flexible moves such as 19. Nc3 are obviously not any worse, but minimax search doesn't take such considerations into account. The endgame resulting from the long forcing line pitted two black rooks against two white knights and a bishop. Black's a-pawn was very strong and after some tedious manoevers white was not able to maintain its blockade.

Rd 8, Rybka-Tzunami 1-0: Tzunami is apparently a new program. It played without an opening book and got into trouble quickly. Interestingly, black had two pieces trapped in this game - first, his knight got stuck on c2, and then his queen was trapped on a1.

Rd 9, Joker-Rybka 0-1: In an open Sicilian, white played the highly enterprising pawn sacrifice 10. e5. Rybka herself thinks that this is ok, but I am skeptical. In the game, white's compensation slowly petered away and Rybka was able to consolidate and win.
Parent - - By Trike (**) Date 2008-11-18 07:57
What are the pros and cons of the cluster version? Does it use multiple engines, eg. rybka default, dynamic or human on different computers and combine them to one?
Parent - By Vasik Rajlich (Silver) [pl] Date 2008-11-18 09:08
The implementation is private for now. Only the default Rybka version is used, although this could change if we can create stronger alternative personalities.

Up Topic Rybka Support & Discussion / Rybka Discussion / Rybka @ 28th Dutch CCC

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill