The 4th World Rapid Computer Chess Championship was held on ICC the weekend of July 17-18 and Rybka won ahead of Deep Sjeng, Crafty and Chess Thinker.
Lukas Cimiotti ran Rybka on his 13-computer 128-core cluster. There were only a couple of hardware tweaks since the Leiden ICT in May, related to memory and the use of large memory pages.
The next step on the clustering side is to reduce the number of draws. Currently, the cluster has no notion of practical chances and is perfectly happy to trade down into dead-drawn endgames when it has no advantage. Adding an understanding of practical chances will have two uses:
1) It will help the cluster win computer chess and freestyle tournaments and show itself in a good light.
2) The main use of the cluster is opening preparation, and those who use it for this purpose have a strong interest in the notion of practical chances. The objective evaluation of an opening position is important, but the opportunities for one side to go wrong are no less important.
Jiri Dufek handled the opening for us. He's been extremely busy recently - Rybka 4 book last winter and spring, followed by helping Topalov for 6 or 7 weeks, followed by catching up at work. He took the safe route in this event and used the well-tested Rybka 4 book, with a few small and safe modifications.
Our longtime book author Jeroen Noomen has informed me (during the World Cup
) that he just can't find the time and energy to continue to stay on top of the opening theory any more. He'll remain on our team and hopefully do some more relaxed and fun things. A huge thanks to Jeroen for his enormous contributions to our team, and the best of luck to him in his other activities. Our opening book is now 100% in the hands of Jiri.
The tournament cross table shows that Rybka and Deep Sjeng got exactly the same result in every round except for the third round, where our draw with Chess Thinker gave us our margin of victory. Crafty was the first non-cluster to win against a cluster since Shredder did so 14 months ago in Pamplona, beating Sjeng in the third round and tying for second place.
Tournament website: http://compchess.org/ACCAWCRCC/2010ACCAWCRCC/WCRCCResults.html
Tournament forum thread: http://rybkaforum.net/cgi-bin/rybkaforum/topic_show.pl?tid=18194
Jiri's opening report: http://rybkaforum.net/cgi-bin/rybkaforum/topic_show.pl?tid=18235
My tribute to Jeroen's Rybka contributions: http://rybkaforum.net/cgi-bin/rybkaforum/topic_show.pl?tid=18220
Jeroen's retirement notice:
Vas
Lukas Cimiotti ran Rybka on his 13-computer 128-core cluster. There were only a couple of hardware tweaks since the Leiden ICT in May, related to memory and the use of large memory pages.
The next step on the clustering side is to reduce the number of draws. Currently, the cluster has no notion of practical chances and is perfectly happy to trade down into dead-drawn endgames when it has no advantage. Adding an understanding of practical chances will have two uses:
1) It will help the cluster win computer chess and freestyle tournaments and show itself in a good light.
2) The main use of the cluster is opening preparation, and those who use it for this purpose have a strong interest in the notion of practical chances. The objective evaluation of an opening position is important, but the opportunities for one side to go wrong are no less important.
Jiri Dufek handled the opening for us. He's been extremely busy recently - Rybka 4 book last winter and spring, followed by helping Topalov for 6 or 7 weeks, followed by catching up at work. He took the safe route in this event and used the well-tested Rybka 4 book, with a few small and safe modifications.
Our longtime book author Jeroen Noomen has informed me (during the World Cup
) that he just can't find the time and energy to continue to stay on top of the opening theory any more. He'll remain on our team and hopefully do some more relaxed and fun things. A huge thanks to Jeroen for his enormous contributions to our team, and the best of luck to him in his other activities. Our opening book is now 100% in the hands of Jiri.The tournament cross table shows that Rybka and Deep Sjeng got exactly the same result in every round except for the third round, where our draw with Chess Thinker gave us our margin of victory.
Crafty was the first non-cluster to win against a cluster since Shredder did so 14 months ago in Pamplona, beating Sjeng in the third round and tying for second place.Tournament website: http://compchess.org/ACCAWCRCC/2010ACCAWCRCC/WCRCCResults.html
Tournament forum thread: http://rybkaforum.net/cgi-bin/rybkaforum/topic_show.pl?tid=18194
Jiri's opening report: http://rybkaforum.net/cgi-bin/rybkaforum/topic_show.pl?tid=18235
My tribute to Jeroen's Rybka contributions: http://rybkaforum.net/cgi-bin/rybkaforum/topic_show.pl?tid=18220
Jeroen's retirement notice:
Vas
Interesting. I wonder why Crafty wasn't running on a cluster? Bob shouldn't have a hard time getting access to hardware and I thought he had his cluster version ready to go...
> Crafty wasn't running on a cluster?
It isn't? And still tied second?
I wonder if thew cluster would have done better (e.g. it was drawing against Rybka, so it could have done it like Buzz).
I was working on a theory of how to develop an algorithm to detect practical chances some years ago.
It had something to do with giving one's own side an extra move after each first move from the given position,
and seeing if it was possible to get an advantage after that. If it wasn't possible to get an advantage even with an extra move,
then putting that move lower in the priority list than moves where it *was* possible to get an advantage if given an extra move.
I think you still have to play good moves, but if move A and move B are equally good, and move A has practical chances
while move B doesn't, choose move A.
Hope that makes sense. I think this idea dates back as far as 1983 or so, I discussed it with a friend who was a strong master
when we visited the ACM championships which were held in San Francisco. He didn't like it, but later the concept of null moves
did seem to catch on, though I haven't followed recent developments closely (haven't done much with it since the 1980s)
It had something to do with giving one's own side an extra move after each first move from the given position,
and seeing if it was possible to get an advantage after that. If it wasn't possible to get an advantage even with an extra move,
then putting that move lower in the priority list than moves where it *was* possible to get an advantage if given an extra move.
I think you still have to play good moves, but if move A and move B are equally good, and move A has practical chances
while move B doesn't, choose move A.
Hope that makes sense. I think this idea dates back as far as 1983 or so, I discussed it with a friend who was a strong master
when we visited the ACM championships which were held in San Francisco. He didn't like it, but later the concept of null moves
did seem to catch on, though I haven't followed recent developments closely (haven't done much with it since the 1980s)
all pls pls me uci wow? www back me?
What do you mean?
He's had six bizarre posts like this!!
>all pls pls me uci wow? www back me?
No we will not help you run World of Warcraft through your chess interface, and the world wide web won't back you up on this either.
hmm interesting how about giving extra pieces too :)
Personally, I think the best approach we've seen in this area was Nick Carlin's. Whatever system of opening book move-prioritization you use is basically fine, except you need to mildly tilt the move-ranks away from the moves with higher draw percentages. It's a trivial adjustment and I think it works quite well in practice. As you know, if you have the hardware-software firepower differential a few slightly suboptimal moves are really no problem if they steer you away from drawish lines effectively. I'm sure you could test this very extensively and wind up in clearly less drawish positions with only a very small penalty in your average first evaluation and probably a similarly trivial opportunity cost on the clock.
In my book the critical lines are very often 40-50% drawn. You really have to avoid that approach and make your line deviation sooner, I think.
In my book the critical lines are very often 40-50% drawn. You really have to avoid that approach and make your line deviation sooner, I think.
What the heck is with these rating for 128 CPU Cluster DeepRybka4 x64??????
Rybka MI USA, Hungary, Netherlands, Great Britain, Germany; IM Vasik Rajlich/Lukas Cimiotti Vasik, WGM Iweta Rajlich, IM Larry Kaufman, Jeroen Noomen, Nick Carlin, Lukas Cimiotti; Nalimov Intel Cluster 128 CPUs; MS Windows 64 bit; WBEC Rating 3090; RWBC Rating 3053; ICC Rating 2813;
http://compchess.org/ACCAWCRCC/2010ACCAWCRCC/WCRCCPart.html
Rybka MI USA, Hungary, Netherlands, Great Britain, Germany; IM Vasik Rajlich/Lukas Cimiotti Vasik, WGM Iweta Rajlich, IM Larry Kaufman, Jeroen Noomen, Nick Carlin, Lukas Cimiotti; Nalimov Intel Cluster 128 CPUs; MS Windows 64 bit; WBEC Rating 3090; RWBC Rating 3053; ICC Rating 2813;
http://compchess.org/ACCAWCRCC/2010ACCAWCRCC/WCRCCPart.html
From ICC: (I wonder who gave Rybka all those losses)
Information about Rybka(C) (Last disconnected Sun Jul 18 2010 12:21):
rating [need] win loss draw total best
Blitz 3006 [8] 599 96 53 748 3170 (05-Mar-2006)
Standard 2813 [6] 258 32 98 388 2920 (03-Apr-2006)
Information about Rybka(C) (Last disconnected Sun Jul 18 2010 12:21):
rating [need] win loss draw total best
Blitz 3006 [8] 599 96 53 748 3170 (05-Mar-2006)
Standard 2813 [6] 258 32 98 388 2920 (03-Apr-2006)
Why the Rybka incorrectly estimates theoretically neutral positions?
Knowledge was removed to make the engine faster.
no, that's not true if you mean eg. endgame knowledge. It was for easier programming.
If it was for easier programming he could have put the code in a different file and problem solved.
Adding the knowledge back would slow down the engine, members have even talked about having a "knowledge" switch to be used for analysis. It's not only endgames, check:
http://rybkaforum.net/cgi-bin/rybkaforum/topic_show.pl?tid=17938
Adding the knowledge back would slow down the engine, members have even talked about having a "knowledge" switch to be used for analysis. It's not only endgames, check:
http://rybkaforum.net/cgi-bin/rybkaforum/topic_show.pl?tid=17938
if you mean endgames, get endgame tablebases. Rybka removed that "knowledge" since it's available as external database.
See http://kirill-kryukov.com/chess/tablebases-online/
See http://kirill-kryukov.com/chess/tablebases-online/
For the same reason as every other engine??
Andrew
Andrew
Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill