I will start this report by giving a short introduction to our team and the tournament, so that you can get an idea how many people were contributing to Rybka’s success. Then the highlight of this report is following, Hans’ diary from Pamplona with a lot of pictures and videos. You can click on most of the pictures to see a larger version.
So let’s introduce our team:
Of course Rybka is improving constantly, as Vasik Rajlich is working on the engine full time. So the Rybka version used in Pamplona was a private development version. The main difference to Rybka 3 was (besides improved playing strength) the cluster support. This means Rybka can use a network of computers. It’s not clear yet when a cluster version of Rybka will be made public, but it works quite well as you may have seen already in the games from Beijing.
Hardware & Testing
This year’s World Computer Chess Championship was held together with the 14th Computer Olympiad.
Prior to the World Championship, there was quite some discussion about a change of the rules, which first limited the hardware to one computer with maximum 8 logical cores and was later modified to allow small clusters with max. 8 cores and allowing hyperthreading (obviously without informing our hardware guru Lukas Cimiotti, who bought a rather expensive Intel Xeon W5580, but would have used a cluster of two quads if he would have known the modified rule changes).
At the Olympiad there was no hardware limit, so Luka’s cluster, consisting of 2x W5580 (12 GB RAM), Skulltrail 4 GHz (8GB RAM), Skulltrail 3.8 GHZ (8 GB RAM), Harpertown 2x X5460 3.8 GHz (8 GB RAM) and 5x Core i7-920 (6 GB RAM, clockspeed 4 GHz – 4.25 GHz) was used.
We had various problems getting everything to work, so in addition to his hardware and tuning work Lukas also contributed a lot of testing and debugging help. This clustering project would probably never have been started without Lukas, and it certainly wouldn’t have come as far along as it has.
We finally got everything working on Tuesday morning before the tournament, and for now have only preliminary performance figures. The cluster won a match against a Nehalem Quad (i7-920) running the latest Rybka version with a score of +37 =25 -2 (+213 Elo). I suspect that with more games this Elo gap will be slightly lower.
Lukas is a regular in the Playchess engine room (handles: Rechenschieber, Victor_Kullberg) and will probably run his cluster there on occasion.
Of course this means that the highest level chess was played at the Olympiad, but since everyone wants the more prestigous World Champion title, there were just 6 participants at the Olympiad and 10 at the World Championship.
Nick Carlin handled our opening book, taking advantage of Jeroen Noomen‘s Rybka 3 Aquarium opening book, and did really well. All of our book positions were equal or better, and all were complex and offered plenty of winning chances.
Nick is from the new breed of computer chess opening authors, who rely on systematic, automated methods and on statistical analysis. He uses all of the available resources, from Jeroen’s work to Playchess games to Aquarium editing tools to Polyglot. For this event, he made algorithmic innovations in the area of sharpening the book exit points – his aim was to drop Rybka off into rich positions with plenty of winning chances, and to prevent opposing authors from doing the opposite. Judging by the games, these methods work quite well.
Unfortunately, Nick has decided to take an extended break from computer chess after these events, as the time required to stay on top of everything is just too high for him. Jeroen is also still taking a break after his last book release and has started to apply his skills to the stock market – hopefully he will soon be rich and will then return to what is best in life 🙂 . All of this should underline just how much work is involved in the book preparation. The responsibility is high, as one mistake can spoil an entire event and wipe out the work of everyone on the team. Modern opening theory is simply a huge load and we will have to think about how to handle it.
Here is the full report.