Saturday, September 8, 2012

ATOMICC Testing Website Out of Business


On August 17 the ATOMICC testing website left the following rather curt message:

No more ATOMICC testing: Hi, friends. I'm done with computer chess and I don't plan to ever return. There will be no more ATOMICC testing. Best wishes!

     This is unfortunate because it was one site where I think testing was more in line with the conditions most people actually use when doing engine analysis: AMD Phenom II 1100T 6 CPUs @ 3.7 GHz 16 GBs of memory @ 2,000 MHz with games being played at 10 minutes per game + 10 seconds per move using a CPU with 64-bit engines- Ponder (permanent brain) = On - Hash = 512 MBs - OS = Windows 7 Pro 64-bit, SP1 and GUI = Fritz 13.
     One thing I really liked about the site was the ease with which you could look up the results of the tests for each engine.  For example, if you wanted to know how Fritz 13 performed against Stockfish you just had to click on the “Fritz” link on the side of the page and there were the results of its 8 matches.  Stockfish won convincingly +46 -9 =45.
      As you can see from the table below, Houdini Pro is easily the best but it’s not free.  I was surprised to see that Strelka did so well.  It easily won all its matches except for narrowly losing to Houdini Pro (+16 -21 =63) and it tied with Komodo64 (a commercial engine) +23 -23 =54.

Houdini Pro
Strelka 5.1
Komodo64
Critter 1.4
Rybka 4.1
Stockfish 2.2.2
Vitruvius
DeepSaros
Fritz 13
Naum
Deep Junior
 
     I never paid much attention to Strelka but here’s some info on it: In May 2007 a new chess engine called Strelka (Russian for "arrow") appeared on the scene, claimed to be written by Yuri Osipov. Soon, there were allegations that Strelka was a clone of Rybka 1.0 beta, in the sense that it was a reverse-engineered and slightly modified version of Rybka. Why does this allegation not surprise me?      Several players found Strelka to yield identical analysis to Rybka in a variety of different situations, even having the same bugs and weaknesses in some cases. Osipov, however, stated repeatedly on discussion boards that Strelka was based on Fruit, not Rybka, and that any similarities was either because Rybka also was based on Fruit, or because he had tuned the evaluation function to be as close to Rybka as possible.
      With the release of Strelka 2.0 beta, source code was included. Rajlich stated that the source made it "obvious" that Strelka 2.0 beta was indeed a Rybka 1.0 beta clone, although not without some improvements in certain areas. On the basis of this, he claimed the source as his own and intended to re-release it under his own name, although he later decided not to do so. He also made allegations that "Yuri Osipov" was a fake name.
      According to Victor Zakharov (Convekta company) in his review for Arena chess website: "I consider that Yuri Osipov (Ivanovich) is real name. He didn't hide it. However I can't state this with 100 percent assurance."
      However, Fruit author Fabien Letouzey has clearly expressed in the open letter mentioned above that Strelka 2.0 beta is a Fruit derivate with some minor changes. Anyway, if you’re not interested in all that, you can download Strelka 5.5 from Chessdotcom.
      I have not used Strelka and it looks like it might be worth considering.  Just for fun I ran a 5 minute match (which is hardly conclusive) with Strelka vs. Houdini 1.5 (32 bit) which was won by Strelka +2 -1 =2
      One thing that stood out when comparing the engine evaluations was that Strelka’s evaluation was often vastly different than Houdini’s.  For example in the following position (Black to move):

     Houdini evaluated the position at about 1.90 in White’s favor while Strelka evaluated it at 3.69 in White’s favor.  Another point worth mentioning is that with the Fritz GUI you cannot increase the number of lines shown with Strelka…you can only see one, which is an inconvenience. Download Strelka from Chessdotcom

No comments:

Post a Comment