CCT11 surprises

[Dann Corbit]

Arasan's performance puts him up with the big boys (IOW, what used to be considered unapproachable Shredder strength). IkarusX similarly. Telepath also has some muscle on his frame, and TwistedLogic is quite high in the rankings. I think that the other placements are not quite so stunning, though Gaviotta's high ranking teeters on the surprising.

Code: Select all

    Program           Elo    +   -   Games   Score   Av.Op.  Draws
  1 Rybka           : 3363  320 296     9    83.3 %   3083   11.1 %
  2 Fruit-CCT11     : 3299  229 205     9    72.2 %   3133   33.3 %
  3 Swaminathan     : 3232  229 205     9    72.2 %   3066   33.3 %
  4 GlaurungCCT11   : 3203  267 247     9    61.1 %   3124   11.1 %
  5 ChessThinkerX   : 3185  214 205     9    61.1 %   3107   33.3 %
  6 ArasanX         : 3169  250 229     9    66.7 %   3049   22.2 %
  7 crafty          : 3168  250 229     9    66.7 %   3047   22.2 %
  8 Sjeng           : 3141  267 247     9    61.1 %   3063   11.1 %
  9 IkarusX         : 3126  250 229     9    66.7 %   3005   22.2 %
 10 thebaron        : 3087  233 226     9    55.6 %   3048   22.2 %
 11 Telepath        : 3060  233 226     9    55.6 %   3021   22.2 %
 12 TwistedLogicCCT : 3056  214 205     9    61.1 %   2977   33.3 %
 13 dshawul         : 2976  250 250     9    50.0 %   2976   11.1 %
 14 Deuterium2009   : 2912  250 250     9    50.0 %   2912   11.1 %
 15 Diep            : 2885  250 250     9    50.0 %   2885   11.1 %
 16 KtuluCCT11      : 2878  207 207     9    50.0 %   2878   33.3 %
 17 Symbolic        : 2858  267 281     9    44.4 %   2897    0.0 %
 18 Buzz            : 2827  226 233     9    44.4 %   2866   22.2 %
 19 Gaviota-CCT     : 2814  267 281     9    44.4 %   2853    0.0 %
 20 Tinker          : 2791  205 214     9    38.9 %   2869   33.3 %
 21 zct             : 2721  250 250     9    50.0 %   2721   11.1 %
 22 Tornado-C       : 2698  226 233     9    44.4 %   2737   22.2 %
 23 Timea           : 2686  226 233     9    44.4 %   2725   22.2 %
 24 HfC             : 2670  226 233     9    44.4 %   2709   22.2 %
 25 Lime            : 2668  205 214     9    38.9 %   2747   33.3 %
 26 ProphetX        : 2627  270 323     9    33.3 %   2747    0.0 %
 27 MatMoi          : 2568  259 309     9    27.8 %   2734   11.1 %
 28 Neurosis        : 2472  242 283     9    22.2 %   2690   22.2 %
 29 Clarabit        : 2463  143 236     9    22.2 %   2681   44.4 %
 30 NoonianChess    : 2396  296 320     9    16.7 %   2676   11.1 %

[/code]

[sje]

Thank you for posting the data.

I'd make one adjustment: decrement each Elo value by 450.

[Matthias Gemuh]

Thank you for posting the data.

I'd make one adjustment: decrement each Elo value by 450.

To understand the data, I had to decrement each Elo value by 350.

Matthias.

[swami]

Arasan's performance puts him up with the big boys (IOW, what used to be considered unapproachable Shredder strength). IkarusX similarly. Telepath also has some muscle on his frame, and TwistedLogic is quite high in the rankings. I think that the other placements are not quite so stunning, though Gaviotta's high ranking teeters on the surprising.

To add more things to your testimonials, I'd say that the latest 2009 version of Deuterium is much improved and comes as quite a surprise for a seemingly 2200 elo engine. Also, ZCT, and Symbolic have made mucho progress relative to their previous performances in online events.

I hope the newer version of Deuterium will be released..

[sje]

The version of Symbolic that's been playing in recent events hasn't been updated for nearly a year other than revisions made to its opening book. The program has been consistently placing more or less in the middle of the pack for a while.

The new C++ half of Symbolic has support for multiple (and distributed) search threads. But it will take more time to copy over the good stuff from the old version.

So far, the new C++ core runs fine under Cygwin. The old version would compile but would sometimes lock up due to Cygwin's wonky support for select(), pipes, and the pthread package.

The new core also runs well on PowerPC hosts as well as Intel/AMD.

Likewise, it tests out okay on both Macintosh and Linux boxes. I might even produce a non-Cygwin Windows version if I can steal decent substitution code for the pthread stuff.

I'm doing the opening book probe code right now; I finished the PGN scanner and book constructor last week. There is support for three command processors: UCI, xboard, and interactive console input.

[Vinvin]

It depends what Elo you are looking for : FIDE ? , ICC blitz ? , ICC standard ? , USCF ? or other pool ...

Thank you for posting the data.

I'd make one adjustment: decrement each Elo value by 450.

[bob]

The version of Symbolic that's been playing in recent events hasn't been updated for nearly a year other than revisions made to its opening book. The program has been consistently placing more or less in the middle of the pack for a while.

The new C++ half of Symbolic has support for multiple (and distributed) search threads. But it will take more time to copy over the good stuff from the old version.

So far, the new C++ core runs fine under Cygwin. The old version would compile but would sometimes lock up due to Cygwin's wonky support for select(), pipes, and the pthread package.

The new core also runs well on PowerPC hosts as well as Intel/AMD.

Likewise, it tests out okay on both Macintosh and Linux boxes. I might even produce a non-Cygwin Windows version if I can steal decent substitution code for the pthread stuff.

I'm doing the opening book probe code right now; I finished the PGN scanner and book constructor last week. There is support for three command processors: UCI, xboard, and interactive console input.

Look at Crafty's thread.c code, which has native windows stuff for threads..

[sje]

I'm doing the opening book probe code right now; I finished the PGN scanner and book constructor last week. There is support for three command processors: UCI, xboard, and interactive console input.

Look at Crafty's thread.c code, which has native windows stuff for threads..

I finished the in-memory book probe code and book move set normalization and weighting. I still need to get the on-disk book probe running although I expect that most usage will involve an in-memory book. Each book record covers one move and takes 24 bytes, so a 24 MB memory allocation at program start time will cover a million move book and offer near instant access.

Book building running on a single thread handles about four hundred games per second and does not require temporary files or an external sort. As with the previous version of Symbolic, the maximum ply length and minimum play count are specified at book creation time and all data is stored WTM. Also, as with the previous version, the book is neutral endian and so has the same coding for both Intel and PowerPC.

Sample book move set for the initial position based on about 9500 games:

Code: Select all

Nc3 [3/12/1] 16 0.00169205 0.21875 0.000370135 0.000709579
Nf3 [370/267/109] 746 0.0788917 0.569035 0.0448921 0.0860618
Nh3 [2/2/1] 5 0.000528765 0.5 0.000264382 0.000506842
a3 [0/3/0] 3 0.000317259 0 0 0
a4 [0/4/0] 4 0.000423012 0 0 0
b3 [6/14/0] 20 0.00211506 0.3 0.000634518 0.00121642
b4 [1/9/0] 10 0.00105753 0.1 0.000105753 0.000202737
c3 [3/10/2] 15 0.00158629 0.266667 0.000423012 0.000810948
c4 [390/216/81] 687 0.0726523 0.626638 0.0455266 0.0872783
d3 [2/6/0] 8 0.000846024 0.25 0.000211506 0.000405474
d4 [1212/1344/436] 2992 0.316413 0.477941 0.151227 0.289914
e3 [1/7/0] 8 0.000846024 0.125 0.000105753 0.000202737
e4 [2239/1917/684] 4840 0.511844 0.533264 0.272948 0.523264
f3 [0/3/0] 3 0.000317259 0 0 0
f4 [21/24/8] 53 0.00560491 0.471698 0.00264382 0.00506842
g3 [13/13/3] 29 0.00306684 0.5 0.00153342 0.00293969
g4 [4/8/0] 12 0.00126904 0.333333 0.000423012 0.000810948
h4 [2/1/2] 5 0.000528765 0.6 0.000317259 0.000608211

The fields are: move, W/L/D, WLD total, fraction of grand total (normalized usage), scoring expectation, merit (fraction times expectation), and playability rate (normalized merit).

In the above, 1 e4 will be picked about 52% of the time and 1 d4 will be selected about 29% of the time.

After 1 e4:

Code: Select all

Na6 [1/2/1] 4 0.000826788 0.375 0.000310045 0.000664011
Nc6 [163/196/87] 446 0.0921869 0.463004 0.0426829 0.0914121
Nf6 [5/25/3] 33 0.006821 0.19697 0.00134353 0.00287738
a6 [1/5/1] 7 0.00144688 0.214286 0.000310045 0.000664011
b6 [0/6/0] 6 0.00124018 0 0 0
b5 [0/3/0] 3 0.000620091 0 0 0
c6 [66/114/14] 194 0.0400992 0.376289 0.0150889 0.0323152
c5 [757/862/252] 1871 0.38673 0.47194 0.182513 0.390881
d6 [150/129/37] 316 0.0653162 0.533228 0.0348284 0.0745905
d5 [16/47/8] 71 0.0146755 0.28169 0.00413394 0.00885347
e6 [204/221/71] 496 0.102522 0.482863 0.0495039 0.10602
e5 [543/597/206] 1346 0.278214 0.479941 0.133526 0.285967
g6 [7/25/4] 36 0.00744109 0.25 0.00186027 0.00398406
g5 [2/1/1] 4 0.000826788 0.625 0.000516742 0.00110668
h5 [1/3/1] 5 0.00103348 0.3 0.000310045 0.000664011

--------

For the thread stuff, I need drop-in replacements for pthread create/join and mutex create/destroy/lock/unlock. Maybe mutex trylock too. I only have the Express version of MSVC++ so I don't expect any breathtaking optimizations. Also, I have only a single core machine (a Toshiba 1.86 GHz Intel Core Solo), so some of the trickier situations seen on a multiple core machine won't appear.