Page 3 of 6

Re: Texel 1.07

Posted: Mon Oct 02, 2017 1:36 pm
by TimoK
Hi Peter,

this is excellent news! Thank you so much for the new Texel package! I think this is one of most interesting engine release since years, especially for people interested in cluster-solutions. Since the publication of Kai Himstedt's cluster-version of Toga there has never been any release of a cluster-approach in chess using MPI.

I just setup a match between Texel running on a Dual Xeon X5670 computer (2x 6 threads on one local machine) and Texel running on a cluster using 5 of these machines. So basically it breaks down to 12 threads vs. 60 threads. I had absolutely no problem setting this up and starting Texel with the mpiexec command. You did a great job!

Image

Nps scaling seems very good, cluster version is running ~4.7 times faster than the local machine. I'm using 4GB Hash on the local machine and 16GB on the cluster. TC is 30m+5s, so I cannot play thousands of games. But I will post results when the match is finished and publish the pgn (with evals and depths).

I hope this encourages other engine developers to think about cluster implementation using MPI, too. Now Texel sources can show the way how to do that. Best thing is that MP approach can stay like it probably already is (Lazy or ABDADA) and it's not necessary to use more complicated approaches like YBW or DTS.

Best regards
Timo

Re: Texel 1.07

Posted: Mon Oct 02, 2017 6:50 pm
by petero2
Hi Timo,
TimoK wrote:this is excellent news! Thank you so much for the new Texel package! I think this is one of most interesting engine release since years, especially for people interested in cluster-solutions. Since the publication of Kai Himstedt's cluster-version of Toga there has never been any release of a cluster-approach in chess using MPI.
Yes, there are very few cluster capable chess engines you can download, but Scorpio has been cluster capable for quite some time. I don't know if cluster enabled binaries are available though. You may have to compile it yourself.
TimoK wrote:I just setup a match between Texel running on a Dual Xeon X5670 computer (2x 6 threads on one local machine) and Texel running on a cluster using 5 of these machines. So basically it breaks down to 12 threads vs. 60 threads. I had absolutely no problem setting this up and starting Texel with the mpiexec command. You did a great job!

Image

Nps scaling seems very good, cluster version is running ~4.7 times faster than the local machine. I'm using 4GB Hash on the local machine and 16GB on the cluster. TC is 30m+5s, so I cannot play thousands of games. But I will post results when the match is finished and publish the pgn (with evals and depths).
Thanks for testing. It will be very interesting to see your results.
TimoK wrote:I hope this encourages other engine developers to think about cluster implementation using MPI, too. Now Texel sources can show the way how to do that. Best thing is that MP approach can stay like it probably already is (Lazy or ABDADA) and it's not necessary to use more complicated approaches like YBW or DTS.
Yes the MP search algorithm does not have to change much for cluster, but you still need quite a bit of infrastructure to handle message passing and transposition table sharing. Calling the implementation "lazy cluster" might be a bit misleading. Even though it is significantly easier to implement than cluster YBW/DTS, it is still a lot more work than "lazy SMP".

Re: Texel 1.07

Posted: Wed Oct 04, 2017 9:04 pm
by tpoppins
Texel may be a marvel in a technical sense, but how reliable is it as an analysis tool?

I've used v1.05 and 1.06 in the past on Chessbase's Let's Check and was disconcerted to see that while Texel's PVs usually matched (the first move, at least; Let's Check stores only three plies of a PV, BTW) those of SF, Houdini and Komodo, its eval was often "off" by 50 cp or more.

After installing v1.07 in the Deep Fritz 14 GUI and running it on a dozen or so positions it looked like the new version was more "conformist" than its predecessors. Then I ran into this 10-men endgame:

[d]1r6/Rb6/3k4/KP5B/P4p1P/8/8/8 b - - 0 65

Texel proclaims what appeared to be a winning advantage for White while the other two existing Let's Check entries for this position (by two versions of Komodo, IIRC) showed 0.00.

For comparison I ran the position through Asmfish and H6 in Tactical mode and this is what I got (4GB hash, 6-men Syzygy and Syzygy probe depth=1, everything else at default for all the engines):

Code: Select all

Analysis by Stockfish 250817 64 POPCNT ASM:

65...Kc5 66.Be2 f3 67.Bf1 f2 68.h5 Kd4 
  +/= (0.66)  Depth: 7/7   00:00:04  10kN, tb=4
65...Kc5 66.Be2 f3 67.Bd3 f2 68.b6 Bd5 69.Bb5 
  +/= (0.61)  Depth: 8/9   00:00:04  11kN, tb=6
65...f3 66.Rxb7 Rxb7 67.Bxf3 Ra7+ 68.Kb4 Rf7 69.Bg2 Rf4+ 70.Ka5 Rxh4 
  +/= (0.60)  Depth: 9/13   00:00:04  23kN, tb=50
65...f3 66.Bg6 Bd5 67.Bd3 Rh8 68.Rh7 Rxh7 69.Bxh7 f2 70.Bd3 
  +/- (1.14)  Depth: 10/13   00:00:04  83kN, tb=182
65...f3 66.Bg4 f2 67.Bh3 Kc5 68.Bf1 Rf8 69.h5 Rh8 70.Be2 Rh7 71.Bd3 
  +/- (1.13)  Depth: 11/15   00:00:04  151kN, tb=211
65...Kc5 66.Be2 f3 67.Bf1 Rh8 68.b6 Bd5 69.Bd3 Kd6 70.Rh7 Rf8 71.Rg7 f2 
  +/- (1.12)  Depth: 12/21   00:00:04  282kN, tb=354
65...Kc5 66.Be2 f3 67.Bf1 Rh8 68.b6 Bd5 69.Rc7+ Kd6 70.Rg7 Rxh4 71.b7 Rh8 72.Bb5 f2 73.Kb4 
  +/- (1.01)  Depth: 13/21   00:00:04  312kN, tb=379
65...f3 66.Kb4 Bd5 67.Bg4 Rf8 68.Rd7+ Ke5 69.Bxf3 Rf4+ 70.Kc5 Rc4+ 71.Kb6 Bxf3 72.a5 Rxh4 73.a6 Ra4 74.a7 Ra1 75.Rd8 Ra3 
  +/- (0.84)  Depth: 14/22   00:00:04  555kN, tb=1317
65...f3 66.Kb4 Bd5 67.Bg4 Rf8 68.Rd7+ Ke5 69.Bxf3 Rf4+ 70.Kc5 Rc4+ 71.Kb6 Bxf3 72.a5 Rxh4 73.a6 Bd5 74.Ka5 Rh2 75.a7 Ra2+ 76.Kb6 Ra4 77.Rd8 
  +/= (0.70)  Depth: 15/26   00:00:04  1057kN, tb=4978
65...f3 66.Kb4 Bd5 67.Bg4 Rf8 68.Rd7+ Ke5 69.Bxf3 Rf4+ 70.Kc5 Rc4+ 71.Kb6 Bxf3 72.a5 Rxh4 73.Rd8 Bd5 74.a6 Ra4 75.Rh8 Kd6 76.Re8 Kd7 77.Rf8 
  +/= (0.55)  Depth: 16/28   00:00:04  1482kN, tb=7784
65...f3 66.Kb4 Bd5 67.Ra6+ Ke5 68.Rg6 Rf8 69.a5 Rf4+ 70.Kc5 Rc4+ 71.Kb6 f2 72.Rg5+ Kd4 73.Be2 Rc2 74.Rg4+ Ke3 75.Bf1 Rc1 76.Rg3+ Kd4 
  +/= (0.48)  Depth: 17/32   00:00:04  2020kN, tb=12500
65...f3 66.Kb4 Bd5 67.Ra6+ Ke5 68.Bxf3 Bxf3 69.a5 Rh8 70.Rg6 Kd5 71.a6 Rxh4+ 72.Ka5 Rh1 73.Kb6 Ra1 74.Rg3 Be2 75.Rg2 Bf3 76.Rg8 Kc4 77.Rf8 Ra3 78.Rf5 
  +/= (0.52)  Depth: 18/32   00:00:04  2713kN, tb=20303
65...f3 66.Kb4 Bd5 67.Ra6+ Ke5 68.Bxf3 Bxf3 69.Rg6 Rh8 70.a5 Rxh4+ 71.Kc5 Rh1 72.a6 Rc1+ 73.Kb6 Ra1 74.Rg3 Bd5 75.Re3+ Kd4 76.Rh3 Ke5 77.Rh8 Kd6 78.Rh4 Ra3 79.Rh6+ Ke5 80.Rh5+ Kd6 81.Rh4 Ra1 82.Rh8 
  +/= (0.46)  Depth: 19/34   00:00:04  3695kN, tb=34417
65...f3 66.Kb4 Bd5 67.Ra6+ Ke5 68.Bxf3 Bxf3 69.Rg6 Rh8 70.a5 Rxh4+ 71.Kc5 Rh1 72.a6 Rc1+ 73.Kb6 Ra1 74.Rg3 Bd5 75.Re3+ Kd6 76.Rd3 Ke5 77.Rh3 Ra4 78.Rh5+ Kd6 79.Rh2 Ke5 80.Rh3 Ke6 81.Rh8 
  +/= (0.46)  Depth: 20/37   00:00:04  4129kN, tb=39655
65...f3 66.Bg4 f2 67.Bh3 Bd5 68.Rd7+ Kc5 69.Rc7+ Kd6 70.Rc2 Ra8+ 71.Kb4 Rh8 72.Bf1 Rxh4+ 73.Ka5 Rf4 74.Kb6 Rxa4 75.Rxf2 Ra8 76.Rf6+ Ke5 77.Rg6 Be4 78.Rg4 Bf3 79.Rh4 
  +/- (0.84)  Depth: 21/44   00:00:04  8067kN, tb=75974
65...f3 66.Kb4 Bd5 67.Ra6+ Ke5 68.Bxf3 Bxf3 69.Rg6 Rh8 70.a5 Kd5 71.a6 Rxh4+ 72.Ka5 Kc5 73.Rg5+ Kc4 74.Rg1 Rh2 75.a7 Kc5 76.Rc1+ Kd6 77.Rc8 Ra2+ 78.Kb6 Ra1 79.Rh8 Bd5 80.Rd8+ Ke5 81.Re8+ Kd4 82.Rh8 Ke5 83.Rc8 Kd4 84.Rf8 
  +/= (0.43)  Depth: 22/44   00:00:04  21272kN, tb=179180
65...f3 66.Bg4 f2 67.Bh3 Kc5 68.h5 Rh8 69.Bf1 Bf3 70.Rc7+ Kd4 71.Rc4+ Ke3 72.Rc3+ Kf4 73.Rc5 Bxh5 74.Rc4+ Ke3 75.b6 Bf3 76.Rc3+ Kf4 77.Bb5 Rh5 78.Rc1 Ke3 79.Ka6 Be4 80.a5 Rh7 81.Bc4 Bg2 82.Kb5 Kd2 83.Bf1 Kxc1 84.Bxg2 
  +/- (0.92)  Depth: 23/48   00:00:05  54271kN, tb=408457
65...f3 66.Bg4 Be4 67.Kb4 Rg8 68.Rd7+ Ke5 69.Re7+ Kd4 70.Bxf3 Bxf3 71.a5 Rh8 72.a6 Rxh4 73.Rd7+ Ke5+ 74.Kc5 Rh5 75.Rd8 Ke6+ 76.Kb6 Ke7 77.Rd2 Ba8 78.Rd3 Rg5 79.Rd4 Bg2 80.a7 Rg3 81.Kc7 Rc3+ 82.Kb8 
  +/- (1.30)  Depth: 24/50   00:00:06  91590kN, tb=688844
65...f3 66.Bg4 f2 67.Bh3 Kc5 68.h5 Rh8 69.Bf1 Rh7 70.h6 Rd7 71.b6 Rh7 72.Bd3 Rf7 73.h7 Rxh7 74.Bxh7 f1Q 75.Rxb7 Qe1+ 76.Ka6 Qe8 77.Bc2 Qa8+ 78.Ra7 Qb8 79.a5 Qc8+ 80.Rb7 Qa8+ 81.Ra7 Qc8+ 
  = (0.00)  Depth: 25/50   00:00:06  96457kN, tb=725901
[...]
65...f3 66.Bg4 Kc5 67.Bh3 Re8 68.Bf1 f2 69.Bd3 Bd5 70.h5 Re3 71.Rc7+ Kd6 72.Rc2 Rxd3 73.Rxf2 Rd4 74.Rf6+ Kc5 75.Rf8 Rh4 76.Rh8 Bb3 77.Rc8+ Kd6 78.Rc6+ Kd5 79.h6 Rxa4+ 80.Kb6 Ke5 81.h7 Rh4 82.h8R Rxh8 
  = (0.00)  Depth: 70/36   00:06:32  13331MN, tb=213475839

Code: Select all

Analysis by Houdini 6.01 Pro x64-popcnt (Tactical=1):

65...f3 66.Kb4 Bd5 67.Bg4 f2 68.Rd7+ Ke5 69.Be2 Rc8 70.b6 Rc1 71.Ba6 Ke6 72.Rc7 Rb1+ 73.Ka5 Kd6 74.h5 Rh1 75.Bb5 
  +/= (0.64)  Depth: 13/33   00:00:00  4148kN, tb=13498
65...f3 66.Bg6 Bd5 67.Bd3 f2 68.Rg7 Rf8 69.h5 f1R 70.Bxf1 Rxf1 71.b6 Rh1 72.Rg6+ Kd7 73.h6 Kc8 74.Ka6 Kb8 75.Rf6 Bb7+ 76.Kb5 Rh5+ 77.Kb4 Rh4+ 78.Kc5 Rh5+ 
  +/= (0.65)  Depth: 14/40   00:00:00  13910kN, tb=76083
65...f3 66.Bg6 Bd5 67.Bd3 f2 68.Rg7 Rf8 69.Bf1 Kc5 70.b6 Ra8+ 71.Ra7 Rb8 72.Rc7+ Bc6 73.h5 Rxb6 74.Bg2 
  +/= (0.64)  Depth: 15/40   00:00:01  14103kN, tb=77019
65...f3 66.Bg4 f2 67.Bh3 Bd5 68.h5 Rh8 69.h6 Kc5 70.Rc7+ Kd6 71.Rc2 Rxh6 72.Bf1 Rf6 73.Kb6 Rf8 74.Rd2 Rb8+ 75.Ka5 Ra8+ 76.Kb4 Rf8 77.b6 Kc6 
  +/= (0.65)  Depth: 16/40   00:00:01  21238kN, tb=115946
65...f3 66.Bg4 Kc5 67.Bh3 Rf8 68.Bf1 f2 69.h5 Rh8 70.h6 Be4 71.Rc7+ Kd4 72.Rc4+ Ke3 73.b6 Rxh6 74.Rxe4+ Kxe4 75.b7 Rh8 76.Ka6 Kd5 
  +/- (0.96)  Depth: 17/44   00:00:01  37396kN, tb=205479
65...f3 66.Bg4 f2 67.Bh3 Kc5 68.Bf1 Rh8 69.h5 Bf3 70.Rc7+ Kd4 71.Rc2 Ke3 72.Rc3+ Kf4 73.Rc5 Bxh5 74.Kb4 Bf3 75.b6 Re8 
  +/- (0.89)  Depth: 18/44   00:00:01  42239kN, tb=234781
65...f3 66.Bg4 Kc5 67.Bh3 f2 68.Bf1 Rh8 69.h5 Bf3 70.Rc7+ Kd4 71.Rc2 Ke3 72.Rc3+ Kf4 73.Rc5 Bxh5 74.b6 Bf3 75.Rc2 Kg3 76.Rc3 Kf4 77.Ka6 Rh1 78.Bd3 Be4 79.Bc4 Bg2 80.Rc2 
  +/- (0.77)  Depth: 19/44   00:00:02  46629kN, tb=259215
65...f3 66.Bg4 Kc5 67.Bh3 f2 68.h5 Rh8 69.Bf1 Be4 70.Rc7+ Kd4 71.Rf7 Ke3 72.b6 Rxh5+ 73.Ka6 Rh1 74.Bb5 Rh8 75.b7 Bd5 76.Rd7 Bxb7+ 77.Kxb7 Rf8 78.Bf1 Rf5 79.Kb6 Rf8 
  +/- (1.04)  Depth: 20/52   00:00:02  70601kN, tb=405711
65...f3 66.Kb4 f2 67.Be2 Bd5 68.Kc3 Rf8 69.Bf1 Rg8 70.h5 Rg1 71.Be2 Rc1+ 72.Kd2 f1Q 73.Bxf1 Rxf1 74.h6 Rf2+ 75.Kc3 Rh2 76.Kd4 Rh4+ 77.Ke3 Bg8 
  = (0.17)  Depth: 21/53   00:00:09  287MN, tb=2254392
65...f3 66.Kb4 f2 67.Be2 Bd5 68.Kc3 Rf8 69.Bf1 Rg8 70.h5 Rg1 71.Be2 Bf3 72.Rf7 Bxe2 73.Rxf2 Bxh5 74.Kb4 Kc7 75.Ka5 Rg8 76.Rf5 Ra8+ 77.Kb4 Be8 
  = (0.00)  Depth: 22/55   00:00:14  443MN, tb=4012256
[...]
65...f3 66.Kb4 f2 67.Be2 Bd5 68.Bf1 Rh8 69.b6 Rxh4+ 70.Kc3 Rh1 71.Bb5 Rc1+ 72.Kd2 f1Q 73.Bxf1 Rxf1 74.b7 Rf2+ 75.Kd3 Rf3+ 76.Kd4 Rf4+ 77.Ke3 Re4+ 78.Kd3 Kc7 79.b8Q+ Kxb8 
  = (0.00)  Depth: 40/75   00:09:22  17996MN, tb=185845905

Code: Select all

Analysis by Texel 1.07:

65...f3 66.Rxb7 Rxb7 67.Bxf3 Ra7+ 68.Kb4 Rf7 69.Be4 
  +- (2.45 --)  Depth: 7   00:00:00  1kN, tb=3
65...f3 66.Rxb7 Rxb7 67.Bxf3 Ra7+ 68.Kb4 Rf7 69.Be4 
  +- (2.46)  Depth: 7   00:00:00  2kN, tb=3
65...f3 66.Rxb7 Rxb7 67.Bxf3 Ra7+ 68.Kb4 Rf7 69.Be4 Rf4 
  +- (2.30 ++)  Depth: 8   00:00:00  3kN, tb=21
65...f3 66.Rxb7 Rxb7 67.Bxf3 Ra7+ 68.Kb4 Rf7 69.Be4 Rf4 
  +- (2.13 ++)  Depth: 8   00:00:00  3kN, tb=22
65...f3 66.Rxb7 Rxb7 67.Bxf3 Ra7+ 68.Kb4 Rf7 69.Be4 Rf1 
  +- (2.19)  Depth: 8   00:00:00  4kN, tb=28
65...f3 66.Rxb7 Rxb7 67.Bxf3 Ra7+ 68.Kb4 Rf7 69.Be4 Rf1 
  +- (2.04 ++)  Depth: 9   00:00:00  9kN, tb=47
65...f3 66.b6 Kc5 67.Bg4 f2 68.Be2 Kd4 69.Kb4 
  +- (2.34 --)  Depth: 9   00:00:00  10kN, tb=48
65...f3 66.Bg4 f2 67.Be2 Bf3 68.Bf1 Rf8 69.b6 Bh5 
  +- (2.22)  Depth: 9   00:00:00  15kN, tb=53
65...f3 66.Bg6 f2 67.Bd3 Be4 68.Bf1 Rf8 69.h5 
  +- (2.37 --)  Depth: 10   00:00:00  30kN, tb=62
65...f3 66.Bg6 f2 67.Bd3 Be4 68.Bf1 Rh8 69.Rf7 Kc5 70.b6 Ra8+ 71.Ba6 Rxa6+ 72.Kxa6 Bd3+ 73.Kb7 
  +- (2.48)  Depth: 10   00:00:00  59kN, tb=94
65...f3 66.Bg6 Kc5 67.h5 Kd4 68.b6 f2 
  +- (2.33 ++)  Depth: 11   00:00:00  152kN, tb=203
65...f3 66.Bg4 f2 67.Bh3 Bf3 68.Rf7 Kc5 69.Rc7+ Kd6 70.Rf7 
  +- (2.37)  Depth: 11   00:00:00  172kN, tb=234
65...f3 66.Bg4 Kc5 67.h5 f2 68.Bh3 Rh8 69.h6 Rxh6 70.Rxb7 Rxh3 71.Rf7 
  +- (2.52 --)  Depth: 12   00:00:00  263kN, tb=442
65...f3 66.Bg4 Kc5 67.h5 f2 68.Bh3 Rh8 69.h6 Rxh6 70.Rxb7 Rxh3 71.Rf7 Rh2 72.b6 Kd6 73.b7 
  +- (2.22 ++)  Depth: 12   00:00:00  366kN, tb=787
65...f3 66.Bg4 Kc5 67.h5 f2 68.Bh3 Rh8 69.Bf1 Be4 70.Rc7+ Kd4 71.Rf7 Ke3 72.b6 Rxh5+ 73.Ka6 Rh1 
  +- (2.34)  Depth: 12   00:00:00  382kN, tb=854
65...f3 66.Bg4 Kc5 67.h5 f2 68.Bh3 Rh8 69.Bf1 Be4 70.Rc7+ Kd4 71.Rf7 Ke3 72.b6 Rxh5+ 73.Ka6 Rh1 
  +- (2.19 ++)  Depth: 13   00:00:00  569kN, tb=1129
65...f3 66.Bg4 Kc5 67.h5 f2 68.Bh3 Rh8 69.Bf1 Be4 70.Rc7+ Kd4 71.Rf7 Ke3 72.b6 Rxh5+ 73.Ka6 Rh1 
  +- (2.04 ++)  Depth: 13   00:00:00  659kN, tb=1509
65...f3 66.Bg4 Kc5 67.h5 f2 68.Bh3 Rh8 69.Bf1 Be4 70.Rc7+ Kd4 71.Rc4+ Ke3 72.Rc5 Bf3 73.b6 Bxh5 74.b7 Bg6 75.Kb6 Be4 
  +- (2.10)  Depth: 13   00:00:00  758kN, tb=1677
65...f3 66.Bg4 Kc5 67.h5 f2 68.Bh3 Rh8 69.Bf1 Be4 70.Rc7+ Kd4 71.Rc4+ Ke3 72.Rc5 Bf3 73.b6 Bxh5 74.b7 Rb8 75.Kb6 Bf3 76.Rc3+ Kf4 77.Rc7 Bxb7 
  +- (1.95 ++)  Depth: 14   00:00:00  938kN, tb=2269
65...f3 66.Bg4 f2 67.Bh3 Bd5 68.Ka6 Rh8 69.Bf1 Rxh4 70.a5 Rh1 
  +- (1.80 ++)  Depth: 14   00:00:00  1203kN, tb=3445
65...f3 66.Bg4 f2 67.Bh3 Kc5 68.Bf1 Rh8 69.b6 Bc6 70.Rf7 Rxh4 71.Rf5+ Kd6 72.Bb5 Bxb5 73.Kxb5 Rh5 
  +- (1.99)  Depth: 14   00:00:00  1454kN, tb=4159
65...f3 66.Bg4 f2 67.Bh3 Bd5 68.h5 Rh8 69.Ra6+ Kc5 70.h6 Bb7 71.Ra7 Be4 72.Rc7+ Kd4 73.Rd7+ Kc5 74.b6 Rxh6 75.b7 Rh8 76.Bf1 Rb8 
  +- (1.84 ++)  Depth: 15   00:00:00  1828kN, tb=4990
65...f3 66.Bg4 f2 67.Be2 Bd5 68.Rg7 Ra8+ 69.Kb4 Rh8 70.Rg6+ Ke5 71.Rg5+ Kd4 72.h5 Bf3 73.Bf1 Rxh5 74.Rxh5 Bxh5 75.b6 Bf3 
  +- (1.80)  Depth: 15   00:00:00  1943kN, tb=5942
65...f3 66.Bg6 f2 67.Bd3 Be4 68.Bf1 Rh8 69.Rf7 Kc5 70.b6 Ra8+ 71.Ra7 Rb8 72.Rc7+ Bc6 73.Rf7 Be4 74.Rxf2 Rxb6 75.Re2 Rb4 76.h5 Rd4 77.h6 
  +- (1.95 --)  Depth: 16   00:00:00  2941kN, tb=9940
65...f3 66.Bg6 f2 67.Bd3 Bd5 68.Rg7 Kc5 69.Rc7+ Kd6 70.b6 Rh8 71.Rc2 Rf8 72.Bf1 Rf4 73.h5 
  +- (1.75)  Depth: 16   00:00:00  4100kN, tb=14935
65...f3 66.Bg6 f2 67.Bd3 Bd5 68.Rg7 Ra8+ 69.Kb4 Rh8 70.Rg4 Ke5 71.b6 Bf3 72.Rg5+ Kd4 73.Rf5 Ke3 74.a5 Rxh4+ 75.Kb5 Kxd3 76.Rxf3+ 
  +- (1.90 --)  Depth: 17   00:00:00  5471kN, tb=20891
65...f3 66.Bg6 f2 67.Bd3 Bd5 68.Rg7 Be4 69.Bf1 Rh8 70.Rf7 Kc5 71.Rc7+ Kd6 72.b6 Bc6 73.Rf7 Ra8+ 74.Ba6 Bb7 75.Kb4 Rxa6 76.a5 Ra8 77.Rf6+ Ke5 78.Rxf2 
  +- (1.83)  Depth: 17   00:00:00  6908kN, tb=30180
65...f3 66.Bg6 f2 67.Bd3 Bd5 68.Rg7 Ra8+ 69.Kb4 Rf8 70.h5 Rf4+ 71.Ka5 Rh4 72.Rg6+ Ke5 73.Rg5+ Kd4 74.Rf5 Be6 
  +- (1.68 ++)  Depth: 18   00:00:00  8986kN, tb=43663
65...f3 66.Bg6 f2 67.Bd3 Bd5 68.Rg7 Ra8+ 69.Kb4 Rf8 70.h5 Rf4+ 71.Ka5 Rh4 72.b6 Rxh5 73.Ka6 Rh1 74.a5 Be4 75.Rg3 f1Q 76.Bxf1 Rxf1 
  +- (1.62)  Depth: 18   00:00:01  9652kN, tb=48687
65...f3 66.Bg6 f2 67.Bd3 Bd5 68.Rg7 Ra8+ 69.Kb4 Rf8 70.h5 Rf4+ 71.Ka5 Rh4 72.b6 Bc6 73.b7 Rxa4+ 74.Kb6 Rb4+ 75.Ka7 Bxb7 
  +/- (1.47 ++)  Depth: 19   00:00:01  15082kN, tb=87644
65...f3 66.Kb4 Bd5 67.Bg4 f2 68.Rd7+ Ke5 69.Be2 Rh8 70.Kc5 Rc8+ 71.Kb6 Bc4 72.Rd2 f1Q 73.Bxf1 Bxf1 74.Rd1 Bc4 75.Rc1 Kd4 76.h5 Ra8 77.h6 Rxa4 78.h7 Ra8 79.Kb7 
  +/- (1.57)  Depth: 19   00:00:01  26516kN, tb=182115
65...f3 66.Kb4 Bd5 67.Bg4 f2 68.Rd7+ Ke5 69.Be2 Rh8 70.Kc5 Rc8+ 71.Kb6 Rc2 72.Bd3 Bc4 73.Bxc4 Rxc4 74.Rf7 Rf4 75.Re7+ Kf6 
  +/- (1.42 ++)  Depth: 20   00:00:02  33183kN, tb=243415
65...f3 66.Bg6 f2 67.Bd3 Bd5 68.Rg7 Ra8+ 69.Kb4 Rf8 70.Rg6+ Ke5 71.Rg5+ Kd6 72.h5 Rf4+ 73.Ka5 Be4 74.Bc4 Bc2 75.Rd5+ Ke7 76.Rc5 Rxc4 
  +/- (1.54)  Depth: 20   00:00:02  55386kN, tb=449276
65...f3 66.Bg6 f2 67.Bd3 Bd5 68.Rg7 Ra8+ 69.Kb4 Rf8 70.Rg6+ Ke5 71.Bf1 Rf4+ 72.Kc5 Bc4 73.Rg5+ Kf6 74.Bxc4 Rxc4+ 
  +/- (1.39 ++)  Depth: 21   00:00:03  74439kN, tb=651163
65...f3 66.Bg6 f2 67.Bd3 Bd5 68.h5 Rh8 69.b6 Rxh5 70.Ka6 Rh1 71.a5 f1Q 72.Bxf1 Rxf1 73.Rg7 Ra1 74.Rg6+ Kc5 75.Rg4 Rb1 76.Rg5 Kd4 77.Ka7 Bc6 
  +/- (1.35)  Depth: 21   00:00:04  82042kN, tb=704212
65...f3 66.Bg6 f2 67.Bd3 Bd5 68.h5 Rh8 69.b6 Rxh5 70.Ka6 Rh1 71.a5 f1Q 72.Bxf1 Rxf1 73.Rg7 Ra1 74.Rg6+ Kc5 75.Rg4 Be6 76.b7 Bxg4 
  +/- (1.20 ++)  Depth: 22   00:00:07  154MN, tb=1691322
65...f3 66.Bg6 f2 67.Bd3 Bd5 68.Rg7 Ra8+ 69.Kb4 Rf8 70.Rg6+ Ke5 71.Rg5+ Kd6 72.Bf1 Rf4+ 73.Ka5 Be4 74.h5 Rh4 75.Bc4 Rh1 76.Rg4 Bd5 77.Bd3 Rxh5 78.Rf4 Rh2 79.Kb6 Ke5 80.Rf5+ Kd4 81.Bf1 Be4 82.Rf6 
  +/- (1.20)  Depth: 22   00:00:07  166MN, tb=1826480
65...f3 66.Bg4 f2 67.Bh3 Bd5 68.Ra6+ Kc5 69.h5 Rh8 70.h6 Bb7 71.Ra7 Be4 72.b6 Rxh6 73.b7 Rh8 74.Bf1 Bd5 75.Bb5 Be4 
  +/- (1.35 --)  Depth: 23   00:00:10  246MN, tb=2688497
65...f3 66.Bg4 f2 67.Bh3 Bd5 68.h5 Rh8 69.h6 Kc5 70.Rc7+ Kd6 71.Rc2 Rxh6 72.Bf1 Rf6 73.Kb6 Rf8 74.Rd2 Rb8+ 75.Ka5 Rf8 76.b6 Kc5 77.Rc2+ Kd4 78.Ka6 
  +/- (1.50 --)  Depth: 23   00:00:11  253MN, tb=2750364
65...f3 66.Bg4 f2 67.Bh3 Bd5 68.h5 Rh8 69.h6 Rxh6 70.Ra6+ 
  +- (1.72 --)  Depth: 23   00:00:11  260MN, tb=2805400
65...f3 66.Bg4 f2 67.Bh3 Bd5 68.h5 Rh8 69.h6 Kc5 70.Rc7+ Kd6 71.Rc2 Rxh6 72.Bf1 Rf6 73.Kb6 Rf8 74.Rd2 Rb8+ 75.Ka5 Rf8 76.b6 Kc5 77.Rc2+ Kd6 78.Kb5 Bb7 79.Rd2+ Ke7 80.a5 Rf5+ 81.Kb4 Rf4+ 82.Kc3 Rf6 83.Rd3 Ke6 84.Re3+ Kd5 85.Kd2 Kc5 86.Re7 Bd5 87.Rc7+ 
  +/- (1.54)  Depth: 23   00:00:12  294MN, tb=3163270
65...f3 66.Bg4 f2 67.Bh3 Bd5 68.h5 Rh8 69.h6 Kc5 70.Rc7+ Kd6 71.Rc2 Rxh6 72.Bf1 Rf6 73.Kb6 Rf8 74.Rd2 Rb8+ 75.Ka5 Rf8 76.b6 Kc5 77.Rc2+ Kd6 78.Kb4 Rf4+ 79.Kb5 Bb7 80.a5 Rf5+ 81.Kb4 Rf4+ 82.Ka3 Rf3+ 83.Ka4 Rf4+ 84.Kb5 
  +/- (1.58)  Depth: 24   00:00:19  468MN, tb=5072216
65...f3 66.Bg4 f2 67.Bh3 Bd5 68.h5 Rh8 69.h6 Kc5 70.Rc7+ Kd6 71.Rc2 Rxh6 72.Bf1 Rf6 73.Kb6 Rf8 74.Rd2 Rb8+ 75.Ka5 Rf8 76.b6 Kc5 77.Rc2+ Kd6 78.Kb5 Bb7 79.Rd2+ Ke6 80.Rd4 
  +/- (1.58)  Depth: 25   00:00:27  648MN, tb=7350538
65...f3 66.Bg4 Bd5 67.h5 f2 68.Bh3 Rh8 69.h6 Kc5 70.Rc7+ Kd6 71.Rc2 Rxh6 
  +/- (1.58)  Depth: 26   00:00:50  1216MN, tb=14101705
65...f3 66.Bg4 Bd5 67.h5 f2 68.Bh3 Rh8 69.h6 Rxh6 70.Ra6+ 
  +- (1.73 --)  Depth: 27   00:00:58  1414MN, tb=16142761
65...f3 66.Bg4 Bd5 67.h5 f2 68.Bh3 Rh8 69.h6 Kc5 70.Rc7+ Kd6 71.Rc2 Rxh6 
  +- (1.88 --)  Depth: 27   00:01:22  2000MN, tb=23133817
65...f3 66.Bg4 Bd5 67.h5 f2 68.Bh3 Rh8 69.h6 Kc5 70.Rc7+ Kd6 71.Rc2 Rxh6 72.Bf1 Rf6 73.Kb6 Rf8 74.Rd2 Rb8+ 75.Ka5 Rf8 76.b6 Kc5 77.Rc2+ Kd6 78.Kb5 Bb7 79.Rd2+ Ke6 80.Rd4 Ra8 81.Kb4 Rc8 82.Rd2 Rf8 83.a5 Rf4+ 84.Kc3 Rf3+ 85.Kb2 Rf5 86.Kb3 Rf3+ 87.Kc2 Ke7 88.Kb2 Rf4 89.Rc2 Kd6 90.Kb3 Rf3+ 91.Rc3 Rxc3+ 
  +/- (1.51)  Depth: 27   00:02:09  3153MN, tb=37829167
65...f3 66.Bg4 Bd5 67.h5 f2 68.Bh3 Rh8 69.h6 Kc5 70.Rc7+ Kd6 71.Rc2 Rxh6 72.Bf1 Rf6 73.Kb6 Rf8 74.Rd2 Rb8+ 75.Ka5 Rf8 76.b6 Kc5 77.Rc2+ Kd6 78.Kb5 Bb7 79.Rd2+ Ke6 80.Rd4 Ra8 81.Kb4 Rc8 82.Rc4 Rxc4+ 83.Bxc4+ Kd6 84.Bf1 Kc6 85.a5 Kd6 86.a6 Bc6 87.Kc3 
  +- (1.66 --)  Depth: 28   00:02:18  3356MN, tb=40200854
65...f3 66.Bg4 Bd5 67.h5 f2 68.Bh3 Rh8 69.h6 Kc5 70.Rc7+ Kd6 71.Rc2 Rxh6 72.Bf1 Rf6 73.Kb6 Rf8 74.Rd2 Rb8+ 75.Ka5 Rf8 76.b6 Kc5 77.Rc2+ Kd6 78.Kb5 Bb7 79.Rd2+ Ke6 80.Rd4 Rf5+ 81.Kb4 Ke5 82.Rd7 Bc6 83.b7 Rf4+ 84.Kb3 Rf8 85.Rh7 Be4 86.Rg7 
  +- (1.61)  Depth: 28   00:02:41  3918MN, tb=49256988
65...f3 66.Bg4 Bd5 67.h5 Kc5 68.h6 f2 69.Bh3 Rh8 70.h7 Bb7 
  +- (1.76 --)  Depth: 29   00:03:30  5110MN, tb=66217142
65...f3 66.Bg4 Bd5 67.h5 f2 68.Bh3 Rh8 
  +/- (1.46 ++)  Depth: 29   00:06:06  8911MN, tb=124343458
65...f3 66.Bg4 Bd5 67.h5 f2 68.Bh3 Rh8 69.h6 Kc5 70.Rc7+ Kd6 71.Rc2 Rxh6 72.Bf1 Rf6 73.Kb6 Ke5+ 74.Kc5 Be4 75.Rd2 Rf5 76.a5 Kf4+ 77.Kb6 
  +- (1.91 --)  Depth: 29   00:07:26  10856MN, tb=152407581
65...f3 66.Bg4 Bd5 67.h5 f2 68.Bh3 Rh8 69.h6 Kc5 70.Rc7+ Kd6 71.Rc2 Rxh6 72.Bf1 Rf6 73.Kb6 Rf8 74.Rd2 Rb8+ 75.Ka5 Rf8 76.b6 Kc5 77.Rc2+ Kd6 78.Kb5 Bb7 79.Rd2+ Ke6 80.Rd4 Rc8 81.Rc4 Ra8 82.Rc1 Kd6 83.a5 
  +/- (1.56)  Depth: 29   00:10:41  15617MN, tb=231720139
65...f3 66.Bg4 Bd5 67.h5 f2 68.Bh3 Rh8 69.h6 Kc5 70.Rc7+ Kd6 71.Rc2 Rxh6 72.Bf1 Rf6 73.Kb6 Rf8 74.Rd2 Rb8+ 75.Ka5 Rf8 76.b6 Kc5 77.Rc2+ Kd6 78.Kb5 Bb7 79.Rd2+ Ke5 80.a5 
  +- (1.71 --)  Depth: 30   00:11:25  16646MN, tb=245396889
65...f3 66.Bg4 Bd5 67.h5 f2 68.Bh3 Rh8 69.h6 Kc5 70.Rc7+ Kd6 71.Rc2 Rxh6 72.Bf1 Rf6 73.Kb6 Rf8 74.Rd2 Rb8+ 75.Ka5 Rf8 76.b6 Kc5 77.Rc2+ Kd6 78.Kb5 Bb7 79.Rd2+ Ke7 80.Rd4 Rc8 81.Rf4 Rc2 82.a5 Kd6 83.Kb4 Ra2 84.Rf7 Bd5 85.Rf6+ Ke5 86.Rf8 Kd4 87.Kb5 Bb7 88.Rf4+ Ke3 89.Rf7 
  +/- (1.51)  Depth: 30   00:13:03  19038MN, tb=286918730
65...f3 66.Bg4 Bd5 67.h5 f2 68.Bh3 Rh8 69.h6 Kc5 70.Rc7+ Kd6 71.Rc2 Rxh6 72.Bf1 Rh2 73.Kb6 Be4 74.Rd2+ Ke5 75.a5 
  +- (1.66 --)  Depth: 31   00:14:53  21690MN, tb=323245383
65...f3 66.Bg4 Be4 67.Kb4 Bd5 68.a5 f2 69.Rd7+ Ke5 70.Be2 Rh8 71.b6 Rxh4+ 72.Kc5 Be6 73.b7 
  +- (1.81 --)  Depth: 31   00:15:25  22455MN, tb=333670718
A bit odd, isn't it? SF and H6 see a draw within seconds, Texel a win for White even after 15 minutes.

Next step was to find what Texel would see in an actual game, preferably at LTC. According to Texel itself, 65...f3 is the critical continuation so let's start with the position after this move.

Same settings as used for the analysis

[pgn][Event "Texel endgame 60m+30s"]
[Site "DellT5600-PC"]
[Date "2017.10.01"]
[Round "1"]
[White "Texel 1.07"]
[Black "Stockfish 300917 x64 POPCNT"]
[Result "1/2-1/2"]
[SetUp "1"]
[FEN "1r6/Rb6/3k4/KP5B/P6P/5p2/8/8 w - - 0 66"]
[PlyCount "34"]
[TimeControl "3600+30"]

{Intel(R) Xeon(R) CPU E5-2670 0 @ 2.60GHz 2593 MHz W=32.8 plies; 23,096kN/s;
-1,074,375,863 TBAs B=72.1 plies; 46,907kN/s; 499,893,314 TBAs} 66. Bg4 {
1.95/27 213 Both last book move} Kc5 {0.00/71 122} 67. Bh3 {2.21/28 144} f2 {
0.00/72 186 (Rh8)} 68. h5 {1.99/27 124 (Bf1)} Rh8 {0.00/62 80} 69. Bf1 {
2.06/27 151} Rf8 {0.00/76 12 (Bf3)} 70. Be2 {0.12/28 121 (h6)} Rd8 {0.00/49 63
(Bf3)} 71. h6 {0.12/27 116 (b6)} Rf8 {0.00/57 86} 72. Bd3 {0.15/28 113 (b6)}
Re8 {0.00/76 169 (Be4)} 73. h7 {0.18/29 122 (Bf1)} Rh8 {0.00/76 67} 74. b6 {
0.18/30 108} Rxh7 {0.00/77 15} 75. Bxh7 {0.18/32 189} f1=Q {0.00/97 79} 76.
Rxb7 {0.00/38 101} Qe1+ {0.00/109 14 (Qf3)} 77. Ka6 {0.20/21 0} Kb4 {0.00/120
79 (Qe2+)} 78. Bd3 {0.19/40 118 (Rc7)} Kxa4 {0.00/125 534} 79. Ka7 {0.19/41 98}
Qa1 {0.00/45 493 (Qa5+)} 80. Rd7 {0.19/41 96 (Rc7)} Qe5 {0.00/38 77 (Kb4+)} 81.
Rf7 {0.19/41 94 (b7)} Qa5+ {0.00/34 459 (Qd4)} 82. Ba6 {0.19/43 90} Qc5 {
0.00/41 45 adjud. Draw accepted} 1/2-1/2

[/pgn]

So, Texel needs to go forward ten plies (to move 70) to produce a drawish score. Why cannot it do it at the root?

Another quibble is the current download's packaging. I've downloaded dozens of engines in the past few years from the links here and elsewhere. The packaging is mostly ZIP and RAR, with a handful of less common formats. Well, this is the first 7Z file that my older WinZIP and WinRAR versions cannot handle. I had to download a trial version of BitZipper to extract the files from the archive. May I suggest repackaging the download with a view to wider compatibility? Not everybody wants to update their utils to the latest and greatest versions because as often as not "newer" means "more bloated".

At the moment I'm running a blitz match (5m+3s) between Texel 1.07 and Fritz 15 on a 16-core box. I picked Fritz 15 because according to Peter's estimate of +35 Elo improvement over v1.06 (admittedly in self-play, thus most likely inflated) and CCRL's 40/4 rating list the gap between F15 and Texel (67 Elo for v1.06, both 4-core) may be of a size that the new version could halve if not bridge. Given the current standings that is not likely:

Code: Select all

1   Fritz 15     +63  +25/=42/-11 58.97%   46.0/78
2   Texel 1.07   -63  +11/=42/-25 41.03%   32.0/78
Elostat:

Code: Select all

Elo        : 3156
Margins    :
 68 %      : (+ 26,- 27) -> [3129,3182]
 95 %      : (+ 52,- 53) -> [3103,3208]
 99.7 %    : (+ 79,- 82) -> [3074,3235]
I'll make a thread for this match in the Tournaments and Matches subforum, complete with the final standings and a link to all the games. This was meant to be a 200-games match; 78 games took about 20 hours. I wonder if I should stop the match after 100 games or let it complete.

Sorry for contributing the sole negative note to this thread but I hope for Texel to evolve into a analysis engine superior to the top dogs.

Re: Texel 1.07

Posted: Wed Oct 04, 2017 9:17 pm
by Paloma
Very interessting,

thanks for posting this.

Re: Texel 1.07

Posted: Thu Oct 05, 2017 12:42 am
by carldaman
Interesting find, although it must be said a +1.81 score in the endgame does not automatically signify a win. Curious to see what Peter has to say.

Re: Texel 1.07

Posted: Thu Oct 05, 2017 2:54 am
by cdani
tpoppins wrote: Texel proclaims what appeared to be a winning advantage for White while the other two existing Let's Check entries for this position (by two versions of Komodo, IIRC) showed 0.00.

...

A bit odd, isn't it? SF and H6 see a draw within seconds, Texel a win for White even after 15 minutes.
Nothing strange. As the engine gets stronger, such things happen less. Andscacs also needs more time than Stockfish to find that the position is drawish. And of course this happens very often :-)

Re: Texel 1.07

Posted: Thu Oct 05, 2017 4:32 am
by tpoppins
The match is complete and the thread is up: Texel 1.07 vs. Fritz 15 / 5m+3s.

I'm aware that a 181 cp score does not necessarily signify a win, and also that one swallow does not a summer make. However, look at how Texel's eval is all over the place from +1.20 to +1.90 in the analysis above. The games from the Fritz 15 match provide more examples.

Also see game 4 of that match for a case where Texel went in for the white queen thinking it led to 0.00 and lost. It took 11 moves to see it was losing. Or game 97 where it was Texel who had the queen - it overrated its position by about 100 cp and dragged out the game almost to 100 moves before acquiescing to a draw. Or game 24: in a level R+B+pawns ending Fritz settles to 0.00 from move 68 for the rest of the game; Texel decides by move 71 that it has a 61 cp advantage and the game drags on for another 95 moves, yet draw it still is.

Re: Texel 1.07

Posted: Sat Oct 07, 2017 12:01 am
by petero2
tpoppins wrote:Texel may be a marvel in a technical sense, but how reliable is it as an analysis tool?

I've used v1.05 and 1.06 in the past on Chessbase's Let's Check and was disconcerted to see that while Texel's PVs usually matched (the first move, at least; Let's Check stores only three plies of a PV, BTW) those of SF, Houdini and Komodo, its eval was often "off" by 50 cp or more.

After installing v1.07 in the Deep Fritz 14 GUI and running it on a dozen or so positions it looked like the new version was more "conformist" than its predecessors. Then I ran into this 10-men endgame:

[d]1r6/Rb6/3k4/KP5B/P4p1P/8/8/8 b - - 0 65

Texel proclaims what appeared to be a winning advantage for White while the other two existing Let's Check entries for this position (by two versions of Komodo, IIRC) showed 0.00.
tpoppins wrote:A bit odd, isn't it? SF and H6 see a draw within seconds, Texel a win for White even after 15 minutes.
tpoppins wrote:So, Texel needs to go forward ten plies (to move 70) to produce a drawish score. Why cannot it do it at the root?
In general I don't find it surprising that texel is much worse than the top 3 engines in many positions, since the rating gap is more than 250 elo on CCRL 40/40.

There might be a few special cases where texel can produce better analysis than the top 3, such as when searching for the fastest mate in a position close to TBs where the 50-move rule plays an important role, but other than that you would be much better off using one of the top 3 engines.
tpoppins wrote:Sorry for contributing the sole negative note to this thread but I hope for Texel to evolve into a analysis engine superior to the top dogs.
I would like that too but realistically that is not likely to happen any time soon.

Thanks for the test position. I will have to analyze this in detail to understand why texel does not see the draw. So far I have found that if I run texel using 16 cores, 4GB hash, 6-men syzygy and "multipv 2" mode, it will get drawish scores much faster than when using non-multipv mode. My guess is that bad move ordering causes the critical lines to be reduced too much, and the wider search tree resulting from multipv mode compensates for this deficiency.

Re: Texel 1.07

Posted: Sat Oct 07, 2017 6:54 am
by Guenther
tpoppins wrote:
Another quibble is the current download's packaging. I've downloaded dozens of engines in the past few years from the links here and elsewhere. The packaging is mostly ZIP and RAR, with a handful of less common formats. Well, this is the first 7Z file that my older WinZIP and WinRAR versions cannot handle. I had to download a trial version of BitZipper to extract the files from the archive. May I suggest repackaging the download with a view to wider compatibility? Not everybody wants to update their utils to the latest and greatest versions because as often as not "newer" means "more bloated".
7zip is completely free and used widely nowadays. No need for Winzip
or Winrar as those are handled by the 7z program itself.
7z will probably replace zip/rar anyhow in a few years.
No idea why you tried that (never heard) 'BitZipper' instead of wellknown
free 7zip program itself?
http://www.7-zip.de/

Re: Texel 1.07

Posted: Sun Oct 08, 2017 6:59 pm
by TimoK
Hi Peter,

here are the results:

TC: 30m+5s
Texel107-Cluster [T60] - Texel107-12C: 41.5 - 28.5 (70: +13, =57, -0)
PGN

TC: 10m+3s
Texel107-Cluster [T60] - Texel107-12C: 111 - 69 (180: +43, =136, -1)
PGN

Of course too few games to make clear statements, but it seems that Texel scales well on 60 cores compared to 12 cores.

Best regards
Timo