TalkChess.com

We have released Komodo 12.1.1, which fixes a bug in the MCTS version that was in both Komodo 12 and Komodo 12.1. The consequence of the bug was that if Komodo actually checkmated the opponent (not sure about stalemate), the game was not terminated properly, which typically resulted in Komodo refusing to move when a new game was started, eventually forfeiting that game. We had noticed some forfeits but wrongly thought that they were due to a tester problem or to insufficient Overhead ms. Testers testing normal Komodo 12.1 have no reason to restart tests, but for those already testing Komodo 12.1 MCTS it is recommended to restart with 12.1.1, especially if you play games until checkmate or if you have had any time forfeits. We also made a one point change to the default setting of MCTS Explore, based on our latest testing. It's probably just a couple elo but since we were releasing a serious bugfix we decided to throw this in at the same time. It is likely that the best setting of this parameter depends on the time limit and number of threads, but this will take some time to determine.
Testers should have the new version already; subscribers and others who got 12.1 from us should have 12.1.1 in their accounts by tomorrow (Tuesday) if it is not there already.
All known bugs have now been fixed, I can't say anything about unknown ones!

Thank U Larry.
I found an error in 12.1 and mailed mark.
Looking forward to analyse with MCTS. Going to use Komodo after a long long time.
Regards
Om

Master Om wrote: ↑Tue Jun 05, 2018 3:36 am Thank U Larry.
I found an error in 12.1 and mailed mark.
Looking forward to analyse with MCTS. Going to use Komodo after a long long time.
Regards
Om

Om seems to have tracked the issue to using a BMI2 version on an AMD Ryzen chip. I know the PEXT instruction used in the BMI2 version runs very slowly on Ryzen (probably due to a microcode emulation instead of hardware like in Intel chips). This slows the program down a lot on Ryzen, so I recommend using the non-BMI2 version of Komodo. I do not know why it could cause a memory error, but I have heard some early Ryzen chips had some issues. Anyway, the non BMI2 version seems to be running fine for him. I will add a note about BMI2 and Ryzen in the Readme file we include with Komodo.

Can some one explain parameters "MCTS Hash=320" and "MCTS Explore=5"? Can we change this parameters if we have 32 or 64 Gb RAM memory?

TasicV wrote: ↑Thu Jun 07, 2018 12:53 pm Can some one explain parameters "MCTS Hash=320" and "MCTS Explore=5"? Can we change this parameters if we have 32 or 64 Gb RAM memory?

Yes you can. MCTS Hash is the amount you want to allocate for the Monte Carlo Search Tree in memory in megabytes. The default is fine for most play, but if you want to analyze for hours on many cored machines, you can raise it. this is explained in the Readme file.

MCTS Explore controls how much Komodo explores the tree instead of exploiting it. More details are give in the Readme.

Mark

lkaufman wrote: ↑Tue Jun 05, 2018 2:41 am We also made a one point change to the default setting of MCTS Explore, based on our latest testing.

Thanks for the 12 threads, Larry!

In Walter Eigenmann's Glarean magazine latest Corr.- Chess Brilliant is this one:

https://glarean-magazin.ch/2018/05/14/f ... s-bccm-07/

Clear single one game changing (from Black's losing to probably keeping the draw) best move is 1...Bxb2+!

Houdini6.03T(actical1) gets it in about 30 seconds on 24 threads, some tactically sharpened SF- and CFish- settings are about as fast and so is a tactical komodo- setting, but with default parameters almost all engines I tried needed several minutes at least.

Best default so far was for me Shredder13:

722: BCCM 07 - Alvarez-Sanakoev, Mail CAPA 1999 2018
7k/6p1/p5p1/4b3/6p1/8/PP2q1QP/K5R1 b - - 0 1

Analysis by Deep Shredder 13 x64:

1...Lxb2+ 2.Kb1 Dxg2 3.Txg2 Le5 4.Txg4 Kh7 5.Th4+ Kg8
+/- (1.28) Tiefe: 7/9 00:00:00 1kN
...
1...Lxb2+ 2.Kb1 Dxg2 3.Txg2 Le5 4.Txg4 Kh7 5.Ta4 a5 6.h3 Lc3 7.Kc2 Lb4 8.a3 Le1 9.Te4
+- (1.97) Tiefe: 12/20 00:00:00 167kN
1...Db5
+- (1.97 ++) Tiefe: 12/20 00:00:00 264kN
...
1...Db5 2.Te1 Lf6 3.De2 Db6 4.Tb1 Db4 5.Tg1 Kh7 6.Kb1 Db7 7.a3 Ld4 8.Txg4 Dh1+ 9.Ka2 Dd5+ 10.b3 Lf6 11.Tg3 Df5 12.Te3 a5 13.Tg3 Lg5 14.Tc3 Lf6 15.Te3 Ld8 16.Dg2 Lg5 17.Tc3 Lf4 18.Tf3 g5 19.Th3+ Kg8
+- (2.41 --) Tiefe: 37/67 00:01:22 2473MN
1...Lxb2+
+- (2.41 ++) Tiefe: 37/67 00:01:23 2527MN
1...Lxb2+ 2.Kb1 De3 3.Kxb2 Dd4+ 4.Kc2 Dc5+ 5.Kd3 Db5+ 6.Kc3 De5+ 7.Kc4 Db5+ 8.Kd4 Db4+ 9.Ke5 Db8+ 10.Ke6 De8+ 11.Kd6 Dd8+ 12.Kc6 Dc8+ 13.Kb6 Dd8+ 14.Kxa6 Dd3+ 15.Ka5 Da3+ 16.Kb5 Dd3+ 17.Kb4 Dd6+ 18.Ka4 Dd4+ 19.Kb3 Dd3+ 20.Kb2 Dd4+ 21.Kb1 Db6+ 22.Kc2 Dc5+ 23.Kb3 Db5+ 24.Kc3 Dc5+ 25.Kd2 Dd4+ 26.Ke1
+- (2.27 ++) Tiefe: 37/87 00:01:26 2591MN
...
1...Lxb2+ 2.Kb1 De3 3.Dxg4
+/- (1.11 ++) Tiefe: 37/95 00:01:45 3170MN
1...Lxb2+ 2.Kb1 De3 3.Kxb2 Dd4+ 4.Kc2 Dc5+ 5.Kd3 Db5+ 6.Kc3 De5+ 7.Kc4 Dc7+ 8.Kd4 Db6+ 9.Kd5
+/= (0.64 ++) Tiefe: 37/95 00:01:53 3390MN
1...Lxb2+ 2.Kb1 De3 3.Dxg4 Le5 4.Dh4+ Kg8 5.Dc4+ Kh7 6.Tg4 Lf6 7.Tg3 De1+ 8.Kc2 Df2+ 9.Kd1 Dxh2 10.Dg4 Dh1+ 11.Ke2 Kg8 12.Dxg6 Dh2+ 13.Kd3 Kf8 14.Te3 Db8 15.Ke2 Dh2+ 16.Kf3 Dh1+ 17.Dg2 Dc1 18.Te2 Db1 19.Dg4 Df1+ 20.Ke3 g5 21.Db4+ Kf7 22.Dc4+ Kf8 23.Kd2 Df5 24.De4 Dc5 25.Kd1 Kg7 26.Tg2 Dc3
+/= (0.69) Tiefe: 37/95 00:04:01 7240MN

But now look here, what komodo12.1.1 with MCTS can do:

7k/6p1/p5p1/4b3/6p1/8/PP2q1QP/K5R1 b - - 0 1

Analysis by Komodo 12.1.1 64-bitM(onte Carlo Tree Search on):

1...Db5 2.Te1 Ld4 3.Dd2 Lf6 4.De2 Db7 5.Tb1 Kh7 6.Dxg4 Db5
+/- (1.11) Tiefe: 22/14 00:00:03 44kN
...
1...Db5 2.Te1 Lf6 3.De2 Db6 4.Kb1 Kh7 5.a3 Dd4 6.Tc1 Le5 7.Tc4 Dg1+ 8.Ka2 Dxh2 9.Dxh2+ Lxh2 10.Txg4 Le5
+/- (1.43) Tiefe: 27/19 00:00:24 506kN
1...Lxb2+ 2.Kb1 De3 3.Dxg4 Le5 4.Dh4+ Kg8 5.Dc4+ Kh7 6.Tg4 Lf6 7.Tg3 De1+ 8.Kc2 Df2+ 9.Kb3 Dxh2 10.Te3 Dh1 11.Ka3 a5 12.Td3
+/- (0.72) Tiefe: 28/24 00:00:27 535kN
...
1...Lxb2+ 2.Kb1 De3 3.Dxg4 Le5 4.Dh4+ Kg8 5.Dc4+ Kh7 6.Tg4 De1+ 7.Kc2 Df2+ 8.Kd1 Df3+ 9.De2 Dh3 10.a3 Lc7 11.Kc2 Le5 12.De4 Dxh2+ 13.Tg2 Dh5 14.Tg1 Df5 15.Th1+ Kg8 16.Te1 Lf6
+/= (0.55) Tiefe: 28/38 00:00:54 743kN
1...Lxb2+ 2.Kb1 De3 3.Dxg4 Le5 4.Dh4+ Kg8 5.Dc4+ Kh7 6.Tg4 De1+ 7.Kc2 Df2+ 8.Kd1 Df3+ 9.De2 Dh3 10.De4 Df1+ 11.Kc2 Df2+ 12.Kc1 Db2+ 13.Kd1 Da1+ 14.Ke2 Dxa2+ 15.Kf1 Da1+ 16.De1 Dxe1+ 17.Kxe1 a5 18.Ke2 Kh6 19.Kd3 g5 20.h3
+/= (0.56) Tiefe: 28/39 00:00:57 776kN
...
1...Lxb2+ 2.Kb1 De3 3.Dxg4 Ld4 4.Tg3 De1+ 5.Kc2 Le3 6.Dd1 Df2+ 7.Kd3 Lh6 8.De2 Db6 9.Kc2 Kh7 10.a3 Dd4 11.Kb1 Df6 12.Dc2 Df1+ 13.Ka2
+/- (0.73) Tiefe: 31/59 00:03:00 2749kN
...
1...Lxb2+ 2.Kb1 De3 3.Dxg4 Ld4 4.Tg3 De1+ 5.Kc2 Le3 6.Dd1 Df2+ 7.Kd3 Lh6 8.De2 Db6 9.Kc2 Kh7 10.a3 Dd6 11.Td3 Df6 12.Kb3 Df7+ 13.Kb2 Df5 14.Td1 Df6+ 15.Kb3
+/- (0.77) Tiefe: 33/59 00:04:24 6073kN

Gave it two more tries, one about the same as quick, third one even better with 18" to stable output on solution with good eval.
It doesn't have to become fully =, I'd say, += to +/- is ok, Walter Eigenmann comments it in his .pgn like this too (+=).

12 threads of the 12core- CPU is only half power here, hyperthreading was on.

Great tactical achievement of the engine with MCTS here, k12.1.1 default didn't "find" 1...Bxb2+ at a single try within 10 minutes, even if it arises early in output, as this is with most other engines I tried as well, but disappears again as long as 2...Qe3! isn't there too.

I just recently purchased a new laptop with an i7-8550U CPU. The Komodo BMI (64bit) version does not seem to work correctly and displays very low depth. The Non-BMI version displays proper depth info. Should not the BMI version be the better version for my CPU? PS: Chessbase GUI & Windows 10 Pro.

peter wrote: ↑Fri Jun 08, 2018 9:37 am

lkaufman wrote: ↑Tue Jun 05, 2018 2:41 am We also made a one point change to the default setting of MCTS Explore, based on our latest testing.

Thanks for the 12 threads, Larry!


12 threads of the 12core- CPU is only half power here, hyperthreading was on.

Great tactical achievement of the engine with MCTS here, k12.1.1 default didn't "find" 1...Bxb2+ at a single try within 10 minutes, even if it arises early in output, as this is with most other engines I tried as well, but disappears again as long as 2...Qe3! isn't there too.

Thanks. Just one comment: using 12 threads on a 12 core with hyperthreading on is not "half power"; it is probably the optimum setting for any chess program even with hyperthreading on. If you have just a quad, it probably pays to run six threads, because six is a lot better than four even with say a 20% or so slowdown, but chess engines gain much less going from 12 to 18 threads than from 4 to 6 so it wouldn't be likely to pay off given the 20% or so slowdown per thread. We may increase the 12 thread limit in the future, but it won't help you.

AdminX wrote: ↑Fri Jun 08, 2018 5:43 pm I just recently purchased a new laptop with an i7-8550U CPU. The Komodo BMI (64bit) version does not seem to work correctly and displays very low depth. The Non-BMI version displays proper depth info. Should not the BMI version be the better version for my CPU? PS: Chessbase GUI & Windows 10 Pro.

I would expect it to be better, so we will have to investigate. But Mark just left for vacation until June 19. In any case, the benefit of BMI for Komodo is less than a 2% speedup, so it's not a big deal.