Test Position: Sion Castro-Rivera, Cordoba 1991

[kgburcham]

[d] r4rk1/1bq1bppp/p1npp3/2nB1NP1/1p2P3/2N1B3/PPP1QP1P/2KR2R1 w - -

Engine: Stockfish 121015 64 BMI2 (8192 MB)
by Tord Romstad, Marco Costalba and Joona


30/47 1:09 -0.56 16.Bxc6 Bxc6 17.Nxe7+ Qxe7 18.Bxc5 dxc5 19.Nb1 a5 20.Qe3 c4 21.c3 Qc7 22.Rd4 Qxh2 23.Qg3 Qh5 24.Rxc4 Rac8 25.Re1 Bb5 26.Rxc8 Rxc8 27.b3 h6 28.c4 Qxg5+ 29.Qxg5 (1.063.270.766) 15375
31/47 1:45 -0.50++ 16.Qh5 (1.627.702.966) 15413
31/47 2:00 -0.43++ 16.Qh5 (1.863.032.360) 15442
31/47 2:11 -0.34++ 16.Qh5 (2.029.331.159) 15463
31/47 2:21 -0.20++ 16.Qh5 (2.191.857.278) 15486
31/47 2:34 0.00++ 16.Qh5 (2.390.548.048) 15512
31/47 2:53 +0.31++ 16.Qh5 (2.703.084.243) 15572
31/47 3:19 +0.78++ 16.Qh5 (3.118.128.261) 15601
31/47 3:58 +1.10 16.Qh5 Ne5 17.Rg3 exf5 18.exf5 Ncd3+ 19.Rxd3 Nxd3+ 20.Kb1 Bxd5 21.Nxd5 Qb7 22.Nxe7+ Qxe7 23.Rh3 h6 24.cxd3 Qb7 25.gxh6 Qh1+ 26.Bc1 Rac8 27.Qg5 Rxc1+ 28.Qxc1 Qxc1+ 29.Kxc1 (3.722.353.054) 15608
32/47 4:19 +1.16++ 16.Qh5 (4.048.762.956) 15621
32/48 4:33 +1.22++ 16.Qh5 (4.278.047.206) 15632
32/48 4:40 +1.18 16.Qh5 Ne5 17.Rg3 exf5 18.exf5 Ncd3+ 19.Rxd3 Nxd3+ 20.Kb1 Bxd5 21.Nxd5 Qb7 22.Nxe7+ Qxe7 23.Rh3 h6 24.cxd3 Qb7 25.gxh6 Qh1+ 26.Bc1 Rac8 27.Qg5 Rxc1+ 28.Qxc1 Qxc1+ 29.Kxc1 (4.388.774.004) 15627
33/51 5:13 +1.24++ 16.Qh5 (4.908.254.975) 15657
33/54 5:40 +1.30++ 16.Qh5 (5.328.965.914) 15659
best move: Qe2-h5 time: 6:04.469 min n/s: 15.659.013 nodes: 5.681.792.755



[Event "Cordoba op"]
[Site "Cordoba"]
[Date "1991.??.??"]
[Round "?"]
[White "Sion Castro, Marcelino"]
[Black "Rivera Kuzawka, Daniel"]
[WhiteElo "2405"]
[BlackElo "2320"]
[ECO "B89"]
[Result "1-0"]

1. e4 c5 2. Nf3 Nc6 3. d4 cxd4 4. Nxd4 Nf6 5. Nc3 d6 6. Bc4
e6 7. Be3 Be7 8. Qe2 a6 9. O-O-O Qc7 10. Bb3 O-O 11. Rhg1
Nd7 12. g4 Nc5 13. Nf5 b5 14. Bd5 Bb7 15. g5 b4 16. Qh5 Ne5
17. Rg3 exf5 18. exf5 bxc3 19. g6 cxb2+ 20. Kb1 hxg6
21. fxg6 Nxg6 22. Qxg6 Ne6 23. Bxe6 Bf6 24. Qxf6 fxe6
25. Qxe6+ Qf7 26. Rxg7+ 1-0

[Eelco de Groot]

Thank you, Eelco!

I have downloaded the files and successfully compiled the engine for 64-bit Linux. It seems to work normally.

Your work, and your answer to me, is much appreciated.

I am glad someone had some fun with this Marek. Thank you and thanks everyone who downloaded or took a look! I have not heard anything back from Damir who reported an exception caused by Rainbow Serpent, probably engine itself crashed, but I have not seen that myself, at least not with all recent versions 001 - 009, 010 beta crashed immediately in the bench but that part of the code was rewritten for the final 010. If anyone sees an exception from engine please make some notes under what circumstances?

The Rainbow Serpent 20150919_001 code as deposited on TDBB is probably ripe for an overhaul, I think one of the later versions; 004, 007 or 011 is a more interesting one now. I am not saying stronger, probably though, because the Singular Extensions were taking up too much resources, the higher the depths used. The SE have been slimmed down and, in fact, in version 11, they have been completely removed

The Singular Extensions were already of a hybrid approach, so that non singular nodes would be reduced and singular ones extended. This gave a nice effect of, in practice, singular extensions that were greater than one In PV nodes no reductions were done. There is something to be said though, for a rule that if you can write an extension as a reduction, you should do so. You only need one set of control mechanisms and there is no danger anymore of search explosions. It greatly changes numbers of the search, even if practical effects are limited. Time to ply for instance can for instance, all else staying the same, be drastically reduced.

Example:

Code: Select all

Stockfish 20150919_004
based on _002
===========================
Total time (ms) : 13031
Nodes searched  : 12204227
Nodes/second    : 936553

Hybrid, both singular extensions and non singular reductions (also known as cheap MultiCut), by far the highest nodes per second reached at time of testing, but a large number of nodes to complete depth 13 on all bench positions. The singular extensions were already trimmed down here but surprisingly I could not see that effect in the bench! 004 had roughly the same benchnumbers as 001.

Code: Select all

Stockfish 20150919_009
based on _004
===========================
Total time (ms) : 4835
Nodes searched  : 3947585
Nodes/second    : 816460


Stockfish 20150919_010
based on _004
===========================
Total time (ms) : 7317
Nodes searched  : 6197342
Nodes/second    : 846978

Two versions with all singular extensions turned into reductions. The second version also does non singular reductions in PV nodes but first did an expensive test if the TTmove at least reached alpha in a PV node. This was so costly that the bench actually was higher instead of lower. Still interesting version though, as is 009.

Code: Select all

Stockfish 20150919_011
based on _004
==========================
Total time (ms) : 4310
Nodes searched  : 3452241
Nodes/second    : 800983

Same features as 010, but a cheaper test and some rewriting, there is not much difference seen in bench though with 009 at this depth. I think reducing even in PV nodes (No extensions!) should make much more difference as depth gets higher though. This is the most recent version.

Not too make things too complex but without much testing, it is probably best to just pick one right now, not several and in that case probably 011 is the more interesting. At high depths it is guaranteed to have much better branching factor (time to depth) than the 001 that is now on TDBB. Not that that always is better, but in this case it is, very likely. Probably tomorrow I will post the 011. Just wanted to explain some of the differences here already.

Thank you for reading, if you have come this far !

Eelco

[Vinvin]

Stockfish has regressed a lot on this position !

It finds no more Qh5!! with a big advantage for white

FEN: r4rk1/1bq1bppp/p1npp3/2nB1NP1/1p2P3/2N1B3/PPP1QP1P/2KR2R1 w - - 0 1

Code: Select all

Stockfish_19081422_x64_modern:
...
 41/64	10:13	4 622 907 059	7 537 807	-0,82	1.Nxe7+ Qxe7 2.Bxc6 Bxc6 3.Bxc5 dxc5 4.Nb1 a5 5.Qe3 a4 6.a3 Rfb8 7.h4 Ra5 8.Nd2 bxa3 9.bxa3 Rab5 10.h5 Qd6 11.h6 Qd4 12.Qxd4 cxd4 13.Rg4 e5 14.Nf3 Rc5 15.hxg7 Kxg7 16.Rh4 Rb3 17.Nxe5 Rxa3 18.Nxc6 Rxc6 19.Rd2 d3 20.Rh3 Ra1+ 21.Kb2 dxc2 22.Rxc2 a3+ 23.Kxa1 Rxc2 24.Rxa3 Rxf2 25.Kb1 Rg2 26.Ra7 Rxg5
 42/59-	10:36	4 804 250 391	7 542 207	-0,92	1.Nxe7+ Qxe7
 42/59-	10:49	4 899 864 766	7 543 429	-1,03	1.Nxe7+ Qxe7

[zullil]

Stockfish has regressed a lot on this position !

It finds no more Qh5!! with a big advantage for white

FEN: r4rk1/1bq1bppp/p1npp3/2nB1NP1/1p2P3/2N1B3/PPP1QP1P/2KR2R1 w - - 0 1

Code: Select all

Stockfish_19081422_x64_modern:
...
 41/64	10:13	4 622 907 059	7 537 807	-0,82	1.Nxe7+ Qxe7 2.Bxc6 Bxc6 3.Bxc5 dxc5 4.Nb1 a5 5.Qe3 a4 6.a3 Rfb8 7.h4 Ra5 8.Nd2 bxa3 9.bxa3 Rab5 10.h5 Qd6 11.h6 Qd4 12.Qxd4 cxd4 13.Rg4 e5 14.Nf3 Rc5 15.hxg7 Kxg7 16.Rh4 Rb3 17.Nxe5 Rxa3 18.Nxc6 Rxc6 19.Rd2 d3 20.Rh3 Ra1+ 21.Kb2 dxc2 22.Rxc2 a3+ 23.Kxa1 Rxc2 24.Rxa3 Rxf2 25.Kb1 Rg2 26.Ra7 Rxg5
 42/59-	10:36	4 804 250 391	7 542 207	-0,92	1.Nxe7+ Qxe7
 42/59-	10:49	4 899 864 766	7 543 429	-1,03	1.Nxe7+ Qxe7

My Cfish-dev failed even to depth 46, but this one is easy for Stockfish-Crystal:

Code: Select all

info depth 30 seldepth 66 multipv 1 score cp -42 lowerbound nodes 1922605629 nps 30006018 hashfull 550 tbhits 0 time 64074 pv e2h5
info depth 30 currmove e2h5 currmovenumber 1
info depth 30 seldepth 66 multipv 1 score cp -13 lowerbound nodes 2040264428 nps 30018456 hashfull 571 tbhits 0 time 67967 pv e2h5
info depth 30 currmove e2h5 currmovenumber 1
info depth 30 seldepth 66 multipv 1 score cp 25 lowerbound nodes 2239444563 nps 30004482 hashfull 615 tbhits 0 time 74637 pv e2h5
info depth 30 currmove e2h5 currmovenumber 1
info depth 30 seldepth 66 multipv 1 score cp 75 lowerbound nodes 2621185444 nps 29983132 hashfull 676 tbhits 0 time 87422 pv e2h5
info depth 30 currmove e2h5 currmovenumber 1
info depth 30 seldepth 66 multipv 1 score cp 140 lowerbound nodes 3041061925 nps 29926998 hashfull 736 tbhits 0 time 101616 pv e2h5
info depth 30 currmove e2h5 currmovenumber 1
info depth 30 seldepth 102 multipv 1 score cp 223 lowerbound nodes 3738588176 nps 30008814 hashfull 803 tbhits 0 time 124583 pv e2h5