hard test position for NNUE engines

[Uri Blass]

[d]2k5/4p2r/3q2pP/p1p3P1/1pP1pP2/1P1rP3/PQ1P4/2KR3R w - - 0 1 bm Qg7

NNUE engine prefer worse winning moves like Qe5 that are clearly slower to mate when every old non NNUE engine can find Qg7 based ion my experience.

I wonder if there are NNUE engine that can find good winning moves that win more material based on short search and in most cases also mate faster or all of them fail(and even if some NNUE engine can find Qg7 after a long search I consider it to be a failure because it is a move that I expect engines to see in less than one second.

After some backward analysis I got mate in at most 21 moves for Qg7

43/41 00:00 5,533k 5,764k +M21 Qb2-g7 Rh7xg7 h6xg7 Qd6-e6 Rh1-h8+ Kc8-d7 g7-g8Q Qe6xg8 Rh8xg8 Rd3-d6 Rd1-h1 a5-a4 b3xa4 b4-b3 a2xb3 Kd7-c7 Rh1-h7 Kc7-d7 a4-a5 Rd6-e6 Rg8-f8 Re6-d6 Rf8-f7 Kd7-c6 Rf7xe7 Rd6-d7 a5-a6 Rd7xe7 Rh7xe7 Kc6-d6 Re7-h7 Kd6-e6 a6-a7 Ke6-f5 a7-a8Q Kf5-g4 Rh7-h2 Kg4-f3 Qa8-h8 Kf3-g4 Qh8-h3+

[peter]

Started with empty hash of 32768Mb and 30 threads of 4.3GHz CPU:

Code: Select all

2k5/4p2r/3q2pP/p1p3P1/1pP1pP2/1P1rP3/PQ1P4/2KR3R w - -

Engine: SlowChess2.9 (32768 MB)
von Jonathan Kreuzer

21/39  0:01   +12.72   1.Dg7 Txg7 2.hxg7 De6 3.Th8+ Kc7 
                       4.g8D Dxg8 5.Txg8 a4 6.bxa4 Ta3 
                       7.Kb2 Txa4 8.Txg6 Kd7 9.Tb6 Ta8 
                       10.Tb5 e5 11.Txc5 exf4 12.exf4 Ta3 
                       13.Td5+ Ke7 14.Tb1 (62.998.410) 38887 

...

29/56  0:34   +22.21   1.Dg7 Txg7 2.hxg7 De6 3.Th8+ Kc7 
                       4.g8D Dxg8 5.Txg8 a4 6.bxa4 Ta3 
                       7.Kb2 Txa4 8.Txg6 Ta8 9.Th6 Kd7 
                       10.Tg1 Tg8 11.g6 Ke6 12.g7+ Kf7 
                       13.Th8 Txg7 14.Txg7+ (1.401.497.263) 41196 

30/67  1:41   +M249++  1.Dg7 (3.876.411.857) 38097 

...

31/67  2:53   +M26     1.Dg7 Txg7 2.hxg7 De6 3.Th8+ Kc7 
                       4.g8D Dxg8 5.Txg8 a4 6.bxa4 b3 
                       7.axb3 Txb3 8.Txg6 Tb4 9.Ta6 Txc4+ 
                       10.Kb2 Tb4+ 11.Kc3 Kb7 12.Te6 Kc7 
                       13.Th1 Kd7 14.Ta6 (6.344.115.429) 36656 

32/67  3:12   +M25     1.Dg7 Txg7 2.hxg7 De6 3.Th8+ Kd7 
                       4.g8D Dxg8 5.Txg8 Td6 6.Th1 a4 
                       7.bxa4 Tb6 8.Th6 Ta6 9.Th7 Tb6 
                       10.Tgg7 Te6 11.a5 b3 12.axb3 Ke8 
                       13.Tg8+ Kd7 14.Tf8 (6.998.868.051) 36414 

33/67  3:23   +M24     1.Dg7 Txg7 2.hxg7 De6 3.Th8+ Kc7 
                       4.g8D Dxg8 5.Txg8 a4 6.bxa4 Ta3 
                       7.Kb2 Txa4 8.Txg6 Ta8 9.Th6 Tg8 
                       10.g6 Kd7 11.Tg1 Ke6 12.Th7 Kf6 
                       13.Tf7+ Ke6 14.f5+ (7.391.316.117) 36269 

...

39/67  3:25   +M24     1.Dg7 Txg7 2.hxg7 De6 3.Th8+ Kc7 
                       4.g8D Dxg8 5.Txg8 a4 6.bxa4 Ta3 
                       7.Kb2 Txa4 8.Txg6 Ta8 9.Th6 Tg8 
                       10.g6 Kd7 11.Tg1 Ke6 12.Th7 Kf6 
                       13.Tf7+ Ke6 14.f5+ (7.452.937.815) 36256 

Getting an at least 3 moves too long DTM after 5 minutes still, I didn't want to wait for #21- output. Anyhow, I guess, time to mate in x- output is a question of search more than a one of eval, finding 1.Qg7 initially could well be the other way round, seeing Huntsman taking some time without NNUE (with MateSearch on). This on (Huntsman) with some FW- BW confirms #21 back to position after full first move but get's stuck (even after extra- FW-BW after this one alternative 1st move of Black's) with 1...a4 and 2 moves longer DTM:

[d]2k5/4p1Qr/3q2pP/2p3P1/ppP1pP2/1P1rP3/P2P4/2KR3R w - - 0 1

Code: Select all

Analysis by The Huntsman 1:

2.Dxh7 axb3 3.axb3 Txb3 4.Dh8+ Kb7 5.De5 Da6 6.Dxe7+ Kb8 7.Dd8+ Kb7 8.Dd7+ Ka8 9.Dc6+ Dxc6 10.h7 Td3 11.h8D+ Kb7 12.Th7+ Td7 13.Tdh1 Dd6 14.Txd7+ Dxd7 15.Th7 Kc7 16.f5 gxf5 17.Txd7+ Kxd7 18.g6 f4 19.g7 fxe3 20.g8D exd2+ 21.Kd1 e3 22.Dhg7+ Kc6 23.De6#
 Tiefe: 65/44 00:01:38 6785MN
 +- (#22)

Edit: SlowChess in its own GUI with new FW- BW yet gets #21 for 1...a4 too:

Code: Select all

SlowChess Blitz 2.9 avx2  (32768 MB hash, 31 threads)
FEN: 2k5/4p1Qr/3q2pP/2p3P1/ppP1pP2/1P1rP3/P2P4/2KR3R w - - 0 2

[43]  M20  00:14.4  492.10M  Qxh7  axb3  Qh8+  Kb7  Qb2  bxa2  Qxa2  Ra3  Qc2  Kc7  Qxe4  Ra1+  Kb2  Ra2+  Kxa2  Qa6+  Kb3  Qa3+  Kc2  Qa4+  Kd3  Qd7+  Ke2  Kb6  h7  Qg4+  Kf2  Qf5  Qxf5  gxf5  Rh6+  e6  h8=Q  Kc6  Qc8+  Kb6  Rxe6+  Ka5  Qa8#  

At very starting position:

Code: Select all

SlowChess Blitz 2.9 avx2  (32768 MB hash, 31 threads)
FEN: 2k5/4p2r/3q2pP/p1p3P1/1pP1pP2/1P1rP3/PQ1P4/2KR3R w - - 0 1

[39]  M21  00:25.7  1031.73M  Qg7  Rxg7  hxg7  Qd8  Rh8  Kc7  g8=Q  Qxg8  Rxg8  Rd6  Rh1  Re6  Rh6  Kd7  Rhxg6  Rxg6  Rxg6  Ke8  Ra6  Kf7  g6+  Kg7  f5  a4  Rxa4  Kh6  Ra7  Kg7  Rxe7+  Kf6  Rf7+  Kg5  g7  Kg4  g8=Q+  Kf3  Qg1  Ke2  Rg7  Kf3  Qg2#