The “Kryptonite” position

Discussion of anything and everything relating to chess playing software and machines.

Moderators: Harvey Williamson, Dann Corbit, hgm

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
Post Reply
dkappe
Posts: 571
Joined: Tue Aug 21, 2018 5:52 pm
Full name: Dietrich Kappe

The “Kryptonite” position

Post by dkappe » Sat Nov 30, 2019 11:14 pm

Scorpio just lost in a real howler to Komodo at CCCC. Here the key position:



Here Scorpio played Kxh4??

I was able to reproduce this, but the amazing thing is that almost all of the leela nets I tried also played this howler —- 11258, 42850, t60, LD2, t58, Bad Gyal 8, etc., and only saw the light after 30M+ nodes or many more. The only exceptions were Good Gyal 7, which got it after 2.5M nodes and Dark Queen 2.1, which took only 150k nodes to discard the move. Both of these last two nets rely heavily of SF’s eval rather than self-play.

I’m adding this to my endgame test suite. Not hard for an ab engine, but kryptonite for an nn engine.

jp
Posts: 1411
Joined: Mon Apr 23, 2018 5:54 am

Re: The “Kryptonite” position

Post by jp » Sun Dec 01, 2019 12:37 pm

What PV does Leela give when it choose Kxh4?

zullil
Posts: 6442
Joined: Mon Jan 08, 2007 11:31 pm
Location: PA USA
Full name: Louis Zulli

Re: The “Kryptonite” position

Post by zullil » Sun Dec 01, 2019 5:02 pm

jp wrote:
Sun Dec 01, 2019 12:37 pm
What PV does Leela give when it choose Kxh4?
Here's the portion of Lc0's log file before and after it realizes the Kxh4 loses (at about 2.25 M nodes and 30 s). This is J13B.3-200.

Code: Select all

1201 11:54:14.551671 140330957842176 ../../src/chess/uciloop.cc:218] << info depth 12 seldepth 30 time 26411 nodes 1872742 score cp 20 hashfull 1 nps 70907 tbhits 95433 multipv 1 pv g4h4 e5d5 h4g4 d5e6 g4g5 e6
f7 g5h6 f7f6 h3h4 f6f5 g2g3 c7a6 h6g7 a6c7 g7f7 c7a6 f7e7 f5g4 e7f6 g4g3
1201 11:54:14.551782 140330957842176 ../../src/chess/uciloop.cc:218] << info depth 12 seldepth 30 time 26411 nodes 1872742 score cp 19 hashfull 1 nps 70907 tbhits 95433 multipv 2 pv d5d6 e5d6 g4h4 c7e6 h4g4 d6
c5 g4f3 c5b5 f3e4 b5a5
1201 11:54:14.551797 140330957842176 ../../src/chess/uciloop.cc:218] << info depth 12 seldepth 30 time 26411 nodes 1872742 score cp 16 hashfull 1 nps 70907 tbhits 95433 multipv 3 pv g4g5 c7d5 a5a6 d5c7 a6a7 e5
e4 g5g6 e4f4 g6h5 f4g3 h5g5 c7a8 g5f5 a8c7 f5e5 g3g2
1201 11:54:14.551806 140330957842176 ../../src/chess/uciloop.cc:218] << info depth 12 seldepth 30 time 26411 nodes 1872742 score cp 2 hashfull 1 nps 70907 tbhits 95433 multipv 4 pv g4f3 e5d5 f3f4 c7e6 f4g4 d5c
6 g4h4 c6b5 h4g4 b5a5
1201 11:54:14.551815 140330957842176 ../../src/chess/uciloop.cc:218] << info depth 12 seldepth 30 time 26411 nodes 1872742 score cp 5 hashfull 1 nps 70907 tbhits 95433 multipv 5 pv g2g3 h4g3 g4g3 e5d5
1201 11:54:14.551824 140330957842176 ../../src/chess/uciloop.cc:218] << info depth 12 seldepth 30 time 26411 nodes 1872742 score cp -37 hashfull 1 nps 70907 tbhits 95433 multipv 6 pv a5a6 c7a6 g4g5 e5e4 g5f6 e
4f4 f6g6 a6c5 d5d6 f4g3 g6h5 c5d7 h5g5 d7e5 g5h5 e5f7
1201 11:54:14.551882 140330957842176 ../../src/chess/uciloop.cc:218] << info string a5a6  (913 ) N:    6677 (+ 0) (P:  8.12%) (Q: -0.12683) (D:  0.833) (U: 0.20188) (Q+U:  0.07506) (V: -0.0009) 
1201 11:54:14.551894 140330957842176 ../../src/chess/uciloop.cc:218] << info string g2g3  (374 ) N:   21033 (+ 0) (P:  7.24%) (Q:  0.01792) (D:  0.959) (U: 0.05720) (Q+U:  0.07512) (V: -0.0011) 
1201 11:54:14.551902 140330957842176 ../../src/chess/uciloop.cc:218] << info string g4f3  (849 ) N:   21410 (+ 0) (P:  8.68%) (Q:  0.00778) (D:  0.924) (U: 0.06735) (Q+U:  0.07513) (V:  0.0054) 
1201 11:54:14.551910 140330957842176 ../../src/chess/uciloop.cc:218] << info string g4g5  (861 ) N:  227117 (+ 0) (P: 25.87%) (Q:  0.05620) (D:  0.837) (U: 0.01891) (Q+U:  0.07512) (V:  0.0698) 
1201 11:54:14.551918 140330957842176 ../../src/chess/uciloop.cc:218] << info string d5d6  (1007) N:  358297 (+ 0) (P: 20.22%) (Q:  0.06574) (D:  0.917) (U: 0.00937) (Q+U:  0.07511) (V:  0.0166) 
1201 11:54:14.551926 140330957842176 ../../src/chess/uciloop.cc:218] << info string g4h4  (858 ) N: 1238207 (+12) (P: 29.86%) (Q:  0.07113) (D:  0.917) (U: 0.00400) (Q+U:  0.07513) (V:  0.4410) 
1201 11:54:19.555215 140330966234880 ../../src/chess/uciloop.cc:218] << info depth 12 seldepth 30 time 31415 nodes 2258517 score cp -12800 hashfull 1 nps 71892 tbhits 115703 multipv 1 pv g4h4 e5d5 h4g4 d5e6 g4
g5 e6f7 g5h6 f7f6 h3h4 f6f5 g2g3 c7a6 h6g7 a6c7 g7f7 c7a6 f7e7 f5g4 e7f6 g4g3
1201 11:54:19.555291 140330966234880 ../../src/chess/uciloop.cc:218] << info depth 12 seldepth 30 time 31415 nodes 2258517 score cp 18 hashfull 1 nps 71892 tbhits 115703 multipv 2 pv d5d6 e5d6 g4h4 c7e6 h4g4 d
6c5 g4f3 c5b5 f3e4 b5a5
1201 11:54:19.555317 140330966234880 ../../src/chess/uciloop.cc:218] << info depth 12 seldepth 30 time 31415 nodes 2258517 score cp 15 hashfull 1 nps 71892 tbhits 115703 multipv 3 pv g4g5 c7d5 a5a6 d5c7 a6a7 e
5e4 g5g6 e4f4 g6h5 f4g3 h5g5 c7a8 g5f5 a8c7 f5e5 g3g2
1201 11:54:19.555327 140330966234880 ../../src/chess/uciloop.cc:218] << info depth 12 seldepth 30 time 31415 nodes 2258517 score cp 5 hashfull 1 nps 71892 tbhits 115703 multipv 4 pv g2g3 h4g3 g4g3 e5d5
1201 11:54:19.555340 140330966234880 ../../src/chess/uciloop.cc:218] << info depth 12 seldepth 30 time 31415 nodes 2258517 score cp 2 hashfull 1 nps 71892 tbhits 115703 multipv 5 pv g4f3 e5d5 f3f4 c7e6 f4g4 d5
c6 g4h4 c6b5 h4g4 b5a5
1201 11:54:19.555349 140330966234880 ../../src/chess/uciloop.cc:218] << info depth 12 seldepth 30 time 31415 nodes 2258517 score cp -39 hashfull 1 nps 71892 tbhits 115703 multipv 6 pv a5a6 c7a6 g4g5 e5e4 g5g4 
a6c5 g4g5 c5d7 g5g6 e4f4 d5d6 f4g3 g6g5

jp
Posts: 1411
Joined: Mon Apr 23, 2018 5:54 am

Re: The “Kryptonite” position

Post by jp » Mon Dec 02, 2019 7:33 am

zullil wrote:
Sun Dec 01, 2019 5:02 pm
Here's the portion of Lc0's log file before and after it realizes the Kxh4 loses (at about 2.25 M nodes and 30 s). This is J13B.3-200.

1201 11:54:14.551926 140330957842176 ../../src/chess/uciloop.cc:218] << info string g4h4 (858 ) N: 1238207 (+12) (P: 29.86%) (Q: 0.07113) (D: 0.917) (U: 0.00400) (Q+U: 0.07513) (V: 0.4410)
1201 11:54:19.555215 140330966234880 ../../src/chess/uciloop.cc:218] << info depth 12 seldepth 30 time 31415 nodes 2258517 score cp -12800 hashfull 1 nps 71892 tbhits 115703 multipv 1 pv g4h4 e5d5 h4g4 d5e6 g4
g5 e6f7 g5h6 f7f6 h3h4 f6f5 g2g3 c7a6 h6g7 a6c7 g7f7 c7a6 f7e7 f5g4 e7f6 g4g3
I see, so it's stuck with the reply Kxd4? at that point and doesn't see the mate in 6.

User avatar
MikeB
Posts: 4489
Joined: Thu Mar 09, 2006 5:34 am
Location: Pen Argyl, Pennsylvania

Re: The “Kryptonite” position

Post by MikeB » Fri Jan 10, 2020 12:52 pm

dkappe wrote:
Sat Nov 30, 2019 11:14 pm
Scorpio just lost in a real howler to Komodo at CCCC. Here the key position:



Here Scorpio played Kxh4??

I was able to reproduce this, but the amazing thing is that almost all of the leela nets I tried also played this howler —- 11258, 42850, t60, LD2, t58, Bad Gyal 8, etc., and only saw the light after 30M+ nodes or many more. The only exceptions were Good Gyal 7, which got it after 2.5M nodes and Dark Queen 2.1, which took only 150k nodes to discard the move. Both of these last two nets rely heavily of SF’s eval rather than self-play.

I’m adding this to my endgame test suite. Not hard for an ab engine, but kryptonite for an nn engine.
lc0 with a sergio net

Code: Select all

dep	score	nodes	time	(not shown:  tbhits	knps	seldep)
 14	+0.07 	9.53M  	1:43.68	d6 Kxd6 Kxh4 Ne6 Kg3 Kc5 Kf3 Kb5 Ke4 Kxa5 
 14	+0.07 	9.18M  	1:38.67	d6 Kxd6 Kxh4 Ne6 Kg3 Kc5 Kf3 Kb5 Ke4 Kxa5 
 14	+0.07 	8.86M  	1:33.66	d6 Kxd6 Kxh4 Ne6 Kg3 Kc5 Kf3 Kb5 Ke4 Kxa5 
 14	+0.08 	8.45M  	1:28.65	d6 Kxd6 Kxh4 Ne6 Kg3 Kc5 Kf3 Kb5 Ke4 Kxa5 
 14	+0.08 	8.03M  	1:23.64	d6 Kxd6 Kxh4 Ne6 Kg3 Kc5 Kf3 Kb5 Ke4 Kxa5 
 14	+0.08 	7.47M  	1:18.64	d6 Kxd6 Kxh4 Ne6 Kg3 Kc5 Kf3 Kb5 Ke4 Kxa5 
 14	+0.08 	7.43M  	1:18.31	d6 Kxd6 Kxh4 Ne6 Kg3 Kc5 Kf3 Kb5 Ke4 Kxa5 
 14	+0.07 	7.47M  	1:18.64	Kg5 Nxd5 a6 Nc7 a7 Ke4 Kxg6 Kf4 Kf6 Kg3 Kg5 Na8 Kf5 Nc7 Kg5 Na8 Kf5 Nc7 Kg5 
 14	+0.07 	7.19M  	1:16.11	Kg5 Nxd5 a6 Nc7 a7 Ke4 Kxg6 Kf4 Kf6 Kg3 Kg5 Na8 Kf5 Nc7 Kg5 Na8 Kf5 Nc7 Kg5 
 14	+0.08 	6.80M  	1:11.11	Kg5 Nxd5 a6 Nc7 a7 Ke4 Kxg6 Kf4 Kf6 Kg3 Kg5 Na8 Kf5 Nc7 Kg5 Na8 Kf5 Nc7 Kg5 
 14	+0.08 	6.72M  	1:10.41	Kg5 Nxd5 a6 Nc7 a7 Ke4 Kxg6 Kf4 Kf6 Kg3 Kg5 Na8 Kf5 Nc7 Kg5 Na8 Kf5 Nc7 Kg5 
 14	+0.08 	6.25M  	1:05.41	Kg5 Nxd5 a6 Nc7 a7 Ke4 Kxg6 Kf4 Kf6 Kg3 Kg5 Na8 Kf5 Nc7 Kg5 Na8 Kf5 Nc7 Kg5 
 14	+0.08 	5.76M  	1:00.41	Kg5 Nxd5 a6 Nc7 a7 Ke4 Kxg6 Kf4 Kf6 Kg3 Kg5 Na8 Kf5 Nc7 Kg5 Na8 Kf5 Nc7 Kg5 
 14	+0.08 	5.43M  	0:55.36	Kg5 Nxd5 a6 Nc7 a7 Ke4 Kxg6 Kf4 Kf6 Kg3 Kg5 Na8 Kf5 Nc7 Kg5 Na8 Kf5 Nc7 Kg5 
 14	+0.08 	4.88M  	0:50.35	Kg5 Nxd5 a6 Nc7 a7 Ke4 Kxg6 Kf4 Kf6 Kg3 Kg5 Na8 Kf5 Nc7 Kg5 Na8 Kf5 Nc7 Kg5 
 14	+0.09 	4.43M  	0:45.35	Kg5 Nxd5 a6 Nc7 a7 Ke4 Kxg6 Kf4 Kf6 Kg3 Kf5 Na8 Kg5 Nc7 Kh5 Na8 Kg5 Nc7 Kf5 Na8 Kg5 
 14	+0.09 	3.95M  	0:40.35	Kg5 Nxd5 a6 Nc7 a7 Ke4 Kxg6 Kf4 Kf6 Kg3 Kf5 Na8 Kg5 Nc7 Kh5 Na8 Kg5 Nc7 Kf5 Na8 Kg5 
 13	+0.09 	3.48M  	0:36.07	Kg5 Nxd5 a6 Nc7 a7 Ke4 Kxg6 Kf4 Kf6 Kg3 Kf5 Na8 Kg5 Nc7 Kh5 Na8 Kg5 Nc7 Kh5 Na8 Kg5 
 13	 -128.00 	3.47M  	0:36.02	Kxh4 Kf4 d6 Ne6 g3+ Kf5 g4+ Kf4 d7 Nd4 d8=Q Nf3# 
 13	 -128.00 	3.00M  	0:31.02	Kxh4 Kf4 d6 Ne6 g3+ Kf5 g4+ Kf4 d7 Nd4 d8=Q Nf3# 
 13	 -128.00 	2.45M  	0:26.02	Kxh4 Kf4 d6 Ne6 g3+ Kf5 g4+ Kf4 d7 Nd4 d8=Q Nf3# 
 13	 -128.00 	1.92M  	0:21.02	Kxh4 Kf4 d6 Ne6 g3+ Kf5 g4+ Kf4 d7 Nd4 d8=Q Nf3# 
 13	+0.16 	1.40M  	0:16.01	Kxh4 Kxd5 Kg4 Ke6 Kg5 Kf7 Kh6 Kf6 h4 Kf5 g3 Na6 Kg7 Nc5 Kf7 Na6 Ke7 Kg4 Kf6 Kxg3 
 13	+0.16 	1.26M  	0:14.66	Kxh4 Kxd5 Kg4 Ke6 Kg5 Kf7 Kh6 Kf6 h4 Kf5 g3 Na6 Kg7 Nc5 Kf7 Na6 Ke7 Kg4 Kf6 Kxg3 
 13	+0.17 	943729	0:11.31	Kxh4 Kxd5 Kg4 Ke6 Kg5 Kf7 Kh6 Kf6 h4 Kf5 g3 Na6 Kh7 Nc5 Kg7 Na6 Kf7 Nc5 Ke7 Na6 Ke8 Kg4 Kf7 Kf5 
 13	+0.17 	827442	0:10.14	Kxh4 Kxd5 Kg4 Ke6 Kg5 Kf7 Kh6 Kf6 h4 Kf5 g3 Na6 Kh7 Nc5 Kg7 Na6 Kf7 Nc5 Ke7 Kg4 Kf6 Kxg3 
 13	+0.17 	800210	0:09.80	Kxh4 Kxd5 Kg4 Ke6 Kg5 Kf7 Kh6 Kf6 h4 Kf5 g3 Na6 Kh7 Nc5 Kg7 Na6 Kf7 Nc5 Ke7 Kg4 Kf6 Kxg3 
 12	+0.18 	512150	0:06.83	Kxh4 Kxd5 Kg4 Ke6 Kg5 Kf7 Kh6 Kf6 h4 Kf5 g3 Na6 Kh7 Nc7 Kg7 Na6 Kf7 Nc7 Ke7 Na6 Ke8 Kg4 Kf7 Kf5 
 12	+0.19 	273615	0:04.31	Kxh4 Kxd5 Kg5 Ne6+ Kg4 Kc6 Kf3 Kb5 Ke4 Kxa5 
 12	+0.18 	147939	0:02.59	Kxh4 Kxd5 Kg5 Ne6+ Kg4 Kc6 Kf3 Kb5 Ke4 Kxa5 
 11	+0.19 	127572	0:02.33	Kg5 Nxd5 a6 Nc7 a7 Ke4 Kxg6 Kf4 Kf6 Kg3 Kg5 Na8 Kh5 Nc7 Kg5 Na8 Kh5 Nc7 Kg5 
 11	+0.20 	105240	0:02.05	Kg5 Nxd5 a6 Nc7 a7 Ke4 Kxg6 Kf4 Kf6 Kg3 Kf5 Na8 Kg5 Nc7 Kh5 Na8 Kg5 Nc7 Kh5 Na8 Kg5 
 11	+0.22 	61057  	0:01.47	Kg5 Nxd5 a6 Nc7 a7 Ke4 Kxg6 Kf4 Kf6 Na8 Ke6 Nb6 Kf6 Na8 Kg6 Nc7 Kh5 Kg3 Kg5 Na8 Kh5 Nc7 Kg5 
 11	+0.24 	38722  	0:01.12	Kg5 Nxd5 a6 Nc7 a7 Ke4 Kxg6 Kf4 Kf6 Na8 Ke6 Nb6 Kf6 Na8 Kg6 Nc7 Kh5 Kg3 Kg5 Na8 Kh5 Nc7 
 11	+0.26 	27970  	0:00.95	Kg5 Nxd5 a6 Nc7 a7 Ke4 Kxg6 Kf4 Kf6 Na8 Ke6 Nb6 Kf6 Na8 Kg6 Nc7 Kh5 Kg3 Kg5 Na8 Kh5 
 11	+0.27 	21555  	0:00.75	Kg5 Nxd5 a6 Nc7 a7 Ke4 Kxg6 Kf4 Kf6 Na8 Ke6 Nb6 Kf6 Na8 Kg6 Nc7 Kh6 Na8 Kh5 Kg3 
 11	+0.18 	134487	0:02.42	Kxh4 Kxd5 Kg5 Ne6+ Kg4 Kc6 Kf3 Kb5 Ke4 Kxa5 
 11	+0.19 	106377	0:02.07	Kxh4 Kxd5 Kg5 Ne6+ Kg4 Kc6 Kf3 Kb5 Ke4 Kxa5 
 11	+0.19 	104959	0:02.05	Kxh4 Kxd5 Kg5 Ne6+ Kg4 Kc6 Kf3 Kb5 Ke4 Kxa5 
 10	+0.37 	8046    	0:00.48	Kg5 Nxd5 a6 Nc7 a7 Ke4 Kxg6 Kf4 Kf6 Na8 Ke6 Nb6 Kf6 Na8 Kg6 Nc7 Kh5 Kg3 Kg5 
 10	+0.45 	4804    	0:00.38	Kg5 Nxd5 a6 Nc7 a7 Ke4 Kf6 Kf4 Kxg6 Kg3 Kg5 Na8 Kf5 Nc7 Kg5 Na8 Kh5 Nc7 Kg5 
  9	+0.45 	4447    	0:00.37	Kg5 Nxd5 a6 Nc7 a7 Ke4 Kf6 Kf4 Kxg6 Kg3 Kg5 Na8 Kf5 Nc7 Kg5 Na8 Kh5 Nc7 Kg5 
  9	+0.47 	3563    	0:00.34	Kg5 Nxd5 a6 Nc7 a7 Ke4 Kxg6 Kf4 Kf6 Na8 Ke6 Nb6 Kf6 Na8 Kg6 Nc7 Kh5 
  9	+0.53 	2376    	0:00.30	Kg5 Nxd5 a6 Nc7 a7 Ke4 Kxg6 Kf4 Kf6 Na8 Ke6 Nb6 Kf7 Kg3 Kg6 Kxg2 
  9	+0.57 	1921    	0:00.28	Kg5 Nxd5 a6 Nc7 a7 Ke4 Kxg6 Kf4 Kf6 Na8 Ke6 Nb6 Kf7 Kg3 Kg6 Kxg2 
  8	+0.57 	1773    	0:00.27	Kg5 Nxd5 a6 Nc7 a7 Ke4 Kxg6 Kf4 Kf6 Na8 Ke6 Nb6 Kf7 Kg3 Kg6 Kxg2 
  8	+0.66 	1243    	0:00.23	Kg5 Nxd5 a6 Nc7 a7 Ke4 Kxg6 Kf4 Kf6 Na8 Ke6 Nb6 Kf7 Kg3 Kg6 
  8	+0.73 	885      	0:00.20	Kg5 Nxd5 a6 Nc7 a7 Ke4 Kxg6 Kf4 Kf6 Na8 Ke6 Kg3 Kf5 Nc7 Kg5 Kh2 
  8	+0.74 	819      	0:00.20	Kg5 Nxd5 a6 Nc7 a7 Ke4 Kxg6 Kf4 Kf6 Na8 Ke6 Kg3 Kf5 Nc7 Kg5 
  7	+0.76 	718      	0:00.19	Kg5 Nxd5 a6 Nc7 a7 Ke4 Kxg6 Kf4 Kf6 Na8 Ke6 Nb6 Kf6 Na8 
  7	+0.80 	607      	0:00.18	Kg5 Nxd5 a6 Nc7 a7 Ke4 Kxg6 Kf4 Kf6 Na8 Ke6 Nb6 Kf6 Kg3 
  7	+0.86 	430      	0:00.17	Kg5 Nxd5 a6 Nc7 a7 Ke4 Kxh4 Kf4 g3+ Kf5 g4+ Kf4 g5 
  6	+0.87 	300      	0:00.16	Kg5 Nxd5 a6 Nc7 a7 Ke4 Kg4 Ke5 Kxh4 Kf4 g3+ Kf5 
  6	+0.58 	251      	0:00.14	Kxh4 Kxd5 Kg5 Ne6+ Kf6 Nf4 a6 Kc6 Kg5 Nxg2 
  6	+0.58 	228      	0:00.13	Kxh4 Kxd5 Kg5 Ne6+ Kf6 Nf4 a6 Kc6 Kg5 Nxg2 
  6	+0.68 	176      	0:00.13	Kxh4 Kxd5 Kg5 Ne6+ Kf6 Nf4 a6 Kc6 a7 Kb7 
  5	+0.73 	89        	0:00.10	Kxh4 Kxd5 Kg5 Ne6+ Kxg6 
  4	+1.43 	28        	0:00.08	Kxh4 Kxd5 Kg5 Ne6+ Kxg6 
  4	+1.33 	45        	0:00.08	Kxh4 Kxd5 Kg5 Ne6+ Kxg6 
  4	+1.06 	64        	0:00.10	Kxh4 Kxd5 Kg5 Ne6+ Kf6 Nf4 a6 Kc6 
  3	+1.86 	21        	0:00.07	Kxh4 Kf5 d6 Na6 d7 
  3	+1.63 	16        	0:00.05	Kxh4 Nxd5 Kg5 Nf4 Kh6 
  3	+1.57 	24        	0:00.07	Kxh4 Kxd5 Kg5 Ne6+ Kxg6 
  2	+1.63 	13        	0:00.05	Kxh4 Nxd5 Kg5 Nf4 
  2	+1.51 	8          	0:00.04	Kxh4 Nxd5 Kg5 
  1	+0.94 	3          	0:00.03	Kxh4 Kxd5 
  0	# 
Image

OneTrickPony
Posts: 136
Joined: Mon Apr 29, 2013 11:29 pm

Re: The “Kryptonite” position

Post by OneTrickPony » Fri Jan 10, 2020 2:08 pm

It looks like possible problem in Leela search. At some point it realizes (quite quickly with Sergio net) that Kxh4 is losing but still lists it at the first option till visit counts catches up:

1. 13 [+0.00] 1.Kxh4 Kf4 2.g3+ Kf5 3.d6 Ne6 4.d7 Ng5 5.g4+ Kf4 6.d8=Q Nf3#
2. 13 [+53.30] 1.d6 Kxd6 2.Kxh4 Ne6 3.Kg4 Kc6 4.Kf3 Kb5 5.Ke4 Kxa5
3. 13 [+53.10] 1.Kg5 Nxd5 2.a6 Nc7 3.a7 Ke4 4.Kxg6 Kf4 5.Kf6 Na8 6.Ke7 Kg3 7.Kd6 Kxg2
4. 13 [+50.97] 1.Kf3 Kxd5 2.Kf4 Ne6+ 3.Kg4 Kc6 4.Kxh4 Kb5 5.Kg4 Kxa5
5. 13 [+50.99] 1.g3 hxg3 2.Kxg3 Kxd5
6. 13 [+42.43] 1.a6 Nxa6 2.Kg5 Ke4 3.Kf6 Kf4 4.Kxg6 Kg3 5.d6 Nc5 6.Kf5 Nd7 7.Kg5 Nb6 8.Kh5
(evals in win percentages)

dkappe
Posts: 571
Joined: Tue Aug 21, 2018 5:52 pm
Full name: Dietrich Kappe

Re: The “Kryptonite” position

Post by dkappe » Sat Jan 11, 2020 12:08 am

OneTrickPony wrote:
Fri Jan 10, 2020 2:08 pm
It looks like possible problem in Leela search. At some point it realizes (quite quickly with Sergio net) that Kxh4 is losing but still lists it at the first option till visit counts catches up:

1. 13 [+0.00] 1.Kxh4 Kf4 2.g3+ Kf5 3.d6 Ne6 4.d7 Ng5 5.g4+ Kf4 6.d8=Q Nf3#
2. 13 [+53.30] 1.d6 Kxd6 2.Kxh4 Ne6 3.Kg4 Kc6 4.Kf3 Kb5 5.Ke4 Kxa5
3. 13 [+53.10] 1.Kg5 Nxd5 2.a6 Nc7 3.a7 Ke4 4.Kxg6 Kf4 5.Kf6 Na8 6.Ke7 Kg3 7.Kd6 Kxg2
4. 13 [+50.97] 1.Kf3 Kxd5 2.Kf4 Ne6+ 3.Kg4 Kc6 4.Kxh4 Kb5 5.Kg4 Kxa5
5. 13 [+50.99] 1.g3 hxg3 2.Kxg3 Kxd5
6. 13 [+42.43] 1.a6 Nxa6 2.Kg5 Ke4 3.Kf6 Kf4 4.Kxg6 Kg3 5.d6 Nc5 6.Kf5 Nd7 7.Kg5 Nb6 8.Kh5
(evals in win percentages)
It’s a problem in all uct/puct implementations I’m aware of. These always picks the move with the most visits. If the refutation shows up late, it can’t take a lot of nodes to catch on.

Post Reply