Discussion of anything and everything relating to chess playing software and machines.
Moderator: Ras
brtzsnr
Posts: 433Joined: Fri Jan 16, 2015 4:02 pm
Post by brtzsnr Sun Jul 19, 2015 10:01 am
[d] 8/3R4/P4pk1/8/4K3/8/8/6b1 b - - 4 73
This a test position that caused a search explosion in zurichess during yesterday's contest.
zurichess
Code: Select all
info depth 18 seldepth 23 score cp -735 nodes 32667001 time 11871 nps 2751685 pv g1b6 a6a7 b6a7 d7a7 f6f5 e4e5 f5f4 a7a4 f4f3 a4f4 f3f2 f4f2 g6g7 f2f6 g7g8 f6f4 g8g7 e5e6 g7g6 f4g4 g6h
7 g4g1 h7h6 e6d5 h6h7 d5e4 h7h6
info depth 19 seldepth 29 score cp -1033 nodes 404429766 time 156144 nps 2590097 pv g1b6 e4d5 g6f5 d5c6 b6g1 d7d5 f5e4 d5c5 g1c5 c6c5 f6f5 a6a7 f5f4 a7a8Q e4e3 a8a3 e3e4 a3a4 e4e3 a4d4
e3f3 c5c4 f3g4 d4d5
stockfish 5 also has a small search explosion (57x more nodes searched)
Code: Select all
info depth 24 seldepth 30 score cp -8313 nodes 1747066 nps 5342709 time 327 multipv 1 pv f6f5 e4f3 g6f6 a6a7 g1a7 d7a7 f6e5 a7a5 e5d6 a5f5 d6e6 f3e4 e6d6 f5d5 d6c6 e4d4 c6b7 d5d6 b7b8
d4e5 b8b7 e5d5 b7c7 d5e6 c7b7 e6d5
info depth 25 seldepth 41 score mate -21 nodes 99873966 nps 6156319 time 16223 multipv 1 pv f6f5 e4f3 g6h6 a6a7 g1a7 d7a7 h6g5 a7a6 f5f4 a6a5 g5f6
MikeGL
Posts: 1010Joined: Thu Sep 01, 2011 2:49 pm
Post by MikeGL Sun Jul 19, 2015 10:21 am
critter32bit in my lowly dual core finds mate, but I think it is not the optimum number of moves.
Laskos
Posts: 10948Joined: Wed Jul 26, 2006 10:21 pmFull name: Kai Laskos
Post by Laskos Sun Jul 19, 2015 10:24 am
Mate in 15, black f5+ for longest DTM.
hgm
Posts: 28489Joined: Fri Mar 10, 2006 10:06 amLocation: AmsterdamFull name: H G Muller
Post by hgm Sun Jul 19, 2015 10:32 am
Who cares? The position is trivially lost. If you see the gain of the Bishop, and play it, you will see the gain of the Pawn later, and after that the opportunity to force the bare King into a corner, and after that the mate. You don't get extra points for announcing it 25 moves in advance. Even Usurpator II would have won this for white against a 6-men tablebase, with its 3-ply search.
Search explosions will happen. In particular when at large depth you get to see something that drastically changes the root score, and all plans laid out in the hash from the previous iteration suddenly become useless, so that it has to do a completey uninformed (or worse: wrongly informed) high-depth search.
Btw, Fruit needs 20 ply to see mate:
Code: Select all
dep score nodes time (not shown: tbhits knps seldep)
20 -99.72 692.3M 8:02.11 Bc5 a7 Bxa7 Rxa7 f5+ Kf4 Kf6 Rd7 Ke6 Rd2 Kf6 Re2 Kg6 Re6+ Kg7 Kxf5 Kf7 Re5 Kg7 Re7+ Kf8 Kf6 Kg8 Re8+ Kh7 Rd8 Kh6 Rh8#
19 -6.25 15.6M 0:07.88 Bc5 a7 Bxa7 Rxa7 f5+ Kf4 Kf6 Rd7 Ke6 Rd2 Kf6 Re2 Kg6 Re6+ Kg7 Kxf5 Kf7 Re5 Kg7 Re7+ Kg8 Rb7 Kf8 Ke4
18 -6.15 12.2M 0:06.28 Bc5 a7 Bxa7 Rxa7 f5+ Kf4 Kf6 Rd7 Ke6 Rd2 Kf6 Re2 Kg6 Re6+ Kg7 Kxf5 Kf7 Re5 Kg8 Re7 Kf8 Rd7
17 -6.15 9.78M 0:05.19 Bc5 a7 Bxa7 Rxa7 f5+ Kf4 Kf6 Rd7 Ke6 Rd2 Kf6 Re2 Kg6 Re6+ Kg7 Kxf5 Kf7 Re4 Kg7 Re7+ Kf8 Rd7
16 -5.84 7.43M 0:04.08 Bc5 a7 Bxa7 Rxa7 f5+ Ke5 f4 Ra6+ Kg5 Rf6 f3 Rxf3 Kg4 Rf4+ Kg3 Kf5 Kg2 Kg4 Kh2 Rf2+ Kg1 Rf3
15 -5.79 6.09M 0:03.45 Bc5 a7 Bxa7 Rxa7 f5+ Ke5 f4 Ra6+ Kg5 Rf6 f3 Rxf3 Kg4 Rf7 Kg3 Ke4 Kg4 Rg7+ Kh4 Rg8
14 -5.96 4.74M 0:02.75 Bc5 a7 Bxa7 Rxa7 f5+ Kf4 Kf6 Ra6+ Ke7 Kxf5 Kd7 Ke5 Kc7 Kd5 Kb7 Rc6 Kb8 Ke4
14 -6.80 2.56M 0:01.55 f5+ Kd3 Kf6 Rd4 Ke5 a7 Bxd4 a8=Q Bc5 Qh8+ Kd5 Qf6 Bd6 Qxf5+ Be5 Qf7+ Kc6 Ke4 Bd6 Qd5+ Kd7
13 -5.56 411883 0:00.28 f5+ Kf3 Kf6 a7 Bxa7 Rxa7 Ke5 Ra5+ Kf6 Kf4 Ke6 Rxf5 Kd6 Ke4 Ke6 Rd5 Kf6
12 -5.47 305523 0:00.22 f5+ Kf3 Kf6 a7 Bxa7 Rxa7 Ke5 Ra5+ Kf6 Kf4 Ke6 Rxf5 Kd6 Ke4 Ke6 Rd5
11 -5.47 218813 0:00.17 f5+ Kf3 Kf6 a7 Bxa7 Rxa7 Ke5 Ra5+ Ke6 Kf4 Kd6 Kxf5 Kc7 Ke4 Kd6 Rd5+ Ke6
10 -5.47 164048 0:00.14 f5+ Kf3 Kf6 a7 Bxa7 Rxa7 Ke5 Ra5+ Ke6 Kf4 Kf6 Rxf5+ Ke6 Ke4 Kd6
9 -5.34 138159 0:00.12 f5+ Kf3 Kf6 a7 Bxa7 Rxa7 Ke5 Ra5+ Ke6 Kf4 Kd6 Kxf5 Kc6
8 -5.26 24967 0:00.02 f5+ Kf3 Kf6 a7 Bxa7 Rxa7 Ke5 Ra5+ Ke6 Kf4 Kd6 Ra6+ Kd5 Kxf5
7 -4.69 13608 0:00.02 f5+ Kf3 Kf6 a7 Bxa7 Rxa7 Kg5 Ra6 Kh5 Kf4
6 -4.30 9194 0:00.02 f5+ Kf3 Kf6 a7 Bxa7 Rxa7 Ke5 Ra5+ Ke6 Kf4
5 -4.00 3923 0:00.00 f5+ Ke5 Bh2+ Ke6 Bg1 a7 Bxa7 Rxa7 f4
4 -4.35 2192 0:00.00 f5+ Ke5 f4 Kxf4 Bc5
3 -4.08 944 0:00.00 f5+ Kd3 f4 a7 Bxa7 Rxa7
2 -3.40 289 0:00.00 f5+ Ke5 Be3
2 -4.42 181 0:00.00 Bc5 a7 f5+ Ke5
1 -3.58 9 0:00.00 Bc5
brtzsnr
Posts: 433Joined: Fri Jan 16, 2015 4:02 pm
Post by brtzsnr Sun Jul 19, 2015 10:55 am
hgm wrote: Search explosions will happen. In particular when at large depth you get to see something that drastically changes the root score, and all plans laid out in the hash from the previous iteration suddenly become useless, so that it has to do a completey uninformed (or worse: wrongly informed) high-depth search.
Btw, Fruit needs 20 ply to see mate:
Code: Select all
dep score nodes time (not shown: tbhits knps seldep)
20 -99.72 692.3M 8:02.11 Bc5 a7 Bxa7 Rxa7 f5+ Kf4 Kf6 Rd7 Ke6 Rd2 Kf6 Re2 Kg6 Re6+ Kg7 Kxf5 Kf7 Re5 Kg7 Re7+ Kf8 Kf6 Kg8 Re8+ Kh7 Rd8 Kh6 Rh8#
19 -6.25 15.6M 0:07.88 Bc5 a7 Bxa7 Rxa7 f5+ Kf4 Kf6 Rd7 Ke6 Rd2 Kf6 Re2 Kg6 Re6+ Kg7 Kxf5 Kf7 Re5 Kg7 Re7+ Kg8 Rb7 Kf8 Ke4
Thanks for confirming. I was not sure whether this issue was a zurichess specific, or more of a problem of the current state of searching heuristics.
stegemma
Posts: 859Joined: Mon Aug 10, 2009 10:05 pmLocation: ItalyFull name: Stefano Gemma
Post by stegemma Sun Jul 19, 2015 10:59 am
Satana 2.0.8 can't go over ply 10, at 5 minutes per game, and see only the same line as your software at the same depth. It doesn't see the mate but for satana it is hard even to mate with KR vs K... so I'm really satisfied!
mar
Posts: 2685Joined: Fri Nov 26, 2010 2:00 pmLocation: Czech RepublicFull name: Martin Sedlak
Post by mar Sun Jul 19, 2015 1:36 pm
brtzsnr wrote: Thanks for confirming. I was not sure whether this issue was a zurichess specific, or more of a problem of the current state of searching heuristics.
Well cheng only sees mated in 16 for f5+
What I do (for very long already) is I score positions where one side has bare king and another has a major piece (that cannot be captured and it's not stalemate)
as "certain win", which is +128.0 in my case. This way it gives -130 at depth 14.
Search will take care of the rest (all you need is bonus for king in center).
kgburcham
Posts: 2016Joined: Sun Feb 17, 2008 4:19 pm
Post by kgburcham Sun Jul 19, 2015 2:57 pm
[d] 8/3R4/P4pk1/8/4K3/8/8/6b1 b - -
zullil
Posts: 6442Joined: Tue Jan 09, 2007 12:31 amLocation: PA USAFull name: Louis Zulli
Post by zullil Sun Jul 19, 2015 3:34 pm
kgburcham wrote: [d] 8/3R4/P4pk1/8/4K3/8/8/6b1 b - -
+M11 makes no sense.
hgm
Posts: 28489Joined: Fri Mar 10, 2006 10:06 amLocation: AmsterdamFull name: H G Muller
Post by hgm Sun Jul 19, 2015 3:40 pm
mar wrote: Well cheng only sees mated in 16 for f5+
Wouldn't it be better to add 128 to the normal evaluation, rather than score the position such? Otherwise the engine wouls stop trying to make progress after you reach such a position in the root.
