Stockfish 7 and partial 6 piece syzygy problem?

[Jouni]

I installed some interesting 6 piece bases (both files of course). After test runs I noticed a problem.
If there is 6 piece position on board, but no corresponding tablebase engine is not using any syzy at all! Komodo is more clever here.

Example 1:

[D]6B1/5P2/5Knk/8/7P/8/8/8 w - - 0 1


Analysis by Stockfish 7 64 BMI2:


1.h5 Nf8 2.Ke7 Kg7 3.Ke8 Ne6 4.h6+ Kh8 5.Kd7 Nf8+ 6.Kd8 Ng6 7.Kc7 Nf8 8.Kb6 Ne6 9.Kb5 Nf8 10.Kc5 Ng6 11.Kd6 Nf8 12.Ke7 Ng6+ 13.Kd8 Nf8 14.Kc7 Ne6+ 15.Kc8 Nf8 16.Kb7 Ne6 17.Kb8 Nf8 18.Ka7 Ne6 19.Kb6 Nf8 20.Kc6 Ne6 21.Kd5 Nf8 22.Ke5 Ng6+ 23.Kf5 Nf8 24.Kf6 Nd7+ 25.Kg5 Nf8 26.Kf4 Nd7 27.Ke4 Nf8
+- (2.82) Depth: 39/54 00:00:07 168mN
1.h5 Nf8 2.Ke7 Kg7 3.Ke8 Ne6 4.h6+ Kh8 5.Kd7 Nf8+ 6.Kd8 Ng6 7.Kc7 Nf8 8.Kb6 Ne6 9.Kb5 Nf8 10.Kc5 Ng6 11.Kd6 Nf8 12.Ke7 Ng6+ 13.Kd8 Nf8 14.Kc7 Ne6+ 15.Kc8 Nf8 16.Kb7 Ne6 17.Kb8 Nf8 18.Ka7 Ne6 19.Kb6 Nf8 20.Kc6 Ne6 21.Kd5 Nf8 22.Ke5 Ng6+ 23.Kf5 Nf8 24.Kf6 Nd7+ 25.Kg5 Nf8 26.Kf4 Nd7 27.Ke4 Nf8 28.Kd5 Ng6
+- (2.82) Depth: 40/56 00:00:10 224mN

Analysis by Komodo 9.2 64-bit:

1.Ke6 Kg7 2.h5 Nf4+ 3.Ke7 Nd5+ 4.Kd6 Nf6 5.h6+ Kf8 6.Ke6 Ne4 7.Kf5 Nf6 8.Kg6 Ng4 9.h7 Ne5+ 10.Kg5 Nf3+ 11.Kf6 Nd2 12.h8Q Ne4+ 13.Kg6 Nc3
+- (250.00) Depth: 26 00:00:00 3218kN, tb=469678

Example2:

[D]1k6/8/1B5p/1R6/6K1/8/8/7r w - - 0 1

Analysis by Stockfish 7 64 BMI2:

1.Kg3 Kb7 2.Be3+ Kc6 3.Ra5 h5 4.Kg2 Re1 5.Bf2 Rb1 6.Rc5+ Kd6 7.Rxh5 Rb3 8.Bg3+ Kc6 9.Bf4 Rb5 10.Rh6+ Kd5 11.Kf3 Rb3+ 12.Kg4 Rb7 13.Rh8 Rg7+ 14.Kf3 Ra7 15.Re8 Ra4 16.Rd8+ Ke6 17.Rd6+ Kf5 18.Rd5+ Ke6 19.Rb5 Ra3+ 20.Be3 Kd7 21.Ke4
+/= (0.68) Depth: 31/45 00:00:07 92907kN
1.Kg3 Kb7 2.Be3+ Kc6 3.Ra5 h5 4.Kg2 Re1 5.Bf2 Rb1 6.Rc5+ Kd6 7.Rxh5 Rb3 8.Bg3+ Kc6 9.Bf4 Rb5 10.Rh6+ Kd5 11.Kf3 Rb3+ 12.Kg4 Rb7 13.Rh8 Rg7+ 14.Kf3 Ra7 15.Re8 Ra4 16.Rd8+ Ke6 17.Rd6+ Kf5 18.Rd5+ Ke6 19.Rb5 Ra3+ 20.Be3 Kd7 21.Ke4 Ra4+ 22.Kd5
+/= (0.68) Depth: 32/45 00:00:09 115mN

Analysis by Komodo 9.2 64-bit:

1.Bg1+ Kc7 2.Rb1 Kd7 3.Re1 Kc6 4.Rd1 Kb5 5.Rc1 h5+ 6.Kg3 Ka4 7.Rb1 h4+ 8.Kg2 h3+
+- (250.00) Depth: 13 00:00:00 3080kN, tb=493450


With 5 piece syzygy SF solves positions correct. Maybe this is easily fixable so You can use 6 piece syzygy without full 150 GB set?

[syzygy]

If there is 6 piece position on board, but no corresponding tablebase engine is not using any syzy at all!

That is not true.

SF simply fails to find the win in these two positions.

[petero2]

I installed some interesting 6 piece bases (both files of course). After test runs I noticed a problem.
If there is 6 piece position on board, but no corresponding tablebase engine is not using any syzy at all!
...
With 5 piece syzygy SF solves positions correct. Maybe this is easily fixable so You can use 6 piece syzygy without full 150 GB set?

I can reproduce the problem in stockfish 7 in droidfish. Here is a fix:

Code: Select all

--- a/search.cpp
+++ b/search.cpp
@@ -271,7 +271,7 @@
               TB::RootInTB = Tablebases::root_probe_wdl(rootPos, rootMoves, TB::Score);
 
               // Only probe during search if winning
-              if &#40;TB&#58;&#58;Score <= VALUE_DRAW&#41;
+              if &#40;TB&#58;&#58;RootInTB && TB&#58;&#58;Score <= VALUE_DRAW&#41;
                   TB&#58;&#58;Cardinality = 0;
           &#125;
 

[Jouni]

Thanks Peter. Hopefully somebody sends fix to Github also.

[syzygy]

Here is a fix:

Yes, that is a good fix.
It might indeed have caused what Jouni observed.

[Jouni]

Will there be a fix soon? Finding this bug took me just 1 minute, because first position was number 3 in AH_endgames.

[syzygy]

Will there be a fix soon? Finding this bug took me just 1 minute, because first position was number 3 in AH_endgames.

Well, you're not really supposed to use incomplete sets

Whether it will be fixed soon depends on the SF maintainers (and whether someone submits a patch, I guess).

[ernest]

Well, you're not really supposed to use incomplete sets

Well, I certainly agree with that, especially for syzygy !!!

[Jouni]

I see there has been syzygy rewriting lately. May be this can be fixed at the same time?

[bob]

I see there has been syzygy rewriting lately. May be this can be fixed at the same time?

This has been a general problem for so long... With nalimov if you are in a KRPKR ending and you do not have KRQKR, you will have a problem. You will avoid promoting since with the pawn on the board, you see a mate-in-N score, but after the promotion you would only see a queen advantage or so, so you don't promote.

With the 6 piece syzygy's being so small, what would be the reason for having an incomplete set? You are introducing overhead. Why try to probe a table you don't have in the first place???

If I were Ronald, I would probably just fix the initialization code so that it would not try to use any 6 piece files unless ALL 6 piece files are available, then this kind of problem goes away.