Skippers sixty memorable bugs: nr 7

[Henk]

In this position Skipper played Kf6 ?? in a two minute game.

[d] 3r4/1p1P4/7p/2KRk3/4p2p/4P3/4P3/8 b - - 1 39

When I tried to reproduce the bug it finds Ke6 only at level 9 which is too late.

Code: Select all

Depth  Value   Time
  1  9192.00   0.020          28   e5f6 
  2  9132.00   0.023         138   e5f6 c5d6 
  3  9147.00   0.027         694   e5f6 d5d4 h4h3 
  4  9129.00   0.034        1753   e5f6 c5d4 f6e7 d5b5 
  5  9178.00   0.067        7828   e5f6 c5d4 f6e7 d4e4 d8d7 
  6  -10.00   0.103       14630   e5f6 d5d6 f6g5 c5b6 g5h5 b6b7 
  7  -13.00   0.191       30027   e5f6 d5d6 f6f5 d6h6 d8d7 h6h4 d7d8 
  8  -63.00   0.380       61975   e5f6 d5d6 f6g7 c5b6 h4h3 b6b7 h3h2 d6d1 
  9  -17.00   1.693      256019   e5e6 d5d6 e6f7 d6h6 d8d7 h6h4 d7d2 h4h2 d2d1 
 10  -37.00   2.572      414664   e5e6 d5d6 e6e5 d6h6 d8d7 h6h5 e5e6 h5h4 d7d5 
 11  -37.00   4.918      799411   e5e6 d5d6 e6e5 d6d5 

Disabling null move helps a bit for it now sees Ke6 at level 8:

Code: Select all

Depth  Value   Time
  1  9168.00   0.000          28   e5f6 
  2  9127.00   0.001         147   e5e6 d5d4 
  3  9137.00   0.005         748   e5f6 d5d6 f6g7 
  4  9128.00   0.010        1627   e5f6 d5d4 f6f5 c5c4 
  5  -10.00   0.023        4437   e5f6 d5d6 f6e5 d6h6 h4h3 
  6  9128.00   0.088       16677   e5f6 d5d6 f6g7 d6d5 g7f8 c5b4 
  7  -15.00   0.180       34448   e5f6 d5d6 f6f7 d6h6 d8d7 h6h4 d7e7 
  8  -18.00   0.501       87527   e5e6 d5d6 e6e5 d6h6 d8d7 h6h4 d7d2 h4h5 
  9  -21.00   0.855      147479   e5e6 d5d6 e6e5 d6h6 d8d7 h6h4 d7c7 c5b6 c7d7 
 10  -83.00   1.648      303479   e5e6 d5d6 e6e5 d6h6 d8d7 h6h4 d7c7 c5b6 c7d7 
 11  -83.00   3.962      719202   e5e6 d5d6 e6e5 d6d5 

[Sven]

You need to find out why iteration 5 (or 6 if you do not try promotions in qsearch) does not find the black loss after 1...Kf6 2.Kd6 h3 3.Kc7 Rg8 4.d8Q. According to the PV you gave Skipper correctly saw an advantage of a pawn for Black after 1...Kf6 2.Kd4 Ke7 3.Kxe4 Rxd7 but "overlooked" (i.e. misevaluated) the much better move 2.Kd6 for White. Why? Print the full (3-ply) search tree after 1...Kf6 2.Kd6 in iteration 5 and look into it.

[Henk]

I first have to write a method that prints the tree for I don't have it.

In the position below it also does not see that h3 is losing at low depths.

[d] 3r4/1p1P4/3K1k1p/3R4/4p2p/4P3/4P3/8 b - - 1 1

Code: Select all

Depth  Value   Time
  1  9216.00   0.016          23   b7b6 
  2  9216.00   0.019         114   b7b6 d5d4 
  3    0.00   0.032        1436   b7b6 d5h5 d8a8 
  4    0.00   0.041        2661   b7b6 d5h5 d8a8 h5h6 
  5    0.00   0.060        6132   b7b6 d5h5 d8a8 h5h6 f6g5 
  6    0.00   0.120       16156   h4h3 d5h5 f6g7 d6c7 g7g6 h5h3 
  7    0.00   0.201       30954   h4h3 d5h5 f6g6 h5h3 d8f8 d6c7 f8f7 
  8    0.00   0.403       68301   h4h3 d5h5 f6g6 h5h3 d8f8 d6c7 f8f7 c7b7 
  9  -30720.00   1.841      296408   h4h3 d5d1 h3h2 d1f1 f6g7 d6c7 d8d7 c7d7 g7h8 
 10  -30720.00   2.725      459426   h4h3 d5d1 h3h2 d1f1 f6g7 d6c7 d8d7 c7d7 g7h8 
 11  -30720.00   4.458      822030   h4h3 d5d1 h3h2 d1f1 f6g7 d6c7 d8a8 d7d8 a8d8 

[Sven]

I first have to write a method that prints the tree for I don't have it.

You only need to print something at each leaf node that is inside the subtree of an interesting position.

In the position below it also does not see that h3 is losing at low depths.

The critical line, in terms of search depth, would be (counted from your initial position in the beginning of this thread):
2...h3 3.Kc7 Ke7 4.Re5+ Kf6 5.Kxd8 Kxe5 6.Ke7 h2 7.d8Q b5 8.Qd7 b4 9.Qb5#
That's 13 plies after 2...h3. I initially thought of a shorter line after 3... Rg8 but 3...Ke7 pushes the horizon quite a bit. The PV above that ends with a dubious Kxb7?? move (instead of Kc8 +-):

Code: Select all

  8    0.00   0.403       68301   h4h3 d5h5 f6g6 h5h3 d8f8 d6c7 f8f7 c7b7

might be caused by a wrong evaluation: the white passer on d7 must get a much higher score than the black one on b7, so exchanging b7 vs. d7 is nonsense.

[Henk]

Yes I already saw that the code that evaluated passers was commented out. I forgot to enable it after a debugging session earlier.

[Sven]

Yes I already saw that the code that evaluated passers was commented out. I forgot to enable it after a debugging session earlier.

So does that also fix this bug nr. 7?

[Henk]

Yes I already saw that the code that evaluated passers was commented out. I forgot to enable it after a debugging session earlier.

So does that also fix this bug nr. 7?

[d] 3r4/1p1P4/7p/2KRk3/4p2p/4P3/4P3/8 b - - 1 39

Yes it does better now. It finds Ke6 at level 6.

Code: Select all

Depth  Value   Time
  1  7266.00   0.019          28   e5f6 
  2  7266.00   0.022          81   e5f6 d5d6 
  3  7266.00   0.028         282   e5f6 d5d6 f6g7 
  4  7266.00   0.031         597   e5f6 d5d6 f6g7 d6d5 
  5  -31.00   0.041        2111   e5f6 d5d4 f6e5 c5b6 h4h3 
  6  1000.00   0.086        8803   e5e6 d5d6 e6e7 d6h6 d8d7 h6h4 
  7  1000.00   0.111       12525   e5e6 d5d6 e6e7 d6h6 d8d7 h6h4 d7c7 
  8  1000.00   0.159       20384   e5e6 d5d6 e6e7 d6h6 d8d7 h6h4 d7c7 c5b4 
  9  1000.00   0.261       37187   e5e6 d5d6 e6e7 d6h6 e7d7 h6h4 d8e8 c5b6 d7c8 
 10  1000.00   0.414       63417   e5e6 d5d6 e6e7 d6h6 e7d7 h6h4 d8e8 c5b6 d7c8 
 11  1000.00   1.090      188970   e5e6 d5d6 e6e5 d6h6 d8d7 h6h4 d7d5 c5b6 d5d7 
 12  1000.00   1.694      299458   e5e6 d5d6 e6e5 d6h6 d8d7 h6h4 d7d5 c5b6 d5d7 
 13  1000.00   3.228      598666   e5e6 d5d6 e6e5 d6h6 d8d7 h6h4 d7d5 c5b6 d5d7 

[d] 3r4/1p1P4/3K1k1p/3R4/4p2p/4P3/4P3/8 b - - 1 1


But it does not consider Kc7 after h3. Might be Kc7 is worse than Rd1.

Code: Select all

Depth  Value   Time
  1  7966.00   0.017          27   h4h3 
  2  7966.00   0.024         129   h4h3 d5d4 
  3  9066.00   0.030         392   h4h3 d5d4 h3h2 
  4  7766.00   0.039        1428   b7b6 d5d4 f6f5 d4d5 
  5  7266.00   0.059        4631   f6f7 d6c7 f7e6 c7d8 e6d5 
  6  -1950.00   0.104       11401   f6f7 d6c7 f7e7 d5d4 b7b6 c7b6 
  7  -3450.00   0.261       33677   h4h3 d5h5 f6g6 h5h3 d8f8 h3h4 f8f6 
  8  -3450.00   0.372       50652   h4h3 d5h5 f6g6 h5h3 d8f8 h3h4 f8f6 d6c5 
  9  -11666.00   0.784      112981   h4h3 d5h5 b7b5 h5h3 b5b4 h3h6 f6g5 h6h7 g5g4 
 10  -24920.00   1.894      295903   h4h3 d5d1 h3h2 d1f1 f6g5 d6c7 d8d7 c7d7 b7b5 
 11  -24420.00   2.742      456194   h4h3 d5d1 h3h2 d1f1 f6g5 d6c7 d8d7 c7d7 b7b5 

[Sven]

But it does not consider Kc7 after h3. Might be Kc7 is worse than Rd1.

Rd1 is winning as well. Kc7 is more obvious for humans, that's why I mentioned it, but for engines there is probably not much difference with a deeper search. And in the PV you have shown White actually plays Kc7 few plies later. So no need to worry in this case