KPK

[hgm]

Do you do this automatically (as I suspect) or manually for each piece configuration?

I never use non-trivial indexing (as if accessing a multidimensional array indexed by the square numbers). Indexes outside the canonical symmetry sector are mapped into it by manipulation of the index, however, to exploit symmetry. For 8x8 boards this easy (e.g. XORing with 070707070).

[Karlo Bala]

Safe promotions to queen are considered a win,

I suppose you do test for stalemate after the promotion, not just for recapture of the Queen. Having a Queen and the move is always a win.

Note that there are also few KPK positions where queening draws due to stalemate but underpromoting to a rook wins, e.g.:
[d]8/k1P5/2K5/8/8/8/8/8 w - - 0 1

Sven

Just out of curiosity, are there any where you need to know about under promotion to judge them accurately, or are they all positions where not knowing about under promotion just causes you to mate a few moves later (like this one)?

-Sam

Probably.

[d]8/k1P5/8/K7/8/8/8/8 w - - 0 1

[mar]

I never use non-trivial indexing (as if accessing a multidimensional array indexed by the square numbers). Indexes outside the canonical symmetry sector are mapped into it by manipulation of the index, however, to exploit symmetry. For 8x8 boards this easy (e.g. XORing with 070707070).

I see, so you mirror horizontally, vertically or both.
I thought that in pawnless positions (8x8 boards), one could try 90-degrees rotations as well (I hope it's not nonsense).

[hgm]

Yes, diagonal mirroring is slightly more expensive:

address = index>>3 & 0707070707 | index<<3 & 0x7070707070

If the board is not 8x8 it is not so straightforward, and the only generator I have for arbitrary boards does not use symmetry at all. It should be possible, however, by keeping the contribution of file and rank coords to the index separately. E.g. on a 10x10 board you could do

indexh = file1 + 100* file2 + 10000*file3 ...
indexv = rank1 + 100*rank2 + 10000*rank3 ...
index = indexh + 10*indexv

Then you can use 90909 - indexh for left-right flip, and 90909 - indexv for top-bottom flip. And index = 10*indexh + indexv for a diagonal reflection. indexh and indexv can then be each updated incrementally in the scanning and moving routines. On non-square boards there is no diagonal symmetry, and you could directly update 10*indexv incrementally.

[BubbaTough]

Safe promotions to queen are considered a win,

I suppose you do test for stalemate after the promotion, not just for recapture of the Queen. Having a Queen and the move is always a win.

Note that there are also few KPK positions where queening draws due to stalemate but underpromoting to a rook wins, e.g.:
[d]8/k1P5/2K5/8/8/8/8/8 w - - 0 1

Sven

Just out of curiosity, are there any where you need to know about under promotion to judge them accurately, or are they all positions where not knowing about under promotion just causes you to mate a few moves later (like this one)?

-Sam

Probably.

[d]8/k1P5/8/K7/8/8/8/8 w - - 0 1

Nice, I like it .

-Sam

[mar]

Yes, diagonal mirroring is slightly more expensive:

address = index>>3 & 0707070707 | index<<3 & 0x7070707070

If the board is not 8x8 it is not so straightforward, and the only generator I have for arbitrary boards does not use symmetry at all. It should be possible, however, by keeping the contribution of file and rank coords to the index separately. E.g. on a 10x10 board you could do

indexh = file1 + 100* file2 + 10000*file3 ...
indexv = rank1 + 100*rank2 + 10000*rank3 ...
index = indexh + 10*indexv

Then you can use 90909 - indexh for left-right flip, and 90909 - indexv for top-bottom flip. And index = 10*indexh + indexv for a diagonal reflection. indexh and indexv can then be each updated incrementally in the scanning and moving routines. On non-square boards there is no diagonal symmetry, and you could directly update 10*indexv incrementally.

I see, clever, so only one diagonal mirroring (or transposition) will do.
So the rotations then become HD, HV and VD.

[lucasart]

Safe promotions to queen are considered a win,

I suppose you do test for stalemate after the promotion, not just for recapture of the Queen. Having a Queen and the move is always a win.

Note that there are also few KPK positions where queening draws due to stalemate but underpromoting to a rook wins, e.g.:
[d]8/k1P5/2K5/8/8/8/8/8 w - - 0 1

Sven

Good spot.

But that is not a problem. In practice the KPK bitbase only tells you if it's a draw or not (1 bit per position). It doesn't tell you that you should do a queen promotion. So it's not a problem.

I don't think there are positions where white can promote (w/o being captured) and it's draw regardless of the promotion piece he chooses (or even if he chooses a king move instead).

[hgm]

I don't think there are positions where white can promote (w/o being captured) and it's draw regardless of the promotion piece he chooses (or even if he chooses a king move instead).

[d]8/1P6/8/8/8/K7/8/k7 w

But it is won after a King move, so still no problem.

[lucasart]

I don't think there are positions where white can promote (w/o being captured) and it's draw regardless of the promotion piece he chooses (or even if he chooses a king move instead).

[d]8/1P6/8/8/8/K7/8/k7 w

But it is won after a King move, so still no problem.

Nice one! Anyway, the Stockfish code is correct to assume that in all cases where white can promote (w/o losing the pawn) it's a win. KPK bitbases don't tell what move you should play to win (but that's trivial for the search to find), they just tell you if it's draw or not. So unless we can find a position where white can promote and it's still a draw, regardless of what white plays, then we're fine with the plain and simple promotion condition.

[Sven]

Safe promotions to queen are considered a win,

I suppose you do test for stalemate after the promotion, not just for recapture of the Queen. Having a Queen and the move is always a win.

Note that there are also few KPK positions where queening draws due to stalemate but underpromoting to a rook wins, e.g.:
[d]8/k1P5/2K5/8/8/8/8/8 w - - 0 1

Sven

Good spot.

But that is not a problem. In practice the KPK bitbase only tells you if it's a draw or not (1 bit per position). It doesn't tell you that you should do a queen promotion. So it's not a problem.

I don't think there are positions where white can promote (w/o being captured) and it's draw regardless of the promotion piece he chooses (or even if he chooses a king move instead).

My point was different. The bitbase generation algorithm "check whether queening creates stalemate; if so => draw, otherwise => won" would misevaluate the position above (provided there is no king move creating a known win), and some similar ones. Of course a full EGTB generation would not suffer from that problem since it would be based on already generated KQK, KRK databases. But as I understood the "short cut" way of generating the KPK bitbase should include to statically evaluate those positions where promotion is possible, and I only wanted to point out that doing this correctly would require more than just checking queen promotions and a possible stalemate.

Sven