Fruit does not see draw - and resigns!

[mclane]

i know exactly what tord is talking about.
here how CSTAL plays in your position.
don't be irritared by the wrong NPS.
the NPS-variable is out of range.


[d]kn6/2B5/8/1P6/1K6/P7/8/8 b - - 0 1

000:00 0000n/s d1 -9.32 b8c6 b5c6 a8a7 b4b5 NULL
000:00 0000n/s d1 ! b8a6 b4c3 NULL
000:00 0000n/s d1 -3.11 b8a6 b4a5 a6c7 b5b6 c7d5 a3a4 NULL
000:00 0000n/s d2 -3.11 b8a6 b4a5 a6c7 b5b6 c7d5 a3a4 NULL
000:00 0000n/s d3 ? b8a6 b4a5 NULL
000:00 0000n/s d3 ? b8a6 b4a5 a6c7 b5b6 c7d5 a5a6 NULL
000:00 0000n/s d4 -2.40 b8a6 b4a5 a6c7 b5b6 c7d5 a5a6 NULL
000:00 0000n/s d5 -2.44 b8a6 b4c4 a6c7 b5b6 c7e6 a3a4 a8b7 a4a5 NULL
000:00 0000n/s d6 -2.44 b8a6 b4c4 a6c7 b5b6 c7e6 a3a4 a8b7 a4a5 NULL
000:00 0000n/s d7 ? b8a6 b4c4 a6c7 b5b6 c7e6 a3a4 a8b7 a4a5 b7a6 c4b4 e6d8 b6b7 NULL
000:00 0000n/s d8 -3.08 b8a6 b4c4 a6c7 b5b6 c7e6 a3a4 a8b7 a4a5 b7a6 c4b4 e6d8 b6b7 NULL
000:00 0000n/s d9 -2.95 b8a6 b4c4 a6c7 b5b6 c7e6 a3a4 a8b7 a4a5 b7a6 c4b4 e6d8 b4a4 NULL
000:00 0000n/s d10 -2.95 b8a6 b4c4 a6c7 b5b6 c7e6 a3a4 a8b7 a4a5 b7a6 c4b4 e6d8 b4a4 NULL
000:00 0000n/s d11 ? b8a6 b4c4 NULL
000:00 0000n/s d11 ? b8a6 b4c4 a6c7 b5b6 a8b7 b6c7 b7c7 c4c5 c7b7
000:01 -11367n/s d12 ? b8a6 b4c4 a6c7 b5b6 a8b7 b6c7 b7c7 c4c5 c7b7
000:01 13199n/s d13 ? b8a6 b4c4 a6c7 b5b6 a8b7 b6c7 b7c7 c4c5 c7b7 a3a4 b7a6
000:02 -12740n/s d14 ? b8a6 b4c4 a6c7 b5b6 a8b7 b6c7 b7c7 c4c5 c7b7 a3a4 b7a6
000:03 -11964n/s d15 ? b8a6 b4c4 a6c7 b5b6 a8b7 b6c7 b7c7 c4c5 c7b7 a3a4 b7a6 a4a5
000:04 -10528n/s d16 ? b8a6 b4c4 a6c7 b5b6 a8b7 b6c7 b7c7 c4c5 c7b7 a3a4 b7a6 a4a5
000:08 -16157n/s d17 ? b8a6 b4c4 a6c7 b5b6 a8b7 b6c7 b7c7 c4c5 c7b7 a3a4 b7a6 a4a5
000:11 -15780n/s d18 ? b8a6 b4c4 a6c7 b5b6 a8b7 b6c7 b7c7 c4c5 c7b7 a3a4 b7a6 a4a5
000:17 -15078n/s d19 ? b8a6 b4c4 a6c7 b5b6 a8b7 b6c7 b7c7 c4c5 c7b7 a3a4 b7a6 a4a5
000:24 -15653n/s d20 ? b8a6 b4c4 a6c7 b5b6 a8b7 b6c7 b7c7 c4c5 c7b7 a3a4 b7a6 a4a5

the evaluation in those areas with the ? is draw or in this case 0.17.

[Tord Romstad]

[d]kn6/2B5/8/1P6/1K6/P7/8/8 b - - 0 1

Tord is speaking about such cases.
He would find the draw immediately, an engine without that knowledge like Spike might play Nd7 and lose.

Hello Mischa,

Thanks for posting this position - it helped me discover and fix a bug! Glaurung didn't see the draw, even though it was supposed to have the necessary knowledge. Now it works as intended.

Tord

[glorfindel]

Not really. Drawn endgames with a doubled rook pawn and a bishop of the wrong color are extremely rare in practice. Implementing specific evaluation rules for all sorts of obscure endgames is boring and tedious, and does not really help playing strength. Most programmers tend to concentrate on more general chess knowledge, which can be applied in a wider variety of positions.

My program actually does evaluate the given position as a draw, but I doubt that it has any effect on playing strength whatsoever. It might save half a point once every 10000 games or so, but on the other hand it also makes the program a tiny bit slower in some endgames, which probably costs a tiny fraction of an Elo point. The net effect is totally insignificant. The only reason I have implemented this evaluation rule at all is that I hate to see my program lose positions which even I know how to draw.

Tord

Tord, are you assuming that people use chess engines for measuring their strength in engine-engine matches only?

Some of us use them for analysing games played by humans, so it is not only the elo rating of an engine that counts. For me it is preferable to have a correct evaluation of a theoretical position, even if it costs some elo points in general.

[Tord Romstad]

Tord, are you assuming that people use chess engines for measuring their strength in engine-engine matches only?

No, I don't - but I am assuming that most engine authors are more concerned about strength in engine vs engine matches than in making their engines better as anslysis tools. This is a fairly safe assumption to make. Just look at how few engine authors implement things like MultiPV mode and the possibility of restricting which root moves are searched. Both of these are entirely trivial to implement, but most authors never bother.

I am not saying that engine authors shouldn't bother about implementing specific evaluation rules for such unusual endgame situations (in fact, I suspect that I have far more such code in my program than most amateur authors do). I am only explaining why it shouldn't come as a surprise when a strong engine doesn't have such knowledge implemented.

Tord

[Uri Blass]

It also can help me to discover a bug because movei without knowledge in evaluation and hash based pruning does not find the move inspite of having very bad score for
Nd7 after some minutes

It seems that it does too much pruning and I do not understand the reason.