Remove KNPK endgame

[protonspring]

This is a functional simplification that removes the KNPK endgame.

In more than 100k local KNPK games, there is no discernible difference between master and master with this endgame removed: master:42971, patch:42973, draws: 3969. Removal does not seem to regress in normal games.

STC
LLR: 2.94 (-2.94,2.94) {-1.50,0.50}
Total: 46390 W: 8998 L: 8884 D: 28508
Ptnml(0-2): 707, 5274, 11163, 5300, 751
https://tests.stockfishchess.org/tests/view/5e83b18ee42a5c3b3ca2ef02

LTC
LLR: 2.94 (-2.94,2.94) {-1.50,0.50}
Total: 44768 W: 5863 L: 5814 D: 33091
Ptnml(0-2): 251, 3918, 14028, 3905, 282
https://tests.stockfishchess.org/tests/view/5e84a82a4411759d9d0984f4

Bench 5147355

[vondele]

No real difference testing locally either, for a set of endgames that might lead to KNPvK:

Score of patch vs master: 6131 - 6188 - 27681  [0.499] 40000
Elo difference: -0.5 +/- 1.9, LOS: 30.4 %, DrawRatio: 69.2 %

[protonspring]

Hmmm. . so there is something. I assume these are KNPsvKPs positions?
Do you have your book that I can verify?

[vondele]

The result is quite consistent with 0 Elo... however, that's the book (basically everything that matched KN?PvK?.epd KN??PvK??.epd) https://www.dropbox.com/s/v81tb5ok1sr7f2q/book_KNPvK_andmore.epd.zip?dl=0 TC=1.0+0.01

[protonspring]

This book is testing very poorly for this patch.
TC2+0.02: patch vs master = 98 - 114 (740).

Should I close this?

EDIT: I believe I had accidentally run patch in debug mode. Better results below.

[vondele]

I'd leave test running for a while... just to make sure this is not noise (cutechess-cli has these nice error bars on results, they are worth reporting...). I'm guessing this is 0 within error bars.

However, I think there might be an important point to make. The endgame code is not only about how one plays the endgame, but also what the balance is between going into an endgame or staying out of it. So, I think one needs to do endgame code testing with 'nearby' endgames.

[protonspring]

Yes. I am learning that. I will let this run for a while and report. If it's bad, i will just close this.

[protonspring]

Sry. . I gotta use the computer for other things. Current results are:
patch vs master: 655-689-3505. It looks like this endgame does at least something. Closing.

EDIT: I wasn't running clean versions. better results below.

[vondele]

And here, using 2+0.2TC:
Score of patch vs master: 5740 - 5741 - 28519 [0.500] 40000
Elo difference: -0.0 +/- 1.8, LOS: 49.6 %, DrawRatio: 71.3 %

[protonspring]

Hmm. . maybe I screwed something up. I will retry when I get another spare computer and let it run overnight.

[protonspring]

After about 5k games: master vs patch: 785 - 776 - 3657.

I guess it looks like this is safe to remove.

[vondele]

Thanks!