This book contains about 34000 engames, which should allow for testing endgame patches.
The positions have been selected out of ~800000 games from fishtest ltc tests and ccrl (>3000Elo),
to make them realistic (because played in actual games). From this set of games,
only decisive games have been retained, to make sure there is win potential.
The first position of each game was selected for which the non-pawn material was small
(knights 3 + bishops 3 + rooks 5 + queens 9 < 20), to yield endgame (or late middlegame) positions.
These positions were subsequently analyzed with SF master up to depth 30 to assert that
abs(score)<150 and abs(score)>20. In this way the selected positions are neither easy wins nor easy draws,
but given the game was decisive likely 'non-trivial' and thus interesting for testing purposes.
In game play this shows as following:
sfdev tc=60.0+0.6 vs sf10 tc=60.0+0.6
Score of master vs sf10: 574 - 434 - 1522 [0.528] 2530
Elo difference: 19.2 +/- 8.5, LOS: 100.0 %, DrawRatio: 60.2 %
i.e. at LTC, reasonable Elo gain between sfdev and sf10,
and with a draw rate that is lower than 8moves_v3.pgn.
Giving sfdev time-odds, it also is clear that these positions are not easy,
but have potential, i.e. sf10 has few wins, but sfdev converts a large fraction:
sfdev tc=20.0+0.2 vs sf10 tc=5+0.05
Score of master vs sf10: 435 - 80 - 585 [0.661] 1100
Elo difference: 116.3 +/- 13.8, LOS: 100.0 %, DrawRatio: 53.2 %
sfdev tc=40.0+0.4 vs sf10 tc=2.5+0.025
Score of master vs sf10: 383 - 22 - 355 [0.738] 760
Elo difference: 179.5 +/- 17.7, LOS: 100.0 %, DrawRatio: 46.7 %
sfdev tc=60.0+0.6 vs sf10 tc=1.0+0.01
Score of master vs sf10: 491 - 4 - 215 [0.843] 710
Elo difference: 291.9 +/- 23.0, LOS: 100.0 %, DrawRatio: 30.3 %
This book contains about 34000 engames, which should allow for testing endgame patches.
The positions have been selected out of ~800000 games from fishtest ltc tests and ccrl (>3000Elo), to make them realistic (because played in actual games). From this set of games, only decisive games have been retained, to make sure there is win potential. The first position of each game was selected for which the non-pawn material was small (knights 3 + bishops 3 + rooks 5 + queens 9 < 20), to yield endgame (or late middlegame) positions. These positions were subsequently analyzed with SF master up to depth 30 to assert that abs(score)<150 and abs(score)>20. In this way the selected positions are neither easy wins nor easy draws, but given the game was decisive likely 'non-trivial' and thus interesting for testing purposes.
In game play this shows as following:
sfdev tc=60.0+0.6 vs sf10 tc=60.0+0.6 Score of master vs sf10: 574 - 434 - 1522 [0.528] 2530 Elo difference: 19.2 +/- 8.5, LOS: 100.0 %, DrawRatio: 60.2 %
i.e. at LTC, reasonable Elo gain between sfdev and sf10, and with a draw rate that is lower than 8moves_v3.pgn.
Giving sfdev time-odds, it also is clear that these positions are not easy, but have potential, i.e. sf10 has few wins, but sfdev converts a large fraction:
sfdev tc=20.0+0.2 vs sf10 tc=5+0.05 Score of master vs sf10: 435 - 80 - 585 [0.661] 1100 Elo difference: 116.3 +/- 13.8, LOS: 100.0 %, DrawRatio: 53.2 %
sfdev tc=40.0+0.4 vs sf10 tc=2.5+0.025 Score of master vs sf10: 383 - 22 - 355 [0.738] 760 Elo difference: 179.5 +/- 17.7, LOS: 100.0 %, DrawRatio: 46.7 %
sfdev tc=60.0+0.6 vs sf10 tc=1.0+0.01 Score of master vs sf10: 491 - 4 - 215 [0.843] 710 Elo difference: 291.9 +/- 23.0, LOS: 100.0 %, DrawRatio: 30.3 %