Stockfish Testing Queue

Pending - 0 tests 0.0 hrs

None

Active - 0 tests

Finished - 1117 tests

19-03-18 31m KDWeak_pawn3 diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 56854 W: 12598 L: 12481 D: 31775
sprt @ 10+0.1 th 1 Tuned values.
19-03-17 31m tune_KDWeak_pawn diff
28460/30000 iterations
59336/60000 games played
60000 @ 20+0.2 th 1 This was very close to passing on March 6 (STC 13K green, LTC 141K yellow) and we haven't had an eval commit since, so it probably still is. With the framework empty, try to tune king danger parameters around it.
19-03-17 31m ThreatByPawnPush_slider diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 28510 W: 6285 L: 6308 D: 15917
sprt @ 10+0.1 th 1 I know I'm more than a week late, but try to rectify a blind spot notice by Bryan in SF 0-1 LC0, Game 17 of TCEC 14 Bonus Final: Giuoco Pianissimo. Count a TBPP square as relatively safe if it is in front of a pawn supported from behind by a rook or queen. Currently this square is not considered "safe" if attacked by enemies, because we only check if the square is directly defended by us and overlook this case.
19-03-17 31m ThreatByPawnPush_slider diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 24454 W: 5432 L: 5475 D: 13547
sprt @ 10+0.1 th 1 Rooks only.
19-03-09 31m KDWeak_pawn2 diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 41494 W: 9230 L: 9189 D: 23075
sprt @ 10+0.1 th 1 Value 190.
19-03-09 31m KDWeak_pawn2^ diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 14215 W: 3074 L: 3168 D: 7973
sprt @ 10+0.1 th 1 Clearly the previous tweaks on this branch were too large. Half effect. Value 180.
19-03-09 31m KDWeak_pawn2^ diff
LLR: -2.94 (-2.94,2.94) [0.50,4.50]
Total: 14359 W: 3135 L: 3228 D: 7996
sprt @ 10+0.1 th 1 Parameter tweak compensation to the coefficient as well. Value 195.
19-03-09 31m KDWeak_pawn2 diff
LLR: -3.41 (-2.94,2.94) [0.50,4.50]
Total: 6549 W: 1392 L: 1550 D: 3607
sprt @ 10+0.1 th 1 Value 175.
19-03-08 31m lmrdepth1 diff
LLR: -2.96 (-2.94,2.94) [-3.00,1.00]
Total: 56256 W: 9408 L: 9654 D: 37194
sprt @ 60+0.6 th 1 LTC for @xoto10: Remove movecount == 1 test in lmr.
19-03-07 31m passed_double^ diff
LLR: -2.94 (-2.94,2.94) [0.50,4.50]
Total: 38045 W: 8482 L: 8457 D: 21106
sprt @ 10+0.1 th 1 Parameter tweaks around @ElbertoOne's 61K STC yellow for a nearly-empty framework. k -= 6
19-03-07 31m passed_double diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 23014 W: 5027 L: 5078 D: 12909
sprt @ 10+0.1 th 1 k -= 2
19-03-07 31m KDWeak_pawn^ diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 33373 W: 7409 L: 7408 D: 18556
sprt @ 10+0.1 th 1 This is extraordinarily close to passing (STC 13K green, LTC 141K yellow). Moreover, the constant term seems incredibly sensitive (STC value 25: 19K red; 30: 13K green; 35: 40K yellow). Since the framework is nearly empty, try some some tweaks around this constant in hopes of finding a slightly better (and therefore sufficient to pass) value. Value 28.
19-03-07 31m KDWeak_pawn diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 23177 W: 5172 L: 5221 D: 12784
sprt @ 10+0.1 th 1 Value 32.
19-03-06 31m KDWeak_pawn diff
LLR: -2.96 (-2.94,2.94) [0.00,3.50]
Total: 141695 W: 23939 L: 23716 D: 94040
sprt @ 60+0.6 th 1 LTC: A square is not "weak" if it is an enemy pawn blocked immediately by our king (with compensation in kingDanger constant).
19-03-06 31m KDWeak_pawn diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 40078 W: 8889 L: 8855 D: 22334
sprt @ 10+0.1 th 1 Double compensation.
19-03-06 31m KDWeak_pawn diff
LLR: 2.94 (-2.94,2.94) [0.50,4.50]
Total: 13073 W: 2976 L: 2748 D: 7349
sprt @ 10+0.1 th 1 With compensation.
19-03-05 31m KDWeak_pawn diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 19141 W: 4197 L: 4267 D: 10677
sprt @ 10+0.1 th 1 I admit I'm not sure why I'm getting inconsistent bench problems--maybe it's the case where ksq + Up goes off the board: relative_rank(Us, ksq) == RANK_8? Try handling this. If this doesn't work, I welcome anyone who can spot the error.
19-03-05 31m KDWeak_pawn diff
LLR: 0.14 (-2.94,2.94) [0.50,4.50]
Total: 345 W: 75 L: 66 D: 204
sprt @ 10+0.1 th 1 A square is not "weak" if it is an enemy pawn blocked immediately by our king.
19-03-05 31m PieceOnQueen_immediate diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 30599 W: 6740 L: 6753 D: 17106
sprt @ 10+0.1 th 1 @Vizvezdenec's suggestion.
19-03-05 31m DoubleAttackK diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 26682 W: 5833 L: 5866 D: 14983
sprt @ 10+0.1 th 1 Opposite effect, which just happens to almost perfectly replicate the largest effect in @SFisGOD's current LTC SPSA. Perhaps combining the best pieces of these two king danger ideas, one [0.5, 4.5] and one [0, 4], will have some synergy...
19-03-05 31m PieceOnQueen_immediate diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 6936 W: 1475 L: 1605 D: 3856
sprt @ 10+0.1 th 1 Expanded the logic of the current best version. Now our pieces that are (a) not pawn-defended (b) doubly enemy attacked, including by the enemy queen are considered under the queen's attack, rather than only ones that are literally hanging.
19-03-05 31m DoubleAttackK diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 25922 W: 5631 L: 5668 D: 14623
sprt @ 10+0.1 th 1 @Vizvezdenec's nearly-passed Elo gainer + compensatory tweak to kingDanger constant, calculated to maintain the same average kingDanger across bench positions.
19-03-05 31m PieceOnQueen_immediate diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 7899 W: 1685 L: 1811 D: 4403
sprt @ 10+0.1 th 1 I've seen a few tests which appear to be based on this aspect of Bryan's analysis of Game 187 of TCEC bonus, but I don't think we've tried the most direct implementation: no KnightOnQueen or SliderOnQueen (for two-step threats) if we already have an immediate threat.
19-03-02 31m Isolated_enemypasser diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 23661 W: 5199 L: 5246 D: 13216
sprt @ 10+0.1 th 1 A pawn is isolated if it is very backward (no friendly pawns behind us or in our pawn attacks) and it cannot be pushed to join friendly pawns without creating enemy passers. (Not sure why bench increases so dramatically...)
19-02-23 31m extend_pawnPushEG diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 27468 W: 5949 L: 5978 D: 15541
sprt @ 10+0.1 th 1 Much less restrictive NPM condition: NPM <= Rook + Bishop for both sides.
19-02-23 31m extend_pawnPushEG diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 36412 W: 8011 L: 7996 D: 20405
sprt @ 10+0.1 th 1 Extension instead (same conditions).
19-02-23 31m lmr_pawnPushEG diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 14393 W: 3140 L: 3233 D: 8020
sprt @ 10+0.1 th 1 Inspired by TCEC Superfinal Game 84, Stockfish 1/2 - 1/2 LC0, where it seems that SF may have missed a win (or at the very least, overlooked that it was headed towards a draw). Eval plummets from +1 to equal in two moves. The first is in response to 35. ...b4, an apparent short-term pawn sacrifice that saddles White with indefensible isolated doubled b-pawns. Try reducing this pawn push less, starting with a very narrow condition and potentially broadening it in future tests. Less LMR for pawn moves in rook-less, queen-less positions that attack both enemy non-pawn pieces (obligating a capture) and the only enemy pawn on adjacent files (creating isolated pawns).
19-02-23 31m SafeCheck_occupied diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 8092 W: 1744 L: 1869 D: 4479
sprt @ 10+0.1 th 1 +25% SafeCheck if the safe check is a threat against non-pawn material.
19-02-22 31m SafeCheck_occupied diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 25373 W: 5596 L: 5635 D: 14142
sprt @ 10+0.1 th 1 Opposite effect.
19-02-22 31m SafeCheck_occupied diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 12980 W: 2790 L: 2890 D: 7300
sprt @ 10+0.1 th 1 Less SafeCheck (RookSafeCheck, KnightSafeCheck, etc.) if all the checking options are occupied by non-pawn material. The idea is that if these pieces can simply move once, the square is suddenly defended; otherwise, the obvious capture probably is handled by search and reducing the static eval is irrelevant.
19-02-22 31m SafeCheck_occupied diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 10388 W: 2165 L: 2278 D: 5945
sprt @ 10+0.1 th 1 Only for rooks and queens.
19-02-22 31m lmr_promotion diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 29839 W: 6550 L: 6567 D: 16722
sprt @ 10+0.1 th 1 Merge the two best-performing versions of the same idea: more extension and less reduction for any promotion.
19-02-20 31m extend_castlefawn diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 25897 W: 5723 L: 5759 D: 14415
sprt @ 10+0.1 th 1 Even more castling extension if castling directly into a fawn pawn. (Note for approver: castling moves are encoded as "king takes rook", so the to_sq for kingside castling is actually "on" the H file (and queenside, A file). Please let me know if someone has a less ugly way of coding this.)
19-02-20 31m lmr_promotion diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 14479 W: 3180 L: 3273 D: 8026
sprt @ 10+0.1 th 1 Less LMR for all promotions was promising, but just for queens was terrible. Try less LMR for underpromotion.
19-02-20 31m lmr_promotion^ diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 60223 W: 13305 L: 13172 D: 33746
sprt @ 10+0.1 th 1 Based on @mstembera/@vondele ideas for extensions, but applied to LMR. Less reduction for promotions.
19-02-20 31m lmr_promotion diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 18835 W: 4113 L: 4184 D: 10538
sprt @ 10+0.1 th 1 Less reduction for promotions (queen promotion only).
19-02-20 31m outpost_futurePasser diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 13339 W: 2909 L: 3008 D: 7422
sprt @ 10+0.1 th 1 Removing eg component. S(10, 0).
19-02-20 31m outpost_futurePasser diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 14382 W: 3107 L: 3200 D: 8075
sprt @ 10+0.1 th 1 Simple S(10, 10) bonus.
19-02-18 31m outpost_futurePasser diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 17424 W: 3787 L: 3865 D: 9772
sprt @ 10+0.1 th 1 Initial attempt based on Bryan's analysis of TCEC 14 Superfinal Game 49. More Outpost bonus for an occupied, pawn-defended rank-6 outpost whose exchange creates a rank-6 passer. Require that the opponent have a minor to exchange for this outpost.
19-02-17 31m TrappedRook_fawn diff
LLR: -2.94 (-2.94,2.94) [0.50,4.50]
Total: 13557 W: 2922 L: 3019 D: 7616
sprt @ 10+0.1 th 1 Bugfix: require that the enemy-pawn-attacked squares on our second rank not be occupied by our pieces.
19-02-17 31m KingMoves diff
LLR: -2.96 (-2.94,2.94) [0.00,3.50]
Total: 19744 W: 3169 L: 3290 D: 13285
sprt @ 60+0.6 th 1 There's no reason to leave the framework empty for hours with a promising test waiting...speculative LTC for @MJZ1977's 170K STC yellow.
19-02-17 31m TrappedRook_fawn diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 10526 W: 2236 L: 2348 D: 5942
sprt @ 10+0.1 th 1 Half effect.
19-02-17 31m TrappedRook_fawn diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 21356 W: 4651 L: 4710 D: 11995
sprt @ 10+0.1 th 1 Double trapped rook if the enemy has a fawn pawn in the area. Inspired by TCEC Superfinal Game 63, Lc0 1-0 SF.
19-02-16 31m lmr_attack diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 13303 W: 2887 L: 2986 D: 7430
sprt @ 10+0.1 th 1 This is by far the largest change I have attempted in search, where I lack experience; please correct me if I have made errors. It has long been suggested to do less reduction for the attacking side if the weak side has high king danger, but king danger is too expensive to calculate in search. Maybe lazy eval isn't! If we have lots of non-pawn material, and an eval well beyond what can be explained by lazy evaluation, we almost certainly have some sort of attack or dynamic advantage. Apply less reduction in this case.
19-02-16 31m lmr_attack diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 9691 W: 2065 L: 2182 D: 5444
sprt @ 10+0.1 th 1 Even more lazy (and hopefully more efficient): do not probe Pawns or evaluate them.
19-02-15 31m combo_TOP_outpost diff
LLR: -2.96 (-2.94,2.94) [0.00,4.00]
Total: 24335 W: 5250 L: 5319 D: 13766
sprt @ 10+0.1 th 1 Combo some old promising tweaks of mine, since the framework is nearly empty. (I have no idea whether they will still perform well...)
19-02-14 31m passed_protection diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 9228 W: 1984 L: 2103 D: 5141
sprt @ 10+0.1 th 1 Check whether this substantially narrower case (candidate passers only) can therefore tolerate larger effect. Double effect (divisor 16 -> 8).
19-02-14 31m passed_protection diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 41791 W: 9259 L: 9217 D: 23315
sprt @ 10+0.1 th 1 Opposite effect: candidate passers only.
19-02-14 31m threat_strongqueen3 diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 26079 W: 5770 L: 5805 D: 14504
sprt @ 10+0.1 th 1 In November, this was a 70K STC and 108K LTC yellow--back when we used [0, 5] bounds for both. Respin with new bounds. Master already has ThreatByRook[QUEEN] but applies this only to "weak" queens, but queens are always weak to threats by lesser pieces. Give ThreatByRook[QUEEN] / 2 bonus for a rook attack against a "strong" queen.
19-02-14 31m passed_protection diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 10366 W: 2217 L: 2330 D: 5819
sprt @ 10+0.1 th 1 My take on MJZ's idea. Don't apply this increased bonus for candidate passers, only for "real" ones.