Stockfish Testing Queue

Finished - 1117 tests

19-02-03 31m PawnEntrySize diff
LLR: -2.95 (-2.94,2.94) [0.00,4.00]
Total: 97377 W: 21484 L: 21264 D: 54629
sprt @ 10+0.1 th 1 This is outside my area of expertise, but I wonder if reducing the unnecessary size of a pawn Entry can replicate the recent SquareDistance[SQUARE_NB][SQUARE_NB] speedup. No functional change.
19-02-03 31m tweak_KingProtector diff
LLR: -2.96 (-2.94,2.94) [0.00,4.00]
Total: 65880 W: 14572 L: 14476 D: 36832
sprt @ 10+0.1 th 1 Opposite effect: increase KingProtector. Test atop PR #1987.
19-02-03 31m tweak_KingProtector diff
LLR: -2.96 (-2.94,2.94) [0.00,4.00]
Total: 26903 W: 5811 L: 5870 D: 15222
sprt @ 10+0.1 th 1 Double effect, tested against PR #1987.
19-02-03 31m PawnEntrySize diff
LLR: -2.96 (-2.94,2.94) [0.00,4.00]
Total: 22879 W: 5049 L: 5123 D: 12707
sprt @ 10+0.1 th 1 Even smaller--thanks @Alayan-stk-2! Since the other test is doing well, see if this is even better. No functional change.
19-02-03 31m MinorBehindPawn_rank1^ diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 10629 W: 2239 L: 2351 D: 6039
sprt @ 10+0.1 th 1 Why do we give MinorBehindPawn bonus for pieces behind friendly pawns and on our home rank? Try excluding this case.
19-02-03 31m MinorBehindPawn_rank1 diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 4525 W: 930 L: 1072 D: 2523
sprt @ 10+0.1 th 1 Same, but home two ranks. For example, both bishops in this position currently receive MinorBehindPawn bonus but arguably should not. 3r2k1/bb1r1ppp/p1p4q/1pP1NP2/4P3/P7/1P2Q1PP/1BR2R1K b - - 0 22
19-02-02 31m tweak_KingProtector diff
LLR: -2.95 (-2.94,2.94) [0.00,4.00]
Total: 15305 W: 3339 L: 3443 D: 8523
sprt @ 10+0.1 th 1 Reduce KingProtector to accommodate the new MinorInKingRing. Test atop PR #1987
19-02-02 31m simplify_KingProtector diff
LLR: -2.96 (-2.94,2.94) [-3.00,1.00]
Total: 3865 W: 778 L: 951 D: 2136
sprt @ 10+0.1 th 1 An aggressive attempt at simplification. Can we use the similarity between the newly-passed MinorInKingRing (less king danger penalty for nearby minors) to simplify away KingProtector (penalty for minors far from the king) entirely? Test atop PR #1987.
19-02-01 31m pawns_rankpenalty diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 15671 W: 3401 L: 3488 D: 8782
sprt @ 10+0.1 th 1 Results seem roughly neutral, so try +50% effect size.
19-02-01 31m KingProtectorKD diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 2923 W: 580 L: 731 D: 1612
sprt @ 10+0.1 th 1 Now that we always evaluate king safety, we can selectively apply certain heuristics only if the king is in danger or only if it safe. HinderMinor penalized minors close to the king if the king was safe...what if we only apply KingProtector (penalizing faraway minors) if the king is in danger?
19-01-31 31m pawns_rankpenalty diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 20228 W: 4456 L: 4520 D: 11252
sprt @ 10+0.1 th 1 Based on two observations: (a) increasing pawn scores with increasing rank, in response to "fawn pawn" positions, nearly passed; (b) many "fawn pawns" seem to be isolated or supported by a backward pawn. Therefore, rather than strictly increasing values for all advanced pawns, merely decrease Isolated, Backward, and Doubled penalties with increasing rank. No change on relative RANK_2.
19-01-31 31m pawns_rankpenalty diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 18997 W: 4146 L: 4217 D: 10634
sprt @ 10+0.1 th 1 Same, but don't change Doubled.
19-01-31 31m simplify_kingdanger2 diff
LLR: 0.05 (-2.94,2.94) [-3.00,1.00]
Total: 418 W: 100 L: 98 D: 220
sprt @ 10+0.1 th 1 Include a constant tweak to CloseEnemies (originally by @snicolet), and make CloseEnemies the only tropism factor.
19-01-29 31m simplify_kingdanger2 diff
LLR: -2.95 (-2.94,2.94) [-3.00,1.00]
Total: 88175 W: 14490 L: 14784 D: 58901
sprt @ 60+0.6 th 1 Merge new master and LTC.
19-01-29 31m simplify_kingdanger2 diff
LLR: 2.96 (-2.94,2.94) [-3.00,1.00]
Total: 48401 W: 10656 L: 10587 D: 27158
sprt @ 10+0.1 th 1 With sincere apologies to @Vizvezdenec...I'm still not sure that we need tropism in both king danger and CloseEnemies. Since CloseEnemies refuses to be simplified, verify that this recent tropism code is still needed in light of even more recent changes to king danger (i.e., applying it in all positions).
19-01-27 31m backward2^ diff
LLR: -2.95 (-2.94,2.94) [0.00,3.50]
Total: 63546 W: 10463 L: 10462 D: 42621
sprt @ 60+0.6 th 1 Speculative LTC for 114K yellow. 1/6 throughput
19-01-27 31m BishopPawnsPassers4 diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 13642 W: 2929 L: 3026 D: 7687
sprt @ 10+0.1 th 1 Take a 68K yellow and double its effect size.
19-01-27 31m simplify_KDeg diff
LLR: -2.94 (-2.94,2.94) [-3.00,1.00]
Total: 11408 W: 2398 L: 2584 D: 6426
sprt @ 10+0.1 th 1 Very aggressive simplification: no endgame kingdanger penalty.
19-01-26 31m backward2 diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 114755 W: 25189 L: 24790 D: 64776
sprt @ 10+0.1 th 1 Try a more complex formula, carefully tuned to provide S(0, 10) extra Backward penalty for rank 2 (which overlaps with the fawn pawn case), then decreases in Backward for more advanced ranks. Rate of increase of bonus (reduction in penalty) slows down as rank increases, reaching a maximum of S(0, 11) at rank 7.
19-01-26 31m backward2 diff
LLR: -2.94 (-2.94,2.94) [0.50,4.50]
Total: 19380 W: 4279 L: 4347 D: 10754
sprt @ 10+0.1 th 1 Still bigger effect: S(0, 10) increase in penalty for rank 2, but S(0, 18) decrease at the maximum (rank 6--we cannot have backward pawns on rank 7).
19-01-26 31m backward2 diff
LLR: -2.95 (-2.94,2.94) [0.00,3.50]
Total: 31472 W: 5153 L: 5241 D: 21078
sprt @ 60+0.6 th 1 LTC: If A or H file, reduce the backward penalty more for pawns on advanced ranks.
19-01-26 31m backward2 diff
LLR: 2.95 (-2.94,2.94) [0.50,4.50]
Total: 34930 W: 7823 L: 7487 D: 19620
sprt @ 10+0.1 th 1 If A or H file, reduce the backward penalty more for pawns on advanced ranks.
19-01-26 31m backward2^ diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 17696 W: 3890 L: 3967 D: 9839
sprt @ 10+0.1 th 1 If not more than one stopper, reduce the backward penalty more for pawns on advanced ranks.
19-01-25 31m BishopPawnsPassers3 diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 11012 W: 2393 L: 2503 D: 6116
sprt @ 10+0.1 th 1 Double effect
19-01-25 31m BishopPawnsPassers3 diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 47863 W: 10518 L: 10447 D: 26898
sprt @ 10+0.1 th 1 Simple per-bishop bonus for seventh-rank passers.
19-01-25 31m BishopPawnsPassers2 diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 40495 W: 8900 L: 8865 D: 22730
sprt @ 10+0.1 th 1 Cleanup by @Rocky640 and bugfix. (I merged @joergoster's bugfix into my other branch locally and confirmed that this one is functionally equivalent.)
19-01-25 31m BishopPawnsPassers diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 52119 W: 11495 L: 11402 D: 29222
sprt @ 10+0.1 th 1 Exclude from pawns on same color as well. (fixed)
19-01-25 31m BishopPawnsPassers diff
LLR: -2.94 (-2.94,2.94) [0.50,4.50]
Total: 68864 W: 15304 L: 15127 D: 38433
sprt @ 10+0.1 th 1 Just exclude (relative) rank 7 passers. Surely we don't want to penalize the defense of these pawns...
19-01-25 31m BishopPawnsPassers diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 39553 W: 8683 L: 8652 D: 22218
sprt @ 10+0.1 th 1 Only exclude highly advanced (relative ranks 6 and 7) passed pawns from BishopPawns.
19-01-25 31m backward1 diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 15878 W: 3433 L: 3519 D: 8926
sprt @ 10+0.1 th 1 Opposite: no mg penalty for singly-stopped backward pawns. Preserve full eg penalty.
19-01-25 31m BishopPawnsPassers diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 10158 W: 2174 L: 2288 D: 5696
sprt @ 10+0.1 th 1 Based on Bryan's analysis of Antifish 125 1-0 SF dev (Game 09 played on 2019.01.18). Don't include our own passed pawns in the BishopPawns penalty--it is, in fact, desirable for our bishop to defend them.
19-01-25 31m backward1 diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 39719 W: 8817 L: 8785 D: 22117
sprt @ 10+0.1 th 1 If there is no more than one stopper, halve the eg penalty of Backward but leave the mg component untouched.
19-01-24 31m backward1 diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 62550 W: 13645 L: 13503 D: 35402
sprt @ 10+0.1 th 1 Half Backward penalty for pawns on the a/h files, since they cannot have as many stoppers.
19-01-24 31m backward1 diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 20475 W: 4461 L: 4524 D: 11490
sprt @ 10+0.1 th 1 Both ideas are related: halve Backward when there aren't many stoppers. So it makes sense to try them together: halve backward for pawns that are on edge files or have no more than one stopper.
19-01-24 31m tweak_TrappedRook diff
LLR: -2.95 (-2.94,2.94) [0.00,4.00]
Total: 30124 W: 6639 L: 6684 D: 16801
sprt @ 10+0.1 th 1 Same as @snicolet's currently promising STC, but experiment with larger effect.
19-01-23 31m TrappedRook_castling3 diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 32601 W: 7123 L: 7127 D: 18351
sprt @ 10+0.1 th 1 Double effect. This is, in effect, similar to @snicolet's currently promising trapped_rook3: TrappedRook if we can castle and 3 * TrappedRook if we cannot. However, it also interpolates between the two cases and applies 2 * TrappedRook if we are partially unable to castle (i.e., able to castle on one side but not the other).
19-01-23 31m backward1 diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 65685 W: 14467 L: 14308 D: 36910
sprt @ 10+0.1 th 1 Half Backward penalty if there is only one stopper (about 10% of bench positions).
19-01-23 31m TrappedRook_castling3 diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 40996 W: 9080 L: 9042 D: 22874
sprt @ 10+0.1 th 1 I generally try to avoid respins of old tests, but there's a fairly obvious, large (multiplicative) interaction between this patch and the recent TrappedRook simplification. Therefore, respin this 60K yellow.
19-01-23 31m backward1 diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 7265 W: 1538 L: 1667 D: 4060
sprt @ 10+0.1 th 1 First attempt looks initially promising. Double effect: no Backward penalty at all if there is only one stopper (but still possibly increment weakUnopposed).
19-01-22 31m pawns_passable~10 diff
LLR: -2.96 (-2.94,2.94) [0.00,3.50]
Total: 12887 W: 2058 L: 2198 D: 8631
sprt @ 60+0.6 th 1 Speculative LTC for a nearly 84K yellow, because it is an endgame-only bonus and therefore more effective at LTC than STC. Low throughput (166), at least to start.
19-01-21 31m pawns_passable diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 83939 W: 18417 L: 18169 D: 47353
sprt @ 10+0.1 th 1 Explicit bonus for a position pointed out by Bryan from TCEC 14 Bonus round 14: Game 60: LC0 1-0 SF dev. Bonus for a pawn that is supported or phalanx and is immediately blocked by an isolated enemy pawn. The idea is that with an assist from other pieces, this can create a passer. Since I suspect this is mostly an endgame advantage, endgame-only bonus. S(0, 15).
19-01-21 31m pawns_passable diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 6696 W: 1388 L: 1519 D: 3789
sprt @ 10+0.1 th 1 Return to the original and best version, but exclude ranks 2 and 3.
19-01-21 31m pawns_passable diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 33145 W: 7300 L: 7301 D: 18544
sprt @ 10+0.1 th 1 Simpler and probably more accurate version suggested by @Rocky640.
19-01-21 31m tweak_candidatePasser diff
LLR: -2.95 (-2.94,2.94) [0.00,4.00]
Total: 30190 W: 6585 L: 6630 D: 16975
sprt @ 10+0.1 th 1 Parameter tweak pointed out by @Rocky640. Allow rank 4 pawns to be candidate passers.
19-01-21 31m pawns_passable diff
LLR: -2.94 (-2.94,2.94) [0.50,4.50]
Total: 52893 W: 11746 L: 11648 D: 29499
sprt @ 10+0.1 th 1 Initial attempt of S(0, 15) looks promising. Try double effect. S(0, 30).
19-01-21 31m pawns_passable^ diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 17256 W: 3834 L: 3913 D: 9509
sprt @ 10+0.1 th 1 S(0, 60)
19-01-21 31m pawns_passable^^ diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 13030 W: 2831 L: 2931 D: 7268
sprt @ 10+0.1 th 1 There is clearly Elo here, but it's difficult to pin down. I typically underestimate the proper size of added evaluation terms, so try increasing further. S(0, 45).
19-01-21 31m pawns_passable diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 3223 W: 671 L: 821 D: 1731
sprt @ 10+0.1 th 1 S(0, 100). Not actually expected to pass, but intended as a quick sanity-check (if it somehow doesn't fail quickly--I need to consider much larger effects than I otherwise would).
19-01-21 31m pawns_passable diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 25247 W: 5470 L: 5510 D: 14267
sprt @ 10+0.1 th 1 S(0, 30), not including the phalanx case. (Require that the pawn be supported.)
19-01-21 31m pawns_passable diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 14086 W: 3035 L: 3130 D: 7921
sprt @ 10+0.1 th 1 Broader than previous tests, but not as broad as the last one: the opponent pawn is immediately blocking us or one push away.