Stockfish Testing Queue

Finished - 1117 tests

19-01-21 31m pawns_passable diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 12646 W: 2739 L: 2841 D: 7066
sprt @ 10+0.1 th 1 S(0, 15) again, but broaden the conditions to include any opposed case, not just an immediate blockade.
19-01-21 31m pawns_passable diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 10764 W: 2343 L: 2454 D: 5967
sprt @ 10+0.1 th 1 Try instead equal mg and eg. S(15, 15).
19-01-20 31m lmr_pawnOnPin diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 32110 W: 7045 L: 7051 D: 18014
sprt @ 10+0.1 th 1 Also use the narrower scope of pawn pushes on this branch.
19-01-20 31m extend_pawnOnPin diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 12497 W: 2719 L: 2822 D: 6956
sprt @ 10+0.1 th 1 Rather than the full pawn attacks span, which may be to expansive, consider just the to_sq(move) and the next push.
19-01-20 31m lmr_pawnOnPin diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 22908 W: 5005 L: 5056 D: 12847
sprt @ 10+0.1 th 1 Exclude pinned enemy pawns.
19-01-20 31m extend_pawnOnPin diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 22132 W: 4829 L: 4884 D: 12419
sprt @ 10+0.1 th 1 Exclude enemy pawns on this branch too.
19-01-20 31m fawn8 diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 15698 W: 3400 L: 3487 D: 8811
sprt @ 10+0.1 th 1 Fawn pawns if they have doubled rooks. Also try smaller effect (S(15, 15)); I suspect 30 may be too large.
19-01-20 31m lmr_pawnOnPin diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 11518 W: 2479 L: 2587 D: 6452
sprt @ 10+0.1 th 1 Decrease reduction for pawn pushes alongside enemy pinned pieces.
19-01-20 31m extend_pawnOnPin^ diff
LLR: -2.94 (-2.94,2.94) [0.50,4.50]
Total: 11337 W: 2448 L: 2556 D: 6333
sprt @ 10+0.1 th 1 Extension for pawn moves on files adjacent to the enemy's pinned pieces, if they appear to be pinned along the neighboring file by our rook or queen.
19-01-20 31m extend_pawnOnPin diff
LLR: -2.94 (-2.94,2.94) [0.50,4.50]
Total: 7469 W: 1575 L: 1702 D: 4192
sprt @ 10+0.1 th 1 No neighboring rook or queen requirement.
19-01-20 31m fawn8 diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 13331 W: 2909 L: 3008 D: 7414
sprt @ 10+0.1 th 1 File-independent fawn pawn definition.
19-01-20 31m fawn8 diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 13049 W: 2783 L: 2883 D: 7383
sprt @ 10+0.1 th 1 First attempt at an idea by Bryan based on Stockfish dev 0 - 1 AntiFish Mark 125, game 16. Fawn pawn bonus only if our rook has a safe landing square on the enemy's back two ranks.
19-01-19 31m ocb_semiDoubled diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 23499 W: 5149 L: 5197 D: 13153
sprt @ 10+0.1 th 1 Subtract the number of semi-doubled pawns from the pawn asymmetry in OCB positions.
19-01-19 31m combo_F_PF diff
LLR: -2.94 (-2.94,2.94) [0.00,4.00]
Total: 9943 W: 2164 L: 2289 D: 5490
sprt @ 10+0.1 th 1 Much smaller combo, of two recent promising LTC [0, 4]s (@snicolet fawn2^, 147K yellow; @Vizvezdenec KDreodree3, 64K yellow).
19-01-18 31m ocb_semiDoubled diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 22016 W: 4788 L: 4844 D: 12384
sprt @ 10+0.1 th 1 Re-attempting this idea: now considering both sides' semi-doubled pawns.
19-01-18 31m tropism_cap diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 21429 W: 4624 L: 4683 D: 12122
sprt @ 10+0.1 th 1 Half effect: 14 (5% of bench positions).
19-01-18 31m tropism_cap diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 11414 W: 2492 L: 2600 D: 6322
sprt @ 10+0.1 th 1 Results so far appear positive but insufficient to pass. No additional tropism for CloseEnemies > 10. Although this doesn't seem much lower than 12, it actually doubles the effect (22% of bench positions are affected, rather than 11%).
19-01-18 31m tropism_cap diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 32750 W: 7283 L: 7285 D: 18182
sprt @ 10+0.1 th 1 Experiment with a maximum tropism value. No additional CloseEnemies for tropism > 12 (about 11% of positions).
19-01-18 31m TrappedRook_castling2 diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 22110 W: 4883 L: 4938 D: 12289
sprt @ 10+0.1 th 1 In master, we double TrappedRook if we cannot castle kingside and we cannot castle queenside. Instead, double if either is true.
19-01-18 31m TrappedRook_castling diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 21222 W: 4599 L: 4659 D: 11964
sprt @ 10+0.1 th 1 Same, but 175% (multiply by 7/4).
19-01-18 31m TrappedRook_castling^ diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 9439 W: 1976 L: 2094 D: 5369
sprt @ 10+0.1 th 1 This branch appears possibly promising, but it seems that the TrappedRook side and other side's castling rights need to be treated the same as each other. However, they do not need to be precisely between 1 and 2 (3/2) as in my first attempt. 100% TrappedRook if we can castle either side and 200% if neither, as in master, but 125% (multiply by 5/4) if we can castle on precisely one side.
19-01-18 31m ThreatByKing_KD diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 12923 W: 2776 L: 2877 D: 7270
sprt @ 10+0.1 th 1 HinderMinor demonstrated that, now that we always evaluate king danger, we have a simple way to add heuristics which should only be applied for a safe king. I wonder if ThreatByKing should be given this treatment: only apply it if kingDanger <= 0.
19-01-17 31m tropism_sqrt diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 7948 W: 1787 L: 1913 D: 4248
sprt @ 10+0.1 th 1 Speed-up and bugfix by @vondele.
19-01-17 31m tropism_sqrt diff
LLR: -0.08 (-2.94,2.94) [0.50,4.50]
Total: 31 W: 4 L: 8 D: 19
sprt @ 10+0.1 th 1 Try scaling CloseEnemies with the square root of tropism rather than linearly, since quadratic scaling seems to be poor.
19-01-17 31m simplify_CloseEnemies2 diff
LLR: -2.96 (-2.94,2.94) [-3.00,1.00]
Total: 1814 W: 356 L: 529 D: 929
sprt @ 10+0.1 th 1 I believe this is a small simplification. Remove CloseEnemies, replacing CloseEnemies * tropism with simply tropism * tropism for mg and 0 for eg. CloseEnemies is S(8, 0), and tropism averages 7.7.
19-01-17 31m TrappedRook_castling diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 60362 W: 13302 L: 13169 D: 33891
sprt @ 10+0.1 th 1 Just as in master, double TrappedRook if we cannot castle at all and do nothing if we can castle either direction. However, if we can only castle one direction, multiply by 3/2 rather than 1.
19-01-17 31m TrappedRook_castling diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 14569 W: 3141 L: 3234 D: 8194
sprt @ 10+0.1 th 1 First attempt looks promising so far. Try weighting so that the TrappedRook's castlingRight is more important. Penalty multiplier 1 if we can castle either way, 4/3 if we only cannot castle on the other side, 5/3 if we only cannot castle on the trapped side, 2 if we cannot castle either way.
19-01-17 31m TrappedRook_castling diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 11699 W: 2471 L: 2578 D: 6650
sprt @ 10+0.1 th 1 Flip the 4/3 and 5/3 weights in the previous attempt.
19-01-17 31m HinderMinorEdge2 diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 10672 W: 2307 L: 2419 D: 5946
sprt @ 10+0.1 th 1 Retest against master rather than the PR. This is my best version against the PR (72K yellow) which did not change the value of HinderMinor itself, which I suspect may have produced my one poor result against master.
19-01-15 31m simplify_outpostB2 diff
LLR: -2.95 (-2.94,2.94) [-3.00,1.00]
Total: 215767 W: 47406 L: 47998 D: 120363
sprt @ 10+0.1 th 1 Remove the bishop outposts but leave the Outpost array intact, in case those specific values are important. (This only differs from the best version so far in the mg component of unsupported case.) Conflicts with PR #1946. This should probably be my last attempt...
19-01-15 31m simplify_outpostB diff
LLR: -2.95 (-2.94,2.94) [-3.00,1.00]
Total: 15588 W: 2414 L: 2591 D: 10583
sprt @ 60+0.6 th 1 LTC: Remove bishop outposts, also converting to a single Score as in PR #1946. Use file-specific PSQT compensation tweaks. Hopefully, this version scales better...
19-01-15 31m simplify_outpostB diff
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 44323 W: 9752 L: 9675 D: 24896
sprt @ 10+0.1 th 1 File-specific PSQT compensation tweaks, since certain files appear much more likely to contain outposts than others in bench. Tweaks average S(3, 1).
19-01-15 31m simplify_outpostB diff
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 56394 W: 12319 L: 12266 D: 31809
sprt @ 10+0.1 th 1 Half PSQT compensation--S(3, 1)--since full S(6, 2) performs very poorly.
19-01-14 31m simplify_outpostB diff
LLR: -2.96 (-2.94,2.94) [-3.00,1.00]
Total: 55100 W: 9018 L: 9260 D: 36822
sprt @ 60+0.6 th 1 LTC: Remove bishop outposts, also converting to a single Score as in PR #1946.
19-01-15 31m simplify_outpostB diff
LLR: -2.96 (-2.94,2.94) [-3.00,1.00]
Total: 4390 W: 873 L: 1046 D: 2471
sprt @ 10+0.1 th 1 After evaluating the frequencies of each of the four bishop outpost possibilities (pawn supported/not, current/reachable), it seems like a tweak of S(6, 2) maintains roughly the same average bonus. PSQT compensation.
19-01-14 31m ocb_semiDoubled^ diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 32104 W: 7059 L: 7065 D: 17980
sprt @ 10+0.1 th 1 Inspired by the endgame of TCEC Div P Game 154, Ethereal - LCZero (currently ongoing), a classic OCB problem in which Ethereal, Stockfish, and Komodo all seem to overestimate the "strong" side. Critical to white's "advantage" is a useless pair of pawns on the h-file. Define "semi-doubled" pawns to be any pawns on the same file as another pawn of the same color, and subtract these from the strongSide pawn count when calculating scale factor. (We already define "doubled" to be the narrower case where one pawn is immediately in front of the other.)
19-01-14 31m mgeg1 diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 32267 W: 7101 L: 7106 D: 18060
sprt @ 10+0.1 th 1 Previous test was neutral, so double effect (20).
19-01-14 31m ocb_semiDoubled diff
LLR: -2.94 (-2.94,2.94) [0.50,4.50]
Total: 18547 W: 4121 L: 4193 D: 10233
sprt @ 10+0.1 th 1 Half effect (treat the doubled pair as one pawn).
19-01-14 31m mgeg1 diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 4350 W: 919 L: 1063 D: 2368
sprt @ 10+0.1 th 1 The original logic currently is at least neutral (unlike the others), with a positive score with effect size 20. Try 30.
19-01-14 31m simplify_outpostB diff
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 21867 W: 4885 L: 4764 D: 12218
sprt @ 10+0.1 th 1 Remove bishop outposts, also converting to a single Score as in PR #1946.
19-01-14 31m mgeg1 diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 8729 W: 1871 L: 1992 D: 4866
sprt @ 10+0.1 th 1 Always use eg in eval (effect size 20), but don't make any tweaks to compensate.
19-01-14 31m hinderMinorEdge^^ diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 9055 W: 1919 L: 2039 D: 5097
sprt @ 10+0.1 th 1 This patch was a 86K yellow against PR #1945. Verify its Elo gain against master.
19-01-14 31m mgeg1 diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 5456 W: 1122 L: 1259 D: 3075
sprt @ 10+0.1 th 1 Continuing this collaboration with @xoto10. Use min on the mg term to ensure the multipliers still add to PHASE_MIDGAME, preserving the faster division by 128 (rather than division by an arbitrary int).
19-01-14 31m hinderMinorEdge diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 19512 W: 4292 L: 4360 D: 10860
sprt @ 10+0.1 th 1 Ranks 1 and 8, excluding the corners. S(15, 23). Test against PR #1945.
19-01-14 31m hinderMinorEdge diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 13094 W: 2881 L: 2981 D: 7232
sprt @ 10+0.1 th 1 After steadily improving this branch, it is on the cusp of passing (75K games: LLR = 0.38). Try only excluding our back rank, not both rank 1 and 8. This is a small, mostly-endgame change--but we're already very close to passing. (Bench also changes substantially.) S(15, 23). Test against PR #1945.
19-01-14 31m hinderMinorEdge diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 86058 W: 18938 L: 18679 D: 48441
sprt @ 10+0.1 th 1 Same, but S(15, 23). Test against PR #1945.
19-01-14 31m hinderMinorEdge diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 72338 W: 15839 L: 15648 D: 40851
sprt @ 10+0.1 th 1 I have thus far considered all edges together, but the A/H files may be different from ranks 1/8. Ranks 1/8 only, no compensation for now. Test against PR #1945.
19-01-14 31m mgeg1 diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 24843 W: 5471 L: 5513 D: 13859
sprt @ 10+0.1 th 1 Variant on @xoto's idea, using std::max rather than +. Always use some eg in eval, minimum 10.
19-01-14 31m hinderMinorEdge^ diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 19546 W: 4245 L: 4313 D: 10988
sprt @ 10+0.1 th 1 Moving to ranks-only (as opposed to edge ranks and edge files) is a 67K+ yellow; compensation tweaks also produced 60-70K yellows. If we try both, maybe there's a green. Try each of the two promising tweaks. S(15, 15). Test against PR #1945.
19-01-14 31m hinderMinorEdge^ diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 25473 W: 5585 L: 5624 D: 14264
sprt @ 10+0.1 th 1 I have thus far considered all edges together, but the A/H files may be different from ranks 1/8. A/H files only, no compensation for now. Test against PR #1945.