Stockfish Testing Queue

Finished - 1408 tests

18-09-30 31m RookOpportunity diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 3811 W: 763 L: 875 D: 2173
sprt @ 10+0.1 th 1 An initial "shot in the dark" based on the CCCC game SF 0-1 Houdini mentioned in the forum. Here, SF is up an exchange, but its two rooks cannot do anything, and SF doesn't appear to sense the danger for some time. Define a "rook opportunity" as an enemy weak unopposed pawn, open file, or friendly passer on a semi-open file (i.e., a file where a rook can accomplish something). If we have only one of these, penalize our rook; if we have none, double penalty. (I have no idea what value this penalty should take. Start with S(10, 10).)
18-09-30 31m BishopPawnsStrong diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 9294 W: 1999 L: 2085 D: 5210
sprt @ 10+0.1 th 1 Increase BishopPawns penalty if there are doubly-pawn-protected enemy pawns in the center files, on the same color as our bishop. Bugfix and larger penalty.
18-09-30 31m BishopPawnsPinned diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 8807 W: 1827 L: 1915 D: 5065
sprt @ 10+0.1 th 1 Another small tweak: treat king-pinned pawns as "blocked" for the purposes of the BishopPawns penalty.
18-09-30 31m BishopPawnsEscape diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 7336 W: 1519 L: 1614 D: 4203
sprt @ 10+0.1 th 1 A small tweak: no BishopPawns penalty if our "blocked" pawn is attacked by an enemy pawn. In this case, it can capture and become unblocked.
18-09-23 31m Probcut5 diff
LLR: -3.66 (-2.94,2.94) [0.00,5.00]
Total: 86836 W: 14180 L: 13992 D: 58664
sprt @ 60+0.6 th 1 LTC for @Vizvezdenec: Take 5.
18-09-18 31m ocb_doubled diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 19618 W: 4190 L: 4226 D: 11202
sprt @ 10+0.1 th 1 Double effect with this narrower case.
18-09-18 31m ocb_doubled^ diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 15436 W: 3269 L: 3325 D: 8842
sprt @ 10+0.1 th 1 Effect size 2, and only consider unsupported doubled pawns. The bench barely changes, but on the basis of discussion in https://github.com/official-stockfish/Stockfish/pull/1767 I think this is because there are few instances of this type of position in the set of bench positions.
18-09-18 31m ocb_doubled diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 5893 W: 1189 L: 1291 D: 3413
sprt @ 10+0.1 th 1 Double effect (size 4).
18-09-17 31m ocb_doubled diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 17697 W: 3728 L: 3773 D: 10196
sprt @ 10+0.1 th 1 Often in OCB endings, a pair of doubled pawns are worth no more than a single pawn. Reduce the scale factor if the strong side has doubled pawns.
18-09-09 31m RvsN10 diff
LLR: -2.97 (-2.94,2.94) [0.00,5.00]
Total: 14863 W: 2370 L: 2435 D: 10058
sprt @ 60+0.6 th 1 I don't think it makes sense to leave the framework empty, so I'll submit a speculative LTC of @Vizvezdenec's 54K yellow. Since this only affects the endgame, there is good reason to suspect it may perform better at LTC. Corrected bench.
18-09-08 31m ROF_openable diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 15657 W: 3332 L: 3387 D: 8938
sprt @ 10+0.1 th 1 Give the open file RookOnFile bonus, rather than the semi-open one, if we can open the file on the next turn (if the enemy blocker, which is a pawn, is pawn-attacked but not pawn-defended). Many thanks to Andrew Grant for spotting an opportunity for substantial simplification of this patch.
18-09-08 31m ROF_openable diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 6970 W: 1439 L: 1536 D: 3995
sprt @ 10+0.1 th 1 If the file is semi-open, but open-able, give a bonus which is the mean of RookOnFile[open] and RookOnFile[!open]. (Note: openable is only true if !open, based on its definition.)
18-09-06 31m BlockedOpen diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 6307 W: 1276 L: 1376 D: 3655
sprt @ 10+0.1 th 1 For the purposes of RookOnFile, check whether an otherwise open file is blocked by an enemy non-RQ piece (possibly a pawn). If it is pawn-defended, and we cannot attack it with a pawn, the file is only semi-open.
18-09-03 31m BlockedSemiOpen4 diff
LLR: -2.94 (-2.94,2.94) [0.00,5.00]
Total: 9323 W: 1928 L: 2013 D: 5382
sprt @ 10+0.1 th 1 The outpost may be defended only once, but it must not be pawn-attacked by us in order to count.
18-09-03 31m BlockedSemiOpen4 diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 6904 W: 1446 L: 1543 D: 3915
sprt @ 10+0.1 th 1 Expand (and simplify) the definition of the best one so far, to include outposts only defended once.
18-09-03 31m BlockedSemiOpen3 diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 11031 W: 2344 L: 2421 D: 6266
sprt @ 10+0.1 th 1 Try reducing by a quarter, rather than a half (as in the previous test on this branch).
18-09-03 31m BlockedSemiOpen2 diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 36968 W: 7879 L: 7831 D: 21258
sprt @ 10+0.1 th 1 Inspired by @Rocky640's tests. If an otherwise open file is blocked by a doubly-pawn-protected enemy (which is neither R nor Q), then for the purposes of RookOnFile, the file must be merely semi-open.
18-09-03 31m BlockedSemiOpen3 diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 12434 W: 2584 L: 2655 D: 7195
sprt @ 10+0.1 th 1 Inspired by @Rocky640's tests. If a semi-open file is blocked by a doubly-pawn-protected enemy (which is neither R nor Q), halve the RookOnFile bonus.
18-09-03 31m BlockedSemiOpen^ diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 29950 W: 6396 L: 6382 D: 17172
sprt @ 10+0.1 th 1 Modify @Rocky640's test: require that the outpost be attacked from in front by an enemy rook. (Please see the comments on his test.)
18-09-03 31m BlockedSemiOpen diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 8035 W: 1655 L: 1747 D: 4633
sprt @ 10+0.1 th 1 Modify @Rocky640's test: require that the file be semi-open for the enemy, directly mirroring the original RookOnFile. (Please see the comments on his test.)
18-09-03 31m IsolatedKProximity5^^ diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 29595 W: 6376 L: 6363 D: 16856
sprt @ 10+0.1 th 1 I thought I had run out of ideas, but I have one more for now. Maybe the bugged patch that went green at STC wasn't due to chance--maybe I do actually need to consider the entire board, not the one passer. Check all passed pawns to see if we have any isolated ones. If not, reduce the king proximity by a quarter.
18-09-03 31m IsolatedKProximity5^ diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 19524 W: 4143 L: 4179 D: 11202
sprt @ 10+0.1 th 1 Reduce the penalty by half.
18-09-03 31m IsolatedKProximity5 diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 19405 W: 4130 L: 4167 D: 11108
sprt @ 10+0.1 th 1 Eliminate the penalty entirely.
18-09-03 31m kingdanger_tweak3 diff
LLR: -2.95 (-2.94,2.94) [0.00,4.00]
Total: 11468 W: 1806 L: 1923 D: 7739
sprt @ 60+0.6 th 1 LTC for @Vizvezdenec. take 3 - middle ground
18-09-02 31m TrappedRook_4 diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 4391 W: 885 L: 994 D: 2512
sprt @ 10+0.1 th 1 An idea based on promising results from @Vizvezdenec, who extended TrappedRook to mob = 4, and comments by @Rocky640, who noted that the resulting penalty was small. Try applying the larger mob = 3 TrappedRook penalty, even if mob = 4.
18-09-02 31m tweak_KingProximity diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 22040 W: 4689 L: 4713 D: 12638
sprt @ 10+0.1 th 1 Removing a quarter of the penalty about half the time was good at STC...try merely reducing the penalty by one eighth in all cases. (Could this be tested as [0, 4]? I'm not sure...)
18-09-02 31m IsolatedKProximity4 diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 26901 W: 5786 L: 5786 D: 15329
sprt @ 10+0.1 th 1 Because this version is so narrow, double the effect size (halve the bonus, rather than reducing by a quarter) to compensate.
18-09-02 31m IsolatedKProximity4^ diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 13856 W: 2929 L: 2993 D: 7934
sprt @ 10+0.1 th 1 A very narrow version. A passer is only connected if other friendly pawns defend it or its blockSq.
18-09-02 31m IsolatedKProximity4 diff
LLR: -2.94 (-2.94,2.94) [0.00,5.00]
Total: 12628 W: 2674 L: 2743 D: 7211
sprt @ 10+0.1 th 1 At the time of writing, larger effect size seems better. Go further: in this narrow definition of connected passed pawns, wipe out the penalty entirely. (The commit message accidentally says wipe out "bonus" instead; please disregard.)
18-09-01 31m IsolatedKProximity2b diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 21811 W: 3441 L: 3479 D: 14891
sprt @ 60+0.6 th 1 Although the performance of a 48K yellow (estimated +1.25 Elo) is generally a bit less than I would submit to speculative LTC, passed pawns are much more important in the endgame, so I am cautiously optimistic. Merge the new master and test at LTC with low throughput (166).
18-09-02 31m IsolatedKProximity2b diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 19619 W: 3160 L: 3206 D: 13253
sprt @ 60+0.6 th 1 LTC: Try a suggestion by @Rocky640, changing the "connected" condition. (I'll leave my speculative LTC also running with its reduced throughput, but it can be stopped if this passes.)
18-09-02 31m IsolatedKProximity2b diff
LLR: 2.96 (-2.94,2.94) [0.00,5.00]
Total: 23826 W: 5208 L: 4962 D: 13656
sprt @ 10+0.1 th 1 Try a suggestion by @Rocky640, changing the "connected" condition.
18-09-01 31m IsolatedKProximity2^ diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 23954 W: 5166 L: 5180 D: 13608
sprt @ 10+0.1 th 1 Reducing by half was bad, but a quarter was close (+1.25 Elo). Try two tweaks on each side of one quarter. Here, one eighth.
18-09-01 31m IsolatedKProximity diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 10447 W: 2155 L: 2235 D: 6057
sprt @ 10+0.1 th 1 Increase by a quarter if isolated, but don't decrease if not.
18-09-01 31m IsolatedKProximity3 diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 8880 W: 1852 L: 1940 D: 5088
sprt @ 10+0.1 th 1 Reduce the king proximity penalty by a quarter if connected, as in the best take so far. However, also reduce the king proximity bonus under the same conditions. (If any of my tests happen to pass in the next 6-7 hours, I may not be available to start my own LTC. I would be grateful to anyone who does.)
18-09-01 31m IsolatedKProximity2 diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 7964 W: 1626 L: 1718 D: 4620
sprt @ 10+0.1 th 1 Reducing by half was bad, but a quarter was close (+1.25 Elo). Try two tweaks on each side of one quarter. Here, three eighths.
18-09-01 31m IsolatedKProximity2 diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 48478 W: 10365 L: 10262 D: 27851
sprt @ 10+0.1 th 1 Half effect (i.e., reduce penalty by a quarter if the passer is connected).
18-09-01 31m IsolatedKProximity2 diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 11170 W: 2320 L: 2397 D: 6453
sprt @ 10+0.1 th 1 Apply only half the king proximity penalty if the passed pawn is connected (there is a friendly pawn one square away, diagonally or along the rank).
18-09-01 31m IsolatedKProximity diff
LLR: -2.97 (-2.94,2.94) [0.00,5.00]
Total: 8081 W: 1686 L: 1778 D: 4617
sprt @ 10+0.1 th 1 Half effect.
18-09-01 31m IsolatedKProximity diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 9161 W: 1912 L: 1998 D: 5251
sprt @ 10+0.1 th 1 My first attempt based on a comment by Bryan on the suggestions thread. Modulate the king-proximity penalty (for passers not yet on the 7th rank) based on whether the passer is isolated--king proximity is more important if it is. +50% effect size if isolated, -50% otherwise.
18-08-31 31m SpaceDivisor diff
LLR: -2.96 (-2.94,2.94) [0.00,4.00]
Total: 34088 W: 7228 L: 7260 D: 19600
sprt @ 10+0.1 th 1 @Rocky640's test was nearly an Elo-gainer, and reduced the size of "bonus" by roughly 20% by restricting the bitboards used to compute it. Perhaps the potential Elo gain comes merely from decreasing the bonus. Increase divisor from 16 to 20.
18-08-31 31m SpaceRankAndPawn diff
LLR: -2.95 (-2.94,2.94) [0.00,4.00]
Total: 17767 W: 3759 L: 3853 D: 10155
sprt @ 10+0.1 th 1 In consultation with Rocky, try an in-between value of 14.
18-08-31 31m SpaceRankAndPawn diff
LLR: -2.94 (-2.94,2.94) [0.00,4.00]
Total: 15416 W: 3279 L: 3382 D: 8755
sprt @ 10+0.1 th 1 My take on @Rocky640's idea. His changes appear to reduce the average value of bonus (across the limited set of bench positions) by about 20%, so I decrease the divisor used to construct the score in order to compensate.
18-08-31 31m combo_pss_SRAP diff
LLR: -2.95 (-2.94,2.94) [0.00,4.00]
Total: 9806 W: 1999 L: 2124 D: 5683
sprt @ 10+0.1 th 1 Two recent tweaks by @Rocky640 and @jerrydonaldwatson performed quite well at STC but gave long yellow runs at LTC. The obvious step is to try to combine them. (Note: Though I wrote neither of these patches, I am submitting this test because the framework is completely empty and SF's division in TCEC is quickly approaching, and I want to make sure that if these tweaks are an Elo gain, we have time to commit them before it starts. I hope this is OK, and please correct me if it is not.)
18-08-31 31m pawndanger diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 25640 W: 5434 L: 5441 D: 14765
sprt @ 10+0.1 th 1 Restore effect size 22, and try adjacent_files_bb rather than kingFlank--perhaps this is more precise.
18-08-31 31m pawndanger^ diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 19410 W: 4128 L: 4165 D: 11117
sprt @ 10+0.1 th 1 Take the best version so far, and experiment with effect size. Double (44).
18-08-31 31m pawndanger^^ diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 14805 W: 3074 L: 3133 D: 8598
sprt @ 10+0.1 th 1 Take the best version so far, and experiment with effect size. Half (11).
18-08-30 31m pawndanger diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 27181 W: 5821 L: 5820 D: 15540
sprt @ 10+0.1 th 1 Scale linearly with the number of enemy majors in our kingFlank.
18-08-30 31m pawndanger^^^ diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 13526 W: 2805 L: 2870 D: 7851
sprt @ 10+0.1 th 1 Inspired by and derived from @protonspring's work. The king danger from doubled pawns on the king file comes from two factors: there is likely a nearby semi-open file, and the pawns themselves may be useful to break open the defense. The former is only useful to the enemy if they have majors. First, try moving this penalty to kingDanger and applying this restriction.
18-08-30 31m pawndanger^^ diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 11767 W: 2499 L: 2573 D: 6695
sprt @ 10+0.1 th 1 An even narrower version. Only penalize if the enemy has a major in our kingFlank.