Stockfish Testing Queue

Finished - 22425 tests

13-12-14 al pval_tunA diff
LLR: -2.96 (-2.94,2.94) [-1.50,3.00]
Total: 6087 W: 1166 L: 1301 D: 3620
sprt @ 15+0.05 th 1 from SPSA run: selectively decrease some Mg piece values (by 0, 2%, or 4%) and increase some Eg values (by 0, 1%, or 2%) ScaleMgValues=-200 bp, ScaleEgValues=100 bp
13-12-14 mb less_king_safety diff
LLR: -2.95 (-2.94,2.94) [-1.50,4.50]
Total: 13598 W: 2728 L: 2797 D: 8073
sprt @ 15+0.05 th 1 Local test at very short tc suggests that lower king safety might be better.
13-12-14 Fi ttSaveNoRead diff
LLR: -2.95 (-2.94,2.94) [-1.50,4.50]
Total: 23861 W: 4686 L: 4727 D: 14448
sprt @ 15+0.05 th 1 Avoid any reading from a ttentry while saving it. This should result in a cache optimization.
13-12-14 vi reduce_dithering diff
LLR: -2.96 (-2.94,2.94) [-1.50,4.50]
Total: 7384 W: 1474 L: 1561 D: 4349
sprt @ 15+0.05 th 1 Trying to reduce the occurrence of very long thinks in an unstable PV situation. The value of 0.7is a guess - could be tuned later. Local 1000-game run at 15+0.05 produced gain of 2 Elo (obviously with big error bars but implies no massive regression at least)
11-12-14 Fi TTstreamline diff
ELO: 1.60 +-2.0 (95%) LOS: 94.4%
Total: 40000 W: 6823 L: 6639 D: 26538
40000 @ 60+0.05 th 1 Measure elo @ LTC w/ 16MB
12-12-14 al pval_spsa diff
9876/10000 iterations
20000/20000 games played
20000 @ 15+0.15 th 1 SPSA tuning of midgame and endgame piece values: Trying to use the results (dynamic) of Gary's SPSA tests but with only two tuning parameters. Anchor values: Mg Bishop and Eg Pawn
12-12-14 sn no_shuffling2 diff
ELO: -1.53 +-3.9 (95%) LOS: 22.4%
Total: 10000 W: 1655 L: 1699 D: 6646
10000 @ 60+0.05 th 1 Estimate the Elo cost of avoiding piece shuffling. Previous try of this idea was -0.56 +-3.9 Elo.
13-12-14 lb storm diff
LLR: -2.96 (-2.94,2.94) [-3.50,0.50]
Total: 41847 W: 8360 L: 8638 D: 24849
sprt @ 15+0.05 th 1 don't special case edge files: take 2.
13-12-14 lb edge diff
LLR: -2.96 (-2.94,2.94) [-3.50,0.50]
Total: 19326 W: 3852 L: 4068 D: 11406
sprt @ 15+0.05 th 1 don't special case edge files: take 1
12-12-14 jo matimb diff
ELO: -1.81 +-2.3 (95%) LOS: 5.9%
Total: 40000 W: 8799 L: 9007 D: 22194
40000 @ 7+0.05 th 1 Test some manually tuned values.
12-12-14 Fi insta_move diff
ELO: 1.00 +-2.5 (95%) LOS: 78.0%
Total: 30000 W: 6233 L: 6147 D: 17620
30000 @ 15+0.05 th 1 Measure a more conservative version w/ proper timing adjustment. Pri -3
12-12-14 jo stormdanger diff
LLR: -2.95 (-2.94,2.94) [-1.50,4.50]
Total: 55175 W: 11231 L: 11184 D: 32760
sprt @ 15+0.05 th 1 Reduced bonus for B/G files, too.
12-12-14 My Kpawn diff
LLR: -2.96 (-2.94,2.94) [-1.50,4.50]
Total: 31242 W: 6351 L: 6371 D: 18520
sprt @ 15+0.05 th 1 Endgame bonus for King attacking enemy Pawns.
12-12-14 sg stormdanger diff
LLR: -2.95 (-2.94,2.94) [-1.50,4.50]
Total: 10472 W: 2075 L: 2153 D: 6244
sprt @ 15+0.05 th 1 25% more Stormdanger bonus for blocked pawn on g6/b6. Try some opposite because this test fails badly: http://tests.stockfishchess.org/tests/view/5489fa5b0ebc591511eb6ef2
10-12-14 lb history diff
ELO: 0.11 +-1.9 (95%) LOS: 54.7%
Total: 40000 W: 6118 L: 6105 D: 27777
40000 @ 120+0.05 th 1 half history max. measure at super LTC to satisfy the hand wavers.
12-12-14 Fi insta_move diff
LLR: -2.96 (-2.94,2.94) [0.00,6.00]
Total: 26775 W: 4608 L: 4580 D: 17587
sprt @ 60+0.05 th 1 This is the tuning suggested by lp that scored best +2.43 elo on 30K games.
11-12-14 sg history_bonus diff
ELO: -3.23 +-3.1 (95%) LOS: 1.9%
Total: 20000 W: 3937 L: 4123 D: 11940
20000 @ 15+0.05 th 1 use half history bonus, so updates can occur up to depth 31 (instead depth 22). I am expect no significant difference on STC, but do a quick measurement as baseline for later attempts (Take 1)
11-12-14 sg stormdanger diff
LLR: -2.96 (-2.94,2.94) [-1.50,4.50]
Total: 45064 W: 9115 L: 9097 D: 26852
sprt @ 15+0.05 th 1 Double Stormdanger bonus for blocked pawn on f5/c5
11-12-14 sg stormdanger diff
LLR: -2.95 (-2.94,2.94) [-1.50,4.50]
Total: 40589 W: 8370 L: 8363 D: 23856
sprt @ 15+0.05 th 1 25% more Stormdanger bonus for blocked f6/c6 pawn (After Garys compile fix)
12-12-14 Fi insta_move diff
LLR: 2.97 (-2.94,2.94) [-1.50,4.50]
Total: 7754 W: 1659 L: 1529 D: 4566
sprt @ 15+0.05 th 1 This is the tuning suggested by lp that scored best +2.43 elo on 30K games.
11-12-14 Fi insta_move diff
ELO: 1.67 +-2.5 (95%) LOS: 90.3%
Total: 30000 W: 6248 L: 6104 D: 17648
30000 @ 15+0.05 th 1 lp idea to bump time factor to .75 to compensate for average time looks over +2 elo. So bump more in the same direction to .78 and see if we improve further. I will resubmit only the stronger to sprt. Pri -1
11-12-14 sg stormdanger diff
LLR: -2.95 (-2.94,2.94) [-1.50,4.50]
Total: 5899 W: 1168 L: 1259 D: 3472
sprt @ 15+0.05 th 1 25% less Stormdanger bonus for blocked pawn on b or g file (After Garys complie fix)
11-12-14 Fi TTstreamline diff
ELO: 2.41 +-2.2 (95%) LOS: 98.6%
Total: 40000 W: 8175 L: 7897 D: 23928
40000 @ 15+0.05 th 1 Measure elo @ STC w/ 4MB
10-12-14 Fi insta_move diff
ELO: 2.43 +-2.5 (95%) LOS: 97.1%
Total: 30000 W: 6293 L: 6083 D: 17624
30000 @ 15+0.05 th 1 Measure a tuning suggested by lp. Pri -1
11-12-14 mc skipEarlyAfterLMR^ diff
LLR: -2.97 (-2.94,2.94) [0.00,6.00]
Total: 45362 W: 7602 L: 7488 D: 30272
sprt @ 60+0.05 th 1 LTC: Skip early pruning at research: take 1
09-12-14 gl master diff
ELO: 41.42 +-1.9 (95%) LOS: 100.0%
Total: 40000 W: 8854 L: 4108 D: 27038
40000 @ 60+0.05 th 1 Regression test against sf5 using 8moves_v3, previous was 39 +- 1.9
10-12-14 mc skipEarlyAfterLMR diff
LLR: -2.95 (-2.94,2.94) [-1.50,4.50]
Total: 5583 W: 1059 L: 1150 D: 3374
sprt @ 15+0.05 th 1 Skip early pruning at research: take 2
10-12-14 mc skipEarlyAfterLMR^ diff
LLR: 2.96 (-2.94,2.94) [-1.50,4.50]
Total: 35805 W: 7263 L: 7057 D: 21485
sprt @ 15+0.05 th 1 Skip early pruning at research: take 1
10-12-14 Fi insta_move diff
ELO: 0.16 +-2.5 (95%) LOS: 55.1%
Total: 30000 W: 6068 L: 6054 D: 17878
30000 @ 15+0.05 th 1 Insta move more conservatively. Leave slow mover at 80 like master. See what elo difference this tuning gives.
10-12-14 My pinned_mob diff
LLR: -2.95 (-2.94,2.94) [-1.50,4.50]
Total: 11139 W: 2226 L: 2302 D: 6611
sprt @ 15+0.05 th 1 Take 3.
09-12-14 Fi insta_move diff
LLR: -2.95 (-2.94,2.94) [0.00,6.00]
Total: 16588 W: 2822 L: 2844 D: 10922
sprt @ 60+0.05 th 1 Try some instant moves. Inspired by watching Gull play on TCEC.
09-12-14 My pinned_mob diff
LLR: -2.95 (-2.94,2.94) [-1.50,4.50]
Total: 92311 W: 18846 L: 18695 D: 54770
sprt @ 15+0.05 th 1 The idea is that when a piece is pinned it's mobility should be devalued.
09-12-14 Ro WeakDefenders diff
LLR: -3.34 (-2.94,2.94) [-1.50,4.50]
Total: 17975 W: 3611 L: 3682 D: 10682
sprt @ 15+0.05 th 1 Undermined pawns makes poor piece defenders. Take this into accounts when computing Threats, Take 1.
09-12-14 Fi insta_move diff
LLR: 2.95 (-2.94,2.94) [-1.50,4.50]
Total: 27719 W: 5749 L: 5564 D: 16406
sprt @ 15+0.05 th 1 Try some instant moves. Inspired by watching Gull play on TCEC.
09-12-14 My pinned_mob2 diff
LLR: -2.95 (-2.94,2.94) [-1.50,4.50]
Total: 18342 W: 3671 L: 3727 D: 10944
sprt @ 15+0.05 th 1 Take 2.
08-12-14 mc skipEarlyPruning diff
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 72007 W: 12133 L: 12095 D: 47779
sprt @ 60+0.05 th 1 LTC: Better clarify when skipping early pruning and go directly to moves loop
06-12-14 gl spnode_extra diff
ELO: -1.63 +-2.8 (95%) LOS: 13.0%
Total: 20000 W: 3445 L: 3539 D: 13016
20000 @ 15+0.05 th 7 Search all captures first in split node children
08-12-14 mc skipEarlyPruning diff
LLR: 2.96 (-2.94,2.94) [-3.00,1.00]
Total: 25282 W: 5136 L: 5022 D: 15124
sprt @ 15+0.05 th 1 Better clarify when skipping early pruning and go directly to moves loop
08-12-14 My defQ_tuned diff
LLR: -2.95 (-2.94,2.94) [0.00,4.00]
Total: 16790 W: 3347 L: 3445 D: 9998
sprt @ 15+0.05 th 1 SPSA values for including Queen attacks on defended pieces.
06-12-14 lb hmax diff
ELO: -5.11 +-2.0 (95%) LOS: 0.0%
Total: 40000 W: 6708 L: 7296 D: 25996
40000 @ 15+0.05 th 7 hmax=250*threads (bcos History is shared in SMP)
07-12-14 jo stormdanger diff
LLR: 2.95 (-2.94,2.94) [0.00,6.00]
Total: 33225 W: 5708 L: 5445 D: 22072
sprt @ 60+0.05 th 1 LTC: Halve StormDanger bonus for blocked pawn on file a or h. See current TCEC game Gull - Stockfish.
07-12-14 My defQ_tune diff
19466/20000 iterations
39900/40000 games played
40000 @ 15+0.05 th 1 Tune defended Queen attacks (excuse hackish setup) low pri.
07-12-14 Fi TTstreamline diff
LLR: 2.97 (-2.94,2.94) [0.00,6.00]
Total: 13021 W: 2238 L: 2073 D: 8710
sprt @ 60+0.05 th 1 Avoid searching the TT twice for the same key/position during both probe() AND store(). Just keep the pointer and remove code from tt.cpp. Maybe this could be tested as a simplification? but I expect some performance gain as well.
06-12-14 gl spnode_extra diff
ELO: -2.66 +-3.5 (95%) LOS: 6.9%
Total: 13076 W: 2230 L: 2330 D: 8516
20000 @ 15+0.05 th 7 Always do verification searches in split node children
07-12-14 jo stormdanger diff
LLR: 2.95 (-2.94,2.94) [-1.50,4.50]
Total: 3410 W: 758 L: 641 D: 2011
sprt @ 15+0.05 th 1 Halve StormDanger bonus for blocked pawn on file a or h. See current TCEC game Gull - Stockfish.
07-12-14 Fi TTstreamline diff
LLR: 2.96 (-2.94,2.94) [-1.50,4.50]
Total: 13620 W: 2810 L: 2665 D: 8145
sprt @ 15+0.05 th 1 Avoid searching the TT twice for the same key/position during both probe() AND store(). Just keep the pointer and remove code from tt.cpp. Maybe this could be tested as a simplification? but I expect some performance gain as well.
07-12-14 jo endgame_scaling diff
LLR: -2.97 (-2.94,2.94) [-1.50,4.50]
Total: 2937 W: 561 L: 661 D: 1715
sprt @ 15+0.05 th 1 Make ScaleFactor depending on the number of remaining pawns. I hope this is the more general approach than pawn_span == 0 or 1. If it passes we can try to raise the number of pawns to 5, 6 or even 7. Take 1.
05-12-14 rn simpleprng diff
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 73703 W: 12281 L: 12245 D: 49177
sprt @ 60+0.05 th 1 LTC: Take 2: Use a simpler PRNG and simplify how it is used. This time, the seed for generating the Zobrist keys was selected to provide good distribution in the key space.
06-12-14 jo spsa_piece_values diff
LLR: -2.96 (-2.94,2.94) [-1.00,3.00]
Total: 40470 W: 8022 L: 8108 D: 24340
sprt @ 15+0.05 th 1 Final try. SPSA tuning oscillated around these values, until suddenly mid- and endgame values drifted apart like in former runs.
07-12-14 My RKonQN_tuned diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 12675 W: 2498 L: 2568 D: 7609
sprt @ 15+0.05 th 1 Final take 'gut feeling' values with tougher bounds to minimise luck chance.