Stockfish Testing Queue

Finished - 40699 tests

15-02-24 zar doubled_tuning diff
8722/30000 iterations
17949/60000 games played
60000 @ 40+0.05 th 1 SPSA tuning for doubled pawns. And try tuning with contempt=50, priority: -2
15-02-25 jos threatened_byPawn diff
ELO: -0.60 +-2.2 (95%) LOS: 29.9%
Total: 40000 W: 8490 L: 8559 D: 22951
40000 @ 9+0.05 th 1 Check new values for Threat and ThreatenedByPawn arrays.
15-02-25 sni knight diff
LLR: 2.97 (-2.94,2.94) [-1.50,4.50]
Total: 27744 W: 5563 L: 5380 D: 16801
sprt @ 15+0.05 th 1 Try tuned values for knight threatened by pawn
15-02-25 sni knight diff
LLR: 2.96 (-2.94,2.94) [0.00,4.00]
Total: 87571 W: 14747 L: 14314 D: 58510
sprt @ 60+0.05 th 1 LTC: Try tuned values for knight threatened by pawn
15-02-26 Fis phaseDivorce diff
LLR: -2.97 (-2.94,2.94) [-1.50,4.50]
Total: 16364 W: 3341 L: 3403 D: 9620
sprt @ 15+0.05 th 1 Separate a pieces material value from it's phase value. Results after tuning.
15-02-26 Roc PawnsKnightsAndBishop diff
LLR: -2.96 (-2.94,2.94) [-1.50,4.50]
Total: 13101 W: 2582 L: 2653 D: 7866
sprt @ 15+0.05 th 1 First attempt on https://groups.google.com/forum/?fromgroups=#!topic/fishcooking/jHLBoTEOAzg
15-02-26 sg long_chain_with_apex diff
LLR: -2.95 (-2.94,2.94) [-1.50,4.50]
Total: 7542 W: 1473 L: 1559 D: 4510
sprt @ 15+0.05 th 1 Add bonus for inner pawns of a long chain. Based on an idea of Lyudmil Tsvetkov.
15-02-26 Roc PawnsKnightsAndBishop diff
LLR: -2.96 (-2.94,2.94) [-1.50,4.50]
Total: 28857 W: 5797 L: 5824 D: 17236
sprt @ 15+0.05 th 1 Take 2. Fixed a bug, and also use a bit more general "blocking" definition.
15-02-26 sg long_chain_with_apex diff
LLR: -2.96 (-2.94,2.94) [-1.50,4.50]
Total: 67777 W: 13347 L: 13269 D: 41161
sprt @ 15+0.05 th 1 Add bonus for inner pawns of a long chain. Based on an idea of Lyudmil Tsvetkov. (Take 2)
15-02-26 Roc OppCastle diff
LLR: -2.94 (-2.94,2.94) [-1.50,4.50]
Total: 10671 W: 2054 L: 2131 D: 6486
sprt @ 15+0.05 th 1 2 Weight adjustments when there is opposite castle. Based on quick local SPSA tuning with the salc500.epd suite.
15-02-26 sni definition_of_isolation diff
LLR: -2.96 (-2.94,2.94) [-1.50,4.50]
Total: 3696 W: 727 L: 825 D: 2144
sprt @ 15+0.05 th 1 Change definition of isolated pawns: we now consider to be isolated any pawn which has no friendly pawn in the three squares on its left or right sides. Divide the Isolated[] penalty by two as there are many more isolated pawns by this definition.
15-02-27 vin blocked_centre_spsa diff
32505/25000 iterations
50000/50000 games played
50000 @ 15+0.05 th 1 Previous tuning produced little movement so hints there is no dependency there. Try the other way around.
15-02-27 vin doubled_passed_pawns diff
LLR: 2.95 (-2.94,2.94) [-1.50,4.50]
Total: 7434 W: 1575 L: 1447 D: 4412
sprt @ 15+0.05 th 1 Try Mindbreaker's suggestion of a penalty for a doubled passed pawn. The front pawn is scored as passed, but with a small penalty because it will be harder to support.
15-02-27 mco eval1 diff
LLR: -2.95 (-2.94,2.94) [-3.00,1.00]
Total: 94440 W: 18608 L: 18941 D: 56891
sprt @ 15+0.05 th 1 Retire KingOnMany
15-02-27 mco eval2 diff
LLR: -2.95 (-2.94,2.94) [-3.00,1.00]
Total: 10185 W: 1953 L: 2132 D: 6100
sprt @ 15+0.05 th 1 Retire unstoppable pawns
15-02-27 sg long_chain diff
LLR: -2.97 (-2.94,2.94) [-1.50,4.50]
Total: 27535 W: 5377 L: 5409 D: 16749
sprt @ 15+0.05 th 1 Add bonus for inner pawns of a long chain. Based on an idea of Lyudmil Tsvetkov. Include new master and normalize long chain stuff like marco did with apex. This changes the bench because the bonus factor is now applied in another order.
15-02-28 jos fianchetto diff
LLR: -2.96 (-2.94,2.94) [-1.50,4.50]
Total: 10899 W: 2115 L: 2192 D: 6592
sprt @ 15+0.05 th 1 Retest this old idea, but only for a kingside fianchettoed bishop.
15-02-28 mco tt_iteration diff
LLR: -2.96 (-2.94,2.94) [-1.50,4.50]
Total: 55631 W: 10128 L: 10090 D: 35413
sprt @ 30+0.05 th 1 Increment TT generation upon reaching hashfull (idea from http://www.talkchess.com/forum/viewtopic.php?t=55501). Test with small hash and a bit longer TC to trigger the condition.
15-02-28 jos fianchetto diff
LLR: -2.97 (-2.94,2.94) [-1.50,4.50]
Total: 20851 W: 4087 L: 4137 D: 12627
sprt @ 15+0.05 th 1 Kingside fianchetto, take 2. Additionally check for a defended pawn on g3/g6.
15-02-28 vin doubled_passed_pawns diff
LLR: 2.95 (-2.94,2.94) [-1.50,4.50]
Total: 14555 W: 2929 L: 2783 D: 8843
sprt @ 15+0.05 th 1 Uh, actually use the correct reweighted values this time instead of greatly increasing the Eg penalty. Final try.
15-02-28 sni sorting_moves4 diff
LLR: -2.95 (-2.94,2.94) [-1.50,4.50]
Total: 7506 W: 1440 L: 1526 D: 4540
sprt @ 15+0.05 th 1 It is often good to capture pawns with pieces rather than with pawns to avoid deteriorating the pawn structure, especially in endgame. Since the see() call will delay the bad captures till the end of the move list anyway, it makes sense to try ordering captures by MVV/MVA instead of MVV/LVA.
15-02-28 sni sorting_moves5 diff
LLR: -2.97 (-2.94,2.94) [-1.50,4.50]
Total: 5643 W: 1104 L: 1196 D: 3343
sprt @ 15+0.05 th 1 Sorting captures in MVV/LVA/history order
15-02-28 zar doubled_tuning diff
33052/30000 iterations
63000/60000 games played
60000 @ 15+0.05 th 1 SPSA tuning for doubled pawns
15-02-28 hxi phase_min diff
LLR: -3.14 (-2.94,2.94) [-1.50,4.50]
Total: 6257 W: 1172 L: 1268 D: 3817
sprt @ 15+0.05 th 1 game phase based on side with less material
15-02-28 jos threatened_byPawn diff
ELO: 0.46 +-2.2 (95%) LOS: 65.8%
Total: 40000 W: 8478 L: 8425 D: 23097
40000 @ 9+0.05 th 1 Measure of fine-tuned values for Threat and ThreatenedByPawn arrays. (Final try.)
15-03-01 zar order diff
ELO: -1.97 +-2.1 (95%) LOS: 3.5%
Total: 41000 W: 8088 L: 8320 D: 24592
40000 @ 15+0.05 th 1 Add psqt weight to sort quites, check value (fixed)
15-03-01 Roc OppCastleV2 diff
LLR: -2.95 (-2.94,2.94) [-1.50,4.50]
Total: 12544 W: 2462 L: 2534 D: 7548
sprt @ 15+0.05 th 1 2 weights have been adjusted when kings are in opposite castle 3x2 corners. Take 2
15-03-01 vin doubled_passed_pawns diff
LLR: -3.94 (-2.94,2.94) [0.00,6.00]
Total: 30360 W: 5031 L: 5023 D: 20306
sprt @ 60+0.05 th 1 Re-test at LTC after correct version of patch passed at STC.
15-03-01 jki avail diff
ELO: -13.14 +-6.9 (95%) LOS: 0.0%
Total: 3200 W: 461 L: 582 D: 2157
20000 @ 15+0.05 th 16 Examine threads in random order when splitting
15-03-01 jos fianchetto diff
LLR: -3.47 (-2.94,2.94) [-1.50,4.50]
Total: 6443 W: 1242 L: 1350 D: 3851
sprt @ 15+0.05 th 1 Check for pawns on f2, g3 and h2, slightly increased bonus. Take 3.
15-03-01 vin time_predictor diff
24048/25000 iterations
46725/50000 games played
50000 @ 30+0.05 th 1 Try using the more consistent estimated node growth rather than BestMoveChanges as a measure of PV instability. See the old thread "Comparative analysis of "big think" positions - suggests new metric?" Difficult to work out the 'right' parameters so begin with a tuning run from a can't-be-awful estimate based on bench output. As this is a time management patch use slightly longer TC.
15-03-01 jki avail2 diff
ELO: 2.40 +-2.8 (95%) LOS: 95.2%
Total: 19000 W: 3166 L: 3035 D: 12799
20000 @ 15+0.05 th 16 Orthodox helpful master concept
15-03-01 jos bishop_pawns diff
LLR: -3.37 (-2.94,2.94) [-3.00,1.00]
Total: 155176 W: 30329 L: 30794 D: 94053
sprt @ 15+0.05 th 1 Remove the little helper "pawns_on_same_color_squares" and directly compute it when needed. Small simplification and speedup as well.
15-03-01 vin blocked_centre diff
LLR: -2.95 (-2.94,2.94) [-1.50,4.50]
Total: 3322 W: 627 L: 725 D: 1970
sprt @ 15+0.05 th 1 Put in the only value from the SPSA session that showed some meaningful change. Adjust stormdanger when centre is open.
15-03-01 jki nolocks diff
ELO: 6.64 +-2.8 (95%) LOS: 100.0%
Total: 20000 W: 3482 L: 3100 D: 13418
20000 @ 15+0.05 th 16 Retire global thread lock
15-03-02 vin time_predictor diff
7122/25000 iterations
14350/50000 games played
50000 @ 30+0.05 th 1 Second run with corrected larger c-value as suggested by Nicklas Persson. Otherwise it'll take forever to converge..
15-03-02 zar one_lever diff
LLR: -2.97 (-2.94,2.94) [-1.50,4.50]
Total: 35125 W: 6914 L: 6925 D: 21286
sprt @ 15+0.05 th 1 Don't bonus lever with two opponents.
15-03-02 vin time_predictor diff
25188/25000 iterations
50000/50000 games played
50000 @ 30+0.05 th 1 Restart SPSA run as one parameter was going to hit the lower limit. (Thanks Binky) Also sync with latest master.
15-03-02 vin structural_mobility diff
LLR: -3.31 (-2.94,2.94) [-1.50,4.50]
Total: 21025 W: 4118 L: 4180 D: 12727
sprt @ 15+0.05 th 1 Try bonus for pawns that are more mobile structurally (e.g. candidate passers) as opposed to currently mobile (the current safe pawn push bonus)
15-03-02 sg long_chain diff
LLR: 2.97 (-2.94,2.94) [-1.50,4.50]
Total: 8225 W: 1708 L: 1578 D: 4939
sprt @ 15+0.05 th 1 Add bonus for inner pawns of a long chain. (even lower bonus)
15-03-02 jos matimb diff
ELO: -3.46 +-2.1 (95%) LOS: 0.1%
Total: 44000 W: 9336 L: 9774 D: 24890
40000 @ 9+0.05 th 1 One last try to improve upon existing values.
15-03-02 lbr 54f8a9cb diff
ELO: -8.27 +-58.9 (95%) LOS: 39.1%
Total: 42 W: 6 L: 7 D: 29
20000 @ 15+0.05 th 1 SF 5: 4 vs. 2 threads
15-03-03 Fis TTflipEval diff
LLR: -2.95 (-2.94,2.94) [-1.50,4.50]
Total: 11453 W: 2226 L: 2301 D: 6926
sprt @ 15+0.05 th 1 Use the NULL move flip trick on TT evals also. 2MB hash
15-03-03 sg long_chain diff
LLR: -2.96 (-2.94,2.94) [0.00,6.00]
Total: 15665 W: 2575 L: 2603 D: 10487
sprt @ 60+0.05 th 1 LTC: Add bonus for inner pawns of a long chain. (even lower bonus)
15-03-03 Fis TTflipEval diff
LLR: -3.09 (-2.94,2.94) [-1.50,4.50]
Total: 10331 W: 1934 L: 2017 D: 6380
sprt @ 15+0.05 th 1 One more try also refreshing the flipped TT entries. 2MB
15-03-03 vin time_predictor diff
26652/25000 iterations
50000/50000 games played
50000 @ 30+0.05 th 1 One parameter has converged it seems, so one more run to stabilise the other. Priority -1 so as not to hold up all the eval/threading work.
15-03-03 sni king2 diff
LLR: -2.95 (-2.94,2.94) [-1.50,4.50]
Total: 8113 W: 1654 L: 1739 D: 4720
sprt @ 15+0.05 th 1 Give less importance to the pawn structure when the kings are separated
15-03-03 sni king2 diff
LLR: -2.94 (-2.94,2.94) [-1.50,4.50]
Total: 11459 W: 2310 L: 2385 D: 6764
sprt @ 15+0.05 th 1 Take 2. Less reduction with king separation.
15-03-03 sni king2 diff
LLR: -2.96 (-2.94,2.94) [-1.50,4.50]
Total: 12731 W: 2514 L: 2586 D: 7631
sprt @ 15+0.05 th 1 Take 3. Quadratic reduction.
15-03-04 Fis tt_iteration2 diff
LLR: -2.96 (-2.94,2.94) [-1.50,4.50]
Total: 2442 W: 433 L: 533 D: 1476
sprt @ 15+0.05 th 1 A modified version of this http://tests.stockfishchess.org/tests/view/54f1b4af0ebc594fbf9a9735 patch. Full credit to Marco and http://www.talkchess.com/forum/viewtopic.php?t=55501 2MB