Stockfish Testing Queue

Pending - 0 tests 0.0 hrs

None

Active - 1 tests

19-09-17 sg tune_complex_eval diff
17856/100000 iterations
37210/200000 games played
200000 @ 20+0.2 th 1 I checked my last tuned values but the values will be for the final values scaled down to zero for all positions on bench. So i experimented a little bit and come up with no scaling between the layers and only the final will be scaled. There the complex_eval values seems reasonable. So redo tuning (C=5, Scale=128)

Finished - 4995 tests

19-09-17 sg complex_eval diff
ELO: -33.93 +-6.0 (95%) LOS: 0.0%
Total: 5609 W: 970 L: 1516 D: 3123
20000 @ 10+0.1 th 1 I checked my last tuned values but the values will be for the final values scaled down to zero for all positions on bench. So i experimented a little bit and come up with no scaling between the layers and only the final will be scaled. There the complex_eval values seems reasonable. The next tuning i will try with this version. But just for information measure for this last tuned values after 200K. C=5, Scale=128 using the fixed version the progress to base version with weights 0.
19-09-16 sg tune_complex_eval diff
95775/100000 iterations
200001/200000 games played
200000 @ 20+0.2 th 1 Setup the other tuning wrong. Also it seems my scaling is now to strong so doesn't do it for the final value so only clamp it. Also use short names. C=5, Scale=128.
19-09-16 sg scale_factor2 diff
LLR: -2.95 (-2.94,2.94) [0.00,4.00]
Total: 24952 W: 5432 L: 5498 D: 14022
sprt @ 10+0.1 th 1 ocb-scaling: weight=6 for passed pawn. Decrease base by one for compensation.
19-09-16 sg scale_factor2^ diff
LLR: -2.95 (-2.94,2.94) [0.00,4.00]
Total: 30398 W: 6661 L: 6705 D: 17032
sprt @ 10+0.1 th 1 ocb-scaling: weight=5 for passed pawn but decrease base by two for compensation.
19-09-16 sg scale_factor2 diff
LLR: -2.96 (-2.94,2.94) [0.00,4.00]
Total: 22099 W: 4717 L: 4795 D: 12587
sprt @ 10+0.1 th 1 ocb-scaling: weight=5 for passed pawn.
19-09-16 sg scale_factor2 diff
LLR: -2.96 (-2.94,2.94) [0.00,4.00]
Total: 58403 W: 12773 L: 12708 D: 32922
sprt @ 10+0.1 th 1 ocb-scaling: more weight for passed pawn but decrease base for compensation.
19-09-16 sg scale_factor2 diff
LLR: -2.95 (-2.94,2.94) [0.00,4.00]
Total: 20382 W: 4433 L: 4517 D: 11432
sprt @ 10+0.1 th 1 ocb-scaling: even more weight=6 for passed pawn but decrease base by two for compensation.
19-09-16 sg scale_factor2^ diff
LLR: -2.95 (-2.94,2.94) [0.00,4.00]
Total: 18049 W: 3913 L: 4006 D: 10130
sprt @ 10+0.1 th 1 ocb-scaling: less weight for passed pawn but increase base for compensation.
19-09-16 sg scale_factor2 diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 8957 W: 1893 L: 2013 D: 5051
sprt @ 10+0.1 th 1 Add passed pawns to strong side pawns but decrease base by 7 as compensation.
19-09-16 sg scale_factor2 diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 8413 W: 1834 L: 1957 D: 4622
sprt @ 10+0.1 th 1 Add to scale factor 4 times passed pawns but decrease base by 4 as compensation.
19-09-16 sg scale_factor2 diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 15978 W: 3479 L: 3565 D: 8934
sprt @ 10+0.1 th 1 Reduce scale factor by 8 if no passed pawns.
19-09-16 sg complex_eval diff
ELO: -30.31 +-3.2 (95%) LOS: 0.0%
Total: 19916 W: 3606 L: 5339 D: 10971
20000 @ 10+0.1 th 1 Measure progress after 200K games tuning against base (NN with weights 0).
19-09-16 sg tune_complex_eval diff
6684/100000 iterations
14423/200000 games played
200000 @ 10+0.1 th 1 Stopped last tuning. After checking i had found that for all positions the same end value is used. Fix scaling and clamp only the end value. Tune again with C=5, Scale=128
19-09-16 sg tune_complex_eval diff
33041/100000 iterations
68678/200000 games played
200000 @ 20+0.2 th 1 Try now even lower C=2 starting from 0 weights.
19-09-15 sg tune_complex_eval diff
96047/100000 iterations
200000/200000 games played
200000 @ 20+0.2 th 1 The value of tuning doesn't move. So try again C=5 but use Scale=1.
19-09-15 sg tweak_contempt diff
LLR: -2.95 (-2.94,2.94) [-3.00,1.00]
Total: 130138 W: 28263 L: 28683 D: 73192
sprt @ 10+0.1 th 1 The standard STC failed yellow. But perhaps this contempt change is good against weaker engines. But before i test this check if this is no regression against master. For static contempt >= 0 (< 0) allow only positive (negative) dynamic contempt.
19-09-15 sg tune_complex_eval diff
46413/100000 iterations
97292/200000 games played
200000 @ 20+0.2 th 1 So the first tuned values are really bad. I had done my first tuning with C=30 which seems to high. So try a second tuning with C=5 from zero values. Everything else keeps the same.
19-09-15 sg tune_complex_eval diff
94166/100000 iterations
196828/200000 games played
200000 @ 20+0.2 th 1 Clamp values to a range and use less games. Hope its works. Experiment: try something new and "train" a simple CNN as additional eval term with SPSA with 200K games (probably several runs are needed but lets see). There are 143 parameters (I hope this is not much). I measured a speed loose of around 3-4% in comparison to master. Comments and discussion are welcome.
19-09-15 sg complex_eval diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 1085 W: 152 L: 308 D: 625
sprt @ 10+0.1 th 1 Test tuned nn values after 182K games to get a first impression of progress but i think we need several million games.
19-09-15 sg scale_factor diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 51856 W: 11329 L: 11239 D: 29288
sprt @ 10+0.1 th 1 Reduce scale factor by one if material equal.
19-09-15 sg scale_factor diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 39751 W: 8618 L: 8588 D: 22545
sprt @ 10+0.1 th 1 Reduce scale factor by 4 if no passed pawns.
19-09-15 sg scale_factor diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 35652 W: 7845 L: 7833 D: 19974
sprt @ 10+0.1 th 1 Reduce scale factor by two if no passed pawns.
19-09-15 sg scale_factor diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 34612 W: 7494 L: 7489 D: 19629
sprt @ 10+0.1 th 1 Reduce scale factor by one if no passed pawns.
19-09-15 sg tweak_contempt diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 34786 W: 7684 L: 7676 D: 19426
sprt @ 10+0.1 th 1 For static contempt >= 0 (< 0) allow only positive (negative) dynamic contempt.
19-09-15 sg scale_factor diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 14963 W: 3261 L: 3352 D: 8350
sprt @ 10+0.1 th 1 Reduce scale factor by two if material equal.
19-09-15 sg tweak_contempt diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 22783 W: 5034 L: 5086 D: 12663
sprt @ 10+0.1 th 1 Add only positive dynamic contempt part.
19-09-15 sg tweak_contempt diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 19184 W: 4120 L: 4190 D: 10874
sprt @ 10+0.1 th 1 Add only positive dynamic contempt part but increase factor to 73 for compensation.
19-09-15 sg unwinnable diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 34010 W: 7473 L: 7470 D: 19067
sprt @ 10+0.1 th 1 33% more unwinnable bonus if material equal. Weight=8.
19-09-15 sg tweak_initiative diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 30497 W: 6568 L: 6583 D: 17346
sprt @ 10+0.1 th 1 Use endgame score without contempt only for bound.
19-09-15 sg unwinnable^ diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 15519 W: 3376 L: 3464 D: 8679
sprt @ 10+0.1 th 1 25% more unwinnable bonus if material equal. Weight=6.
19-09-15 sg tweak_initiative^ diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 19211 W: 4074 L: 4144 D: 10993
sprt @ 10+0.1 th 1 Use endgame score without contempt only for sign.
19-09-15 sg tweak_initiative diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 17830 W: 3814 L: 3891 D: 10125
sprt @ 10+0.1 th 1 Use midgame score withput contempt only for sign.
19-09-15 sg tweak_initiative^ diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 10859 W: 2305 L: 2416 D: 6138
sprt @ 10+0.1 th 1 Use midgame score withput contempt only for bound.
19-09-15 sg unwinnable diff
LLR: -2.94 (-2.94,2.94) [0.50,4.50]
Total: 41567 W: 9224 L: 9182 D: 23161
sprt @ 10+0.1 th 1 25% more unwinnable bonus if material equal. Weight=7.
19-09-15 sg unwinnable^ diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 30726 W: 6723 L: 6736 D: 17267
sprt @ 10+0.1 th 1 25% more unwinnable bonus if material equal. Weight=9.
19-09-15 sg tweak_initiative diff
LLR: -2.94 (-2.94,2.94) [0.50,4.50]
Total: 27902 W: 6026 L: 6053 D: 15823
sprt @ 10+0.1 th 1 Subtract contempt from midgame score in initiative.
19-09-15 sg unwinnable^ diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 52463 W: 11641 L: 11546 D: 29276
sprt @ 10+0.1 th 1 33% more unwinnable bonus if material equal. Weight=9.
19-09-15 sg tweak_initiative diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 38609 W: 8492 L: 8466 D: 21651
sprt @ 10+0.1 th 1 Use endgame sign for midgame initiative.
19-09-15 sg unwinnable diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 14924 W: 3207 L: 3298 D: 8419
sprt @ 10+0.1 th 1 33% more unwinnable bonus if material equal. Weight=12.
19-09-15 sg unwinnable diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 29703 W: 6475 L: 6493 D: 16735
sprt @ 10+0.1 th 1 Consider a position only unwinnable if material difference <= difference of bishop and knight.
19-09-15 sg unwinnable^ diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 17164 W: 3713 L: 3793 D: 9658
sprt @ 10+0.1 th 1 50% more unwinnable bonus if material equal. Weight=18.
19-09-15 sg unwinnable diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 16864 W: 3634 L: 3715 D: 9515
sprt @ 10+0.1 th 1 50% more unwinnable bonus if material equal. Weight=12.
19-09-15 sg unwinnable diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 11761 W: 2469 L: 2575 D: 6717
sprt @ 10+0.1 th 1 Consider a position only unwinnable if material is equal.
19-09-15 sg tune_complex_eval diff
143/500000 iterations
399/1000000 games played
1000000 @ 20+0.2 th 1 Sorry forget to set scale to 128 and fix also a bug there. Experiment: try something new and "train" a simple CNN as additional eval term with SPSA with 1M games (probably several runs are needed but lets see). There are 143 parameters (I hope this is not much). I measured a speed loose of around 3-4% in comparison to master. Comments and discussion are welcome.
19-09-14 sg tune_complex_eval diff
1904/500000 iterations
4163/1000000 games played
1000000 @ 20+0.2 th 1 Experiment: try something new and "train" a simple CNN as additional eval term with SPSA with 1M games (probably several runs are needed but lets see). There are 143 parameters (I hope this is not much). I measured a speed loose of around 3-4% in comparison to master. Comments and discussion are welcome.
19-09-14 sg initiative_flanks3 diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 8492 W: 1778 L: 1900 D: 4814
sprt @ 10+0.1 th 1 Give 20% more pawns on both flanks bonus if pawn outside of center files exists. Weight=3. Test against Viz's passed patch.
19-09-14 sg initiative_flanks diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 84999 W: 18755 L: 18501 D: 47743
sprt @ 10+0.1 th 1 Changes in pawnsOnBothFlanks affect also almostUnwinnable. So try my 190K yellow condition only for the pawnsOnBothFlanks part.
19-09-14 sg initiative_flanks diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 64316 W: 14237 L: 14084 D: 35995
sprt @ 10+0.1 th 1 Give 50% more pawns on both flanks bonus if pawn outside of center files exists. Weight=6.
19-09-14 sg initiative_unwinnable diff
LLR: -2.95 (-2.94,2.94) [0.50,4.50]
Total: 30792 W: 6767 L: 6779 D: 17246
sprt @ 10+0.1 th 1 Changes in pawnsOnBothFlanks affect also almostUnwinnable. So try my 190K yellow condition only for almostUnwinnable.
19-09-14 sg initiative_flanks diff
LLR: -2.96 (-2.94,2.94) [0.50,4.50]
Total: 23825 W: 5220 L: 5267 D: 13338
sprt @ 10+0.1 th 1 Give 50% more pawns on both flanks bonus if pawn outside of center files exists. Weight=9.