Stockfish Testing Queue

Pending - 0 tests 0.0 hrs

None

Active - 1 tests

26-03-17 sg update_stats4 diff
LLR: -0.38 (-2.94,2.94) [0.00,4.00]
Total: 64492 W: 11688 L: 11489 D: 41315
sprt @ 10+0.1 th 1 Test now 3/16 solution which is in the middle of best attempts

Finished - 937 tests

24-03-17 sg update_stats diff
LLR: -2.96 (-2.94,2.94) [0.00,4.00]
Total: 40998 W: 7216 L: 7235 D: 26547
sprt @ 10+0.1 th 1 Half for all history stats the resolution.
25-03-17 sg update_stats3 diff
LLR: -2.95 (-2.94,2.94) [0.00,4.00]
Total: 174899 W: 31910 L: 31469 D: 111520
sprt @ 10+0.1 th 1 Last test struggles, so try an eighth resolution.
24-03-17 sg update_stats2 diff
LLR: -2.95 (-2.94,2.94) [0.00,4.00]
Total: 56416 W: 10138 L: 10103 D: 36175
sprt @ 10+0.1 th 1 My first two tests indicate that lower resolution is better, so go further and try quarter resolution. Put other test on prio -1.
26-03-17 sg stats_bonus diff
LLR: -2.95 (-2.94,2.94) [0.00,4.00]
Total: 10021 W: 1755 L: 1877 D: 6389
sprt @ 10+0.1 th 1 My resolution series has the intention to lower the noise in computed stats (shows some very little gain so far). Now i try to go from the opposite site and filter the noise of the update values by decreasing the stats bonus by 1. For depth 1 (the noisiest one) we get then a bonus of zero, which avoids stats updates.
25-03-17 sg update_stats4 diff
LLR: -2.96 (-2.94,2.94) [0.00,4.00]
Total: 23034 W: 4166 L: 4244 D: 14624
sprt @ 10+0.1 th 1 Go now to the extreme 1/32 resolution (update weight 1).
25-03-17 sg update_stats4 diff
LLR: -2.96 (-2.94,2.94) [0.00,4.00]
Total: 25366 W: 4478 L: 4549 D: 16339
sprt @ 10+0.1 th 1 1/32 seems to low so try 1/16 resolution
24-03-17 sg update_stats2 diff
LLR: -2.95 (-2.94,2.94) [0.00,4.00]
Total: 18828 W: 3353 L: 3445 D: 12030
sprt @ 10+0.1 th 1 Double stats resolution.
20-03-17 sg initiative diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 63116 W: 11350 L: 11208 D: 40558
sprt @ 10+0.1 th 1 Cap pawns up to a maximum of 12.
21-03-17 sg initiative2 diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 20284 W: 3628 L: 3667 D: 12989
sprt @ 10+0.1 th 1 Increasing weight failed, so try decreasing weight to 11
21-03-17 sg initiative2 diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 14120 W: 2516 L: 2581 D: 9023
sprt @ 10+0.1 th 1 weight 13 with more compensation
21-03-17 sg initiative2 diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 24882 W: 4451 L: 4471 D: 15960
sprt @ 10+0.1 th 1 Weight 13 with compensation
21-03-17 sg initiative2 diff
LLR: -2.97 (-2.94,2.94) [0.00,5.00]
Total: 14201 W: 2449 L: 2515 D: 9237
sprt @ 10+0.1 th 1 My first test struggles. Increase weight to 13.
20-03-17 sg stat_bonus diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 11587 W: 2040 L: 2116 D: 7431
sprt @ 10+0.1 th 1 Inspired by VoyagerOne comment (my values are higher for low depths than master formula) i test a hybrid: For d < 5 take d*log(d*d) formula and else the old one.
19-03-17 sg stat_bonus diff
LLR: -2.97 (-2.94,2.94) [0.00,5.00]
Total: 27779 W: 3535 L: 3564 D: 20680
sprt @ 60+0.6 th 1 LTC: d*log(d*d)
19-03-17 sg stat_bonus diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 36710 W: 6575 L: 6545 D: 23590
sprt @ 10+0.1 th 1 Go further and try the form d*log(d*d*d)
19-03-17 sg stat_bonus diff
LLR: 2.96 (-2.94,2.94) [0.00,5.00]
Total: 18535 W: 3433 L: 3229 D: 11873
sprt @ 10+0.1 th 1 d*log(d*d)
19-03-17 sg stat_bonus diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 8003 W: 1378 L: 1469 D: 5156
sprt @ 10+0.1 th 1 Try slower growing formula for stat bonus: d*log(d) instead of d*d.
18-03-17 sg null_verify diff
LLR: -2.97 (-2.94,2.94) [0.00,5.00]
Total: 18259 W: 3180 L: 3229 D: 11850
sprt @ 10+0.1 th 1 Combine now with or
17-03-17 sg null_verify diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 41226 W: 7246 L: 7199 D: 26781
sprt @ 10+0.1 th 1 Combine my two patches of this topic
17-03-17 sg probcut diff
LLR: -2.94 (-2.94,2.94) [0.00,4.00]
Total: 12345 W: 2152 L: 2266 D: 7927
sprt @ 10+0.1 th 1 Skip early pruning in probcut
17-03-17 sg null_verify diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 51245 W: 9155 L: 9064 D: 33026
sprt @ 10+0.1 th 1 Verify null move always if tt move exists
17-03-17 sg null_verify diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 35587 W: 6282 L: 6258 D: 23047
sprt @ 10+0.1 th 1 Verify null move always at all nodes
17-03-17 sg null_verify diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 12392 W: 2183 L: 2256 D: 7953
sprt @ 10+0.1 th 1 Verify null move always if no tt move exists
16-03-17 sg defended diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 13843 W: 2430 L: 2497 D: 8916
sprt @ 10+0.1 th 1 Only end game bonus
16-03-17 sg defended diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 21765 W: 3811 L: 3845 D: 14109
sprt @ 10+0.1 th 1 Only mid game bonus
16-03-17 sg defended diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 22839 W: 4051 L: 4080 D: 14708
sprt @ 10+0.1 th 1 Now try a decrease of the bonus by 50%. Take 3
16-03-17 sg defended diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 60306 W: 10841 L: 10711 D: 38754
sprt @ 10+0.1 th 1 Quadratic bonus for bad defended opponent pieces.
16-03-17 sg defended diff
LLR: -2.97 (-2.94,2.94) [0.00,5.00]
Total: 7685 W: 1310 L: 1403 D: 4972
sprt @ 10+0.1 th 1 First test struggles so try 50% more bonus. Take 2
15-03-17 sg null_move diff
LLR: -2.97 (-2.94,2.94) [0.00,5.00]
Total: 12024 W: 2066 L: 2141 D: 7817
sprt @ 10+0.1 th 1 Inspired by game from SF 8 against Fire on a recently Graham tournament i tried something for null move avoidance in end games. No null move pruning if only one additional piece (no queen) present and the king is immobile.
09-03-17 sg master diff
ELO: 10.84 +-1.6 (95%) LOS: 100.0%
Total: 40000 W: 4817 L: 3569 D: 31614
40000 @ 60+0.6 th 1 Regression test until "Helper functions to count material for both sides"
14-03-17 sg mate_history diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 10515 W: 1830 L: 1911 D: 6774
sprt @ 10+0.1 th 1 Try the opposite and double the update weight. Take 3
14-03-17 sg mate_history diff
LLR: -2.97 (-2.94,2.94) [0.00,5.00]
Total: 6520 W: 1123 L: 1221 D: 4176
sprt @ 10+0.1 th 1 Half update weight. Take 2
13-03-17 sg mate_history diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 21049 W: 3721 L: 3757 D: 13571
sprt @ 10+0.1 th 1 Fixed broken merge. Add mate move history for quiet move ordering.
12-03-17 sg tune_closedness_mob diff
6785/50000 iterations
14311/100000 games played
100000 @ 10+0.1 th 1 Abandon the four phase idea and try instead a closedness index. First tune mobility. Only new open and closed values (starting with zero) are tuned. ck=10.
10-03-17 sg master_without_simpl diff
ELO: 12.03 +-1.5 (95%) LOS: 100.0%
Total: 39811 W: 4750 L: 3372 D: 31689
40000 @ 60+0.6 th 1 For real comparison i started a regression test against SF 8 with the stripped master (with low throughput). That could then directly compared to current running regression test (base on same master).
10-03-17 sg master_without_simpl diff
ELO: 0.96 +-3.4 (95%) LOS: 70.9%
Total: 14151 W: 2529 L: 2490 D: 9132
30000 @ 10+0.1 th 1 Because it gives concerns about simplifications i have removed from current master all patches which passed with SPRT[-3,1] and are functional changes for one core since SF 8. Now measure the elo difference to the master. Low throughput
10-03-17 sg four_phases_tuned_pawns diff
ELO: -0.23 +-2.9 (95%) LOS: 43.9%
Total: 19941 W: 3536 L: 3549 D: 12856
20000 @ 10+0.1 th 1 Test now the four phases tuned values for pawn structure
07-03-17 sg tune_four_phases_pawns diff
48131/50000 iterations
99375/100000 games played
100000 @ 10+0.1 th 1 The tuning of the pawn psqt with four phases gives neutral result (possible to elo-insensitive). So i will dp a further try, but with a different parameter set. This time the pawn structure eval is tuned (36 parameters). ck=5 and without nodestime
09-03-17 sg futility diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 16364 W: 2852 L: 2908 D: 10604
sprt @ 10+0.1 th 1 Try now my last condition for depth < 7 with a third less margin.
08-03-17 sg futility diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 52027 W: 9301 L: 9207 D: 33519
sprt @ 10+0.1 th 1 Extend to depth 7 if ttMove is a vapture and last move was not. This condition have a very low wrong cutoff rate (2%) in comparison to unconstrained depth 7 (11%).
08-03-17 sg futility2 diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 31725 W: 5624 L: 5616 D: 20485
sprt @ 10+0.1 th 1 The idea seems good, but for more effect try to extend to depth < 9
07-03-17 sg futility diff
LLR: -2.95 (-2.94,2.94) [0.00,4.00]
Total: 21099 W: 3654 L: 3739 D: 13706
sprt @ 10+0.1 th 1 Base on stats measurements of wrong cut off rates i have attempted to reduce the depth but all tests failed. So try the opposite and simple raise the depth bound to 8.
07-03-17 sg futility diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 17731 W: 3117 L: 3167 D: 11447
sprt @ 10+0.1 th 1 If at expected all node allow only depth < 6
06-03-17 sg null_verify diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 17041 W: 2951 L: 3005 D: 11085
sprt @ 10+0.1 th 1 Always verify null move if last move was first move.
06-03-17 sg futility diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 6348 W: 1033 L: 1131 D: 4184
sprt @ 10+0.1 th 1 Do futility pruning only if last two moves was not the first move
06-03-17 sg futility diff
LLR: -2.95 (-2.94,2.94) [0.00,5.00]
Total: 6098 W: 1040 L: 1139 D: 3919
sprt @ 10+0.1 th 1 Do futility pruning only if last move was not the first move
06-03-17 sg futility diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 15051 W: 2632 L: 2694 D: 9725
sprt @ 10+0.1 th 1 Less futility pruning if ttMove exists.
05-03-17 sg move_order diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 21052 W: 3686 L: 3723 D: 13643
sprt @ 10+0.1 th 1 Measured stats indicated the using countermove before second killer gives better move ordering so try this. Secondly i rewrite the first killer as explicit state for more clarity.
04-03-17 sg capture_killer diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 72772 W: 12970 L: 12790 D: 47012
sprt @ 10+0.1 th 1 Now extend my best version to two capture killers.
04-03-17 sg capture_killer diff
LLR: -2.96 (-2.94,2.94) [0.00,5.00]
Total: 47999 W: 8585 L: 8508 D: 30906
sprt @ 10+0.1 th 1 My tests on quiet killers indicate that reduction of reach seems good. Try this now also for my best capture killer patch.