Active : 0 Machines / 0 Threads / 0 MNPS
Priority 0
RosentWinterdisallow_consecutive_null_moves *diff40.0+0.40LLR: -0.30 (-2.94, 2.94) [-4.00, 1.00]
Games: 1584 W: 366 L: 381 D: 837
Ptnml(0-2): 8, 200, 394, 179, 11
Likely the change is exactly neutral at STC, as consecutive null moves require quite a bit of depth. Lets try LTC!
RosentWinterdisallow_consecutive_null_moves *diff8.0+0.08LLR: -1.84 (-2.94, 2.94) [-4.00, 1.00]
Games: 25332 W: 5810 L: 5958 D: 13564
Ptnml(0-2): 487, 3094, 5617, 3016, 452
Consecutive null moves can happen as perspective can make both values fall above beta. This would erroneously register as a position repetition. For this reason we try disallowing repetitions.
Finished
RosentWintermove_order_bugfix *diff40.0+0.40LLR: 2.95 (-2.94, 2.94) [-4.00, 1.00]
Games: 6020 W: 1159 L: 1064 D: 3797
Ptnml(0-2): 40, 628, 1593, 695, 54
Fixed move ordering bug where history heuristics were ignored if previous move was a null move.
RosentWintermove_order_bugfix *diff8.0+0.08LLR: 2.95 (-2.94, 2.94) [-4.00, 1.00]
Games: 11914 W: 2756 L: 2661 D: 6497
Ptnml(0-2): 199, 1398, 2699, 1431, 230
Restarted the test as a non-regression patch. It is unclear how much Elo the change should be worth, but it is the more reasonable form, so it should be added even if Elo neutral or slightly negative.
RosentWintermove_order_bugfix *diff8.0+0.08LLR: 0.04 (-2.94, 2.94) [0.00, 5.00]
Games: 130 W: 30 L: 28 D: 72
Ptnml(0-2): 1, 15, 31, 17, 1
Fixed move ordering bug where history heuristics were ignored if previous move was a null move.
RosentWinterhistory_reductions *diff8.0+0.08LLR: -2.96 (-2.94, 2.94) [0.00, 5.00]
Games: 7890 W: 1770 L: 1863 D: 4257
Ptnml(0-2): 146, 997, 1749, 910, 143
Support for history reductions
RosentWinterqnet_402_d32 *diff40.0+0.40Elo: 71.79 +- 6.17 (95%) [N=4000]
Games: 4000 W: 1273 L: 458 D: 2269
Ptnml(0-2): 12, 185, 915, 752, 136
Regression test vs v5.0
RosentWinterqnet_402_d32 *diff40.0+0.40LLR: 2.95 (-2.94, 2.94) [0.00, 5.00]
Games: 4092 W: 872 L: 744 D: 2476
Ptnml(0-2): 39, 381, 1085, 495, 46
d32 net update.
RosentWinternet382_d24f64b *diff8.0+0.08LLR: 0.79 (-2.94, 2.94) [0.00, 5.00]
Games: 9980 W: 2443 L: 2362 D: 5175
Ptnml(0-2): 212, 1170, 2133, 1275, 200
Added net definition
RosentWinterqnet_382_d32 *diff40.0+0.40LLR: -0.55 (-2.94, 2.94) [0.00, 5.00]
Games: 20744 W: 3953 L: 3894 D: 12897
Ptnml(0-2): 161, 2233, 5548, 2246, 184
STC was passed against older master branch, but this is likely to scale even harder than the new master, so trying directly at LTC.
RosentWinterqnet_382_d32 *diff8.0+0.08LLR: 2.97 (-2.94, 2.94) [0.00, 5.00]
Games: 17098 W: 4326 L: 4111 D: 8661
Ptnml(0-2): 375, 2021, 3560, 2200, 393
Even larger quantized net. Might not yet be sufficiently mature.
RosentWinterqnet_382_d24 *diff40.0+0.40LLR: 2.95 (-2.94, 2.94) [0.00, 5.00]
Games: 3404 W: 749 L: 622 D: 2033
Ptnml(0-2): 28, 331, 873, 426, 44
Now that net is quantized, we try the larger size again.
RosentWinterqnet_382_d24 *diff8.0+0.08LLR: 2.96 (-2.94, 2.94) [0.00, 5.00]
Games: 12870 W: 3262 L: 3070 D: 6538
Ptnml(0-2): 265, 1499, 2728, 1665, 278
Now that net is quantized, we try the larger size again.
RosentWinterquantize_net *diff40.0+0.40LLR: 2.96 (-2.94, 2.94) [0.00, 5.00]
Games: 1382 W: 324 L: 201 D: 857
Ptnml(0-2): 7, 121, 325, 218, 20
Quantized network. Should be faster, at the cost of a little accuracy.
RosentWinterquantize_net *diff8.0+0.08LLR: 2.97 (-2.94, 2.94) [0.00, 5.00]
Games: 1012 W: 301 L: 160 D: 551
Ptnml(0-2): 10, 80, 211, 169, 36
Quantized network. Should be faster, at the cost of a little accuracy.
RosentWinternet382_W50a *diff8.0+0.08LLR: -2.96 (-2.94, 2.94) [0.00, 5.00]
Games: 4510 W: 1066 L: 1178 D: 2266
Ptnml(0-2): 109, 580, 971, 504, 91
Net trained with more uniform data. If this is not negative we are happy because it simplifies our life. If this is negative it is fine as it means we were doing something reasonably fine.
RosentWinternet347_d24f64 *diff40.0+0.40LLR: -2.95 (-2.94, 2.94) [0.00, 5.00]
Games: 4604 W: 872 L: 969 D: 2763
Ptnml(0-2): 60, 555, 1154, 488, 45
Another day another net.
RosentWinternet347_d24f64 *diffN=40000Elo: 25.02 +- 7.56 (95%) [N=4000]
Games: 4020 W: 1277 L: 988 D: 1755
Ptnml(0-2): 95, 414, 767, 575, 159
Preliminary fixed node test with a potential larger net.
RosentWinternet347_W50 *diff40.0+0.40LLR: 2.96 (-2.94, 2.94) [0.00, 5.00]
Games: 4838 W: 986 L: 852 D: 3000
Ptnml(0-2): 44, 487, 1244, 579, 65
Trained with more data.
RosentWinternet347_W50 *diff8.0+0.08LLR: 3.03 (-2.94, 2.94) [0.00, 5.00]
Games: 2440 W: 644 L: 499 D: 1297
Ptnml(0-2): 38, 241, 541, 338, 62
Trained with more data.
RosentWintermaster *diff40.0+0.40Elo: 26.89 +- 6.25 (95%) [N=4000]
Games: 4000 W: 1002 L: 693 D: 2305
Ptnml(0-2): 33, 354, 963, 571, 79
Regression Test vs v5.0
RosentWinternet311_ace501v2 *diff40.0+0.40LLR: 2.95 (-2.94, 2.94) [0.00, 5.00]
Games: 7672 W: 1608 L: 1462 D: 4602
Ptnml(0-2): 83, 791, 1949, 923, 90
This is also trained with WSD but more data. I think I need to train with a longer cooldown phase and have a run with the old scheduler which will hopefully finish this afternoon.
RosentWinternet311_ace501v3 *diff8.0+0.08LLR: -3.05 (-2.94, 2.94) [0.00, 5.00]
Games: 29654 W: 7302 L: 7301 D: 15051
Ptnml(0-2): 656, 3723, 6123, 3614, 711
Comparison of old scheduler vs WSD. Based on my understanding the new scheduler is not properly tuned at the moment, so this should pass.
RosentWinternet311_ace501v2 *diff8.0+0.08LLR: 3.02 (-2.94, 2.94) [0.00, 5.00]
Games: 19066 W: 4895 L: 4667 D: 9504
Ptnml(0-2): 418, 2294, 3920, 2444, 457
This is also trained with WSD but more data. I think I need to train with a longer cooldown phase and have a run with the old scheduler which will hopefully finish this afternoon.
RosentWinternet311_ace501v2 *diff8.0+0.08LLR: -3.04 (-2.94, 2.94) [0.00, 5.00]
Games: 4132 W: 961 L: 1078 D: 2093
Ptnml(0-2): 111, 506, 918, 451, 80
Slight reduction in dataset size, but hopefully higher quality for training. Using WSD scheduler instead of linear. Hoping for small rating gain!
RosentWinternet311_ace501 *diff40.0+0.40LLR: 2.97 (-2.94, 2.94) [-5.00, 0.00]
Games: 15212 W: 2871 L: 2820 D: 9521
Ptnml(0-2): 109, 1621, 4104, 1654, 118
New approach to training net. Hopefully at least Elo neutral but less blunder prone.
RosentWinternet311_ace501 *diff8.0+0.08LLR: 2.95 (-2.94, 2.94) [-5.00, 0.00]
Games: 56412 W: 13466 L: 13594 D: 29352
Ptnml(0-2): 1179, 6842, 12275, 6748, 1162
New approach to training net. Hopefully at least Elo neutral but less blunder prone.