| Active : 5 Machines / 84 Threads / 119.86 MNPS | ||||||
|---|---|---|---|---|---|---|
| Priority 0 | ||||||
| Mhoupp | Stash | thread_voting | diff | 5.0+0.05 | LLR: 0.99 (-2.94, 2.94) [0.00, 5.00] Games: 4056 W: 1034 L: 979 D: 2043 Ptnml(0-2): 28, 482, 976, 491, 51 | Add thread voting to search (Note: impl derived from Stormphrax which is itself ported from Stockfish) |
| Finished | ||||||
| Rosent | Winter | net_end_p * | diff | 8.0+0.08 | LLR: -2.98 (-2.94, 2.94) [-5.00, 0.00] Games: 1216 W: 243 L: 393 D: 580 Ptnml(0-2): 54, 189, 236, 111, 18 | Regression testing net training idea. |
| Kieren | Halogen | spsa_tune_21 | diff | 40.0+0.40 | Tuning 122 Parameters 1501/25000 Iterations 24024/400000 Games Played | |
| Mhoupp | Stash | node_timeman | diff | 40.0+0.40 | LLR: -3.00 (-2.94, 2.94) [0.00, 5.00] Games: 3628 W: 777 L: 871 D: 1980 Ptnml(0-2): 29, 436, 958, 382, 9 | Add node repartition scaling to time management |
| Rosent | Winter | net_end_p * | diff | 8.0+0.08 | LLR: 3.08 (-2.94, 2.94) [-5.00, 0.00] Games: 13636 W: 3293 L: 3219 D: 7124 Ptnml(0-2): 272, 1627, 2966, 1661, 292 | Non-regression sanity check. |
| Mhoupp | Stash | node_timeman | diff | 8.0+0.08 | LLR: -3.04 (-2.94, 2.94) [0.00, 5.00] Games: 8042 W: 1941 L: 2029 D: 4072 Ptnml(0-2): 81, 1028, 1906, 910, 96 | Add node repartition scaling to time management |
| Mhoupp | Stash | lmr_less_for_checks | diff | 40.0+0.40 | LLR: -2.95 (-2.94, 2.94) [0.00, 5.00] Games: 20020 W: 4691 L: 4723 D: 10606 Ptnml(0-2): 107, 2377, 5054, 2385, 87 | Decrease LMR if the move gives check |
| Rosent | Winter | net_end_p | diff | N=64000 | Elo: 7.25 +- 13.90 (95%) [N=1000] Games: 1006 W: 272 L: 251 D: 483 Ptnml(0-2): 19, 118, 213, 129, 24 | Sanity test of net training |
| Kieren | Halogen | r140 | diff | 40.0+0.40 | LLR: -3.00 (-2.94, 2.94) [0.00, 3.00] Games: 23402 W: 5751 L: 5874 D: 11777 Ptnml(0-2): 58, 2740, 6228, 2617, 58 | |
| Kieren | Halogen | r140 | diff | 8.0+0.08 | LLR: -2.96 (-2.94, 2.94) [0.00, 3.00] Games: 21304 W: 5467 L: 5604 D: 10233 Ptnml(0-2): 179, 2559, 5287, 2474, 153 | |
| Kieren | Halogen | r141 | diff | 40.0+0.40 | LLR: -2.96 (-2.94, 2.94) [0.00, 3.00] Games: 16686 W: 4066 L: 4202 D: 8418 Ptnml(0-2): 43, 2006, 4373, 1886, 35 | |
| Vshche | Igel | ttfix | diff | 10.0+0.10 | LLR: -1.04 (-2.94, 2.94) [0.00, 3.00] Games: 8730 W: 2238 L: 2284 D: 4208 Ptnml(0-2): 68, 1081, 2115, 1031, 70 | 3.6.14: improve tt store algo |
| Rosent | Winter | net_502_24x64 * | diff | 8.0+0.08 | LLR: -2.96 (-2.94, 2.94) [0.00, 5.00] Games: 6688 W: 1605 L: 1709 D: 3374 Ptnml(0-2): 168, 856, 1382, 788, 150 | Warmup Stable Decay seems to fit the training pipeline better than the previous linear scheduler based on metrics. Does this hold in self play? |
| Kieren | Halogen | r141 | diff | 8.0+0.08 | LLR: -2.96 (-2.94, 2.94) [0.00, 3.00] Games: 17808 W: 4497 L: 4643 D: 8668 Ptnml(0-2): 153, 2180, 4360, 2082, 129 | |
| Kieren | Halogen | r141 | diff | N=40000 | Elo: 1.53 +- 2.08 (95%) [N=40000] Games: 40038 W: 11527 L: 11351 D: 17160 Ptnml(0-2): 706, 4651, 9046, 4993, 623 | |
| Vshche | Igel | ttfix | diff | 60.0+0.60 | LLR: 0.00 (-2.94, 2.94) [0.00, 3.00] Games: 0 W: 0 L: 0 D: 0 Ptnml(0-2): 0, 0, 0, 0, 0 | 3.6.14: improve tt store algo |
| Kieren | Halogen | r140 | diff | N=40000 | Elo: 4.78 +- 2.11 (95%) [N=40000] Games: 40232 W: 11932 L: 11379 D: 16921 Ptnml(0-2): 715, 4570, 9052, 5005, 774 | |
| Rosent | Winter | net_502_24x64 * | diff | N=40000 | Elo: 2.86 +- 7.28 (95%) [N=4000] Games: 4006 W: 1111 L: 1078 D: 1817 Ptnml(0-2): 102, 478, 824, 483, 116 | Get an understanding of how far apart the based increased size net is compared to the current version. Fixed nodes in this case as system load would otherwise distort results. |
| Rosent | Winter | net_502_16x96 * | diff | 10.0+0.10 | LLR: -2.95 (-2.94, 2.94) [-5.00, 0.00] Games: 8948 W: 1999 L: 2170 D: 4779 Ptnml(0-2): 186, 1157, 1942, 1020, 169 | The larger 24x64 net architecture is failing STC so we instead try 16x96. This has far fewer parameters but still has a 5% slowdown we need to overcome. Based on training metrics this lies squarely between the baseline and 24x64. |
| Kieren | Halogen | extmove-int16-sse4 | diff | 8.0+0.08 | LLR: -1.47 (-2.94, 2.94) [0.00, 3.00] Games: 3118 W: 755 L: 839 D: 1524 Ptnml(0-2): 19, 409, 787, 325, 19 | SIMD movepicker + clamp scores to int16 to compress |
| Rosent | Winter | net_502_24x64 * | diff | 10.0+0.10 | LLR: -2.95 (-2.94, 2.94) [-5.00, 0.00] Games: 1566 W: 309 L: 453 D: 804 Ptnml(0-2): 65, 210, 333, 154, 21 | Having passed the fixed node test as expected, we try an STC test with [-5, 0] bounds, as the slowdown is expected to hurt more than in actual play. |
| Rosent | Winter | net_502_24x64 * | diff | N=40000 | LLR: 2.99 (-2.94, 2.94) [0.00, 5.00] Games: 2282 W: 745 L: 582 D: 955 Ptnml(0-2): 69, 209, 455, 306, 102 | Larger net, hopefully a clear improvement at fixed nodes. Around 16% less Nps. |
| Vshche | Igel | nn-epoch1109 | diff | 10.0+0.10 | LLR: -3.06 (-2.94, 2.94) [0.00, 3.00] Games: 9958 W: 2520 L: 2697 D: 4741 Ptnml(0-2): 94, 1311, 2350, 1126, 98 | 3.6.12: fix tb tt bug (#324) |
| Mhoupp | Stash | lmr_less_for_checks | diff | 8.0+0.08 | LLR: 2.98 (-2.94, 2.94) [0.00, 5.00] Games: 23534 W: 5929 L: 5710 D: 11895 Ptnml(0-2): 251, 2814, 5469, 2931, 302 | Decrease LMR if the move gives check |
| Kieren | Halogen | tt-prefetch-key-after | diff | 4.0+0.04 | LLR: 3.05 (-2.94, 2.94) [0.00, 3.00] Games: 11928 W: 3316 L: 3096 D: 5516 Ptnml(0-2): 100, 1244, 3063, 1450, 107 | Prefetch TT entry before apply_move using precomputed key AVX2: +2.10% +/- 0.19% |
| Kieren | Halogen | [Berserk] main | diff | 8.0+0.08 | Elo: -31.54 +- 11.66 (95%) [N=10000] Games: 1204 W: 273 L: 382 D: 549 Ptnml(0-2): 38, 154, 292, 115, 3 | Update default network to ed164cf2.nn (#890) |