Certified dual-price MoE routing / bench_report.json
Failed on benchmark
1{
2 "bench_version": 1,
3 "track": "router_regime_regression",
4 "model": "custom_routed_experts",
5 "metric_direction": "lower is better",
6 "n_seeds": 8,
7 "baseline": {
8 "best_cfg": {
9 "lr": 0.01,
10 "balance_weight": 0.0
11 },
12 "sweep": [
13 {
14 "cfg": {
15 "lr": 0.001,
16 "balance_weight": 0.0
17 },
18 "mean": 1.2769302129745483
19 },
20 {
21 "cfg": {
22 "lr": 0.001,
23 "balance_weight": 0.01
24 },
25 "mean": 1.2769302129745483
26 },
27 {
28 "cfg": {
29 "lr": 0.001,
30 "balance_weight": 0.1
31 },
32 "mean": 1.2769302129745483
33 },
34 {
35 "cfg": {
36 "lr": 0.003,
37 "balance_weight": 0.0
38 },
39 "mean": 1.0836826115846634
40 },
41 {
42 "cfg": {
43 "lr": 0.003,
44 "balance_weight": 0.01
45 },
46 "mean": 1.0836826115846634
47 },
48 {
49 "cfg": {
50 "lr": 0.003,
51 "balance_weight": 0.1
52 },
53 "mean": 1.0836826115846634
54 },
55 {
56 "cfg": {
57 "lr": 0.01,
58 "balance_weight": 0.0
59 },
60 "mean": 0.5929363369941711
61 },
62 {
63 "cfg": {
64 "lr": 0.01,
65 "balance_weight": 0.01
66 },
67 "mean": 0.5929363369941711
68 },
69 {
70 "cfg": {
71 "lr": 0.01,
72 "balance_weight": 0.1
73 },
74 "mean": 0.5929363369941711
75 }
76 ],
77 "full": {
78 "per_seed": [
79 0.6337619423866272,
80 0.5049523711204529,
81 0.5780711770057678,
82 0.6549598574638367,
83 0.7300989627838135,
84 0.5037400126457214,
85 0.5650720000267029,
86 0.5289599895477295
87 ],
88 "mean": 0.5874520391225815,
89 "std": 0.08006617146577943,
90 "details": [
91 {
92 "metric": 0.6337619423866272,
93 "history": [
94 1.4282171487808228,
95 1.190228533744812,
96 1.092739050388336,
97 1.0344352579116822,
98 0.9784821963310242,
99 0.9456612777709961,
100 0.8985520505905151,
101 0.8510248947143555,
102 0.8331455397605896,
103 0.7536739921569824,
104 0.6857570123672485,
105 0.6390636825561523,
106 0.5909974932670593,
107 0.5384618306159973,
108 0.4952736568450928,
109 0.45394895553588865,
110 0.4213700973987579,
111 0.38855039596557617
112 ],
113 "counts": [
114 12.0,
115 119.0,
116 215.0,
117 54.0
118 ],
119 "prices": [
120 0.0,
121 0.0,
122 0.0,
123 0.0
124 ],
125 "dual_gap": 36.18646240234375,
126 "price_excess_corr": 0.0
127 },
128 {
129 "metric": 0.5049523711204529,
130 "history": [
131 1.5119562721252442,
132 1.1794532775878905,
133 1.0590163993835449,
134 0.9910790395736694,
135 0.9429091310501099,
136 0.8942043042182922,
137 0.8424692082405091,
138 0.815048189163208,
139 0.7707590198516846,
140 0.7232303953170777,
141 0.6701731777191162,
142 0.6477069449424744,
143 0.6097545051574706,
144 0.5813015913963318,
145 0.551750590801239,
146 0.5354827451705932,
147 0.49303611755371096,
148 0.4774340856075287
149 ],
150 "counts": [
151 122.0,
152 67.0,
153 88.0,
154 123.0
155 ],
156 "prices": [
157 0.0,
158 0.0,
159 0.0,
160 0.0
161 ],
162 "dual_gap": -3.1641159057617188,
163 "price_excess_corr": 0.0
164 },
165 {
166 "metric": 0.5780711770057678,
167 "history": [
168 1.818830304145813,
169 1.3692575263977051,
170 1.3353008460998534,
171 1.2090500450134278,
172 1.153546187877655,
173 1.094208278656006,
174 1.0585458588600158,
175 1.0022277736663818,
176 0.9496053886413575,
177 0.9010961532592774,
178 0.8399820518493653,
179 0.7756313908100129,
180 0.7183408498764038,
181 0.6691022443771363,
182 0.6206990718841553,
183 0.5716083288192749,
184 0.5291094565391541,
185 0.5017254376411437
186 ],
187 "counts": [
188 38.0,
189 97.0,
190 37.0,
191 228.0
192 ],
193 "prices": [
194 0.0,
195 0.0,
196 0.0,
197 0.0
198 ],
199 "dual_gap": 20.31761932373047,
200 "price_excess_corr": 0.0
201 },
202 {
203 "metric": 0.6549598574638367,
204 "history": [
205 1.8027701520919799,
206 1.3253757810592652,
207 1.2136915063858031,
208 1.1932248497009277,
209 1.1371701192855834,
210 1.071843571662903,
211 1.003528470993042,
212 0.9638576173782348,
213 0.8964797353744507,
214 0.8383043766021728,
215 0.7723504877090455,
216 0.7259118270874023,
217 0.6827260518074035,
218 0.6220569729804992,
219 0.5784211730957032,
220 0.5290492010116578,
221 0.4989316189289093,
222 0.4740072989463806
223 ],
224 "counts": [
225 31.0,
226 161.0,
227 54.0,
228 154.0
229 ],
230 "prices": [
231 0.0,
232 0.0,
233 0.0,
234 0.0
235 ],
236 "dual_gap": -5.1824951171875,
237 "price_excess_corr": 0.0
238 },
239 {
240 "metric": 0.7300989627838135,
241 "history": [
242 1.599121561050415,
243 1.2852731895446778,
244 1.2019276189804078,
245 1.1541751861572265,
246 1.1143722343444824,
247 1.0603200483322144,
248 1.0094259405136108,
249 0.9526594161987305,
250 0.8896933722496033,
251 0.8333265924453735,
252 0.7677580070495605,
253 0.7211516761779785,
254 0.6531311893463134,
255 0.6131453013420105,
256 0.5731891989707947,
257 0.5463045930862427,
258 0.503038067817688,
259 0.5016564512252808
260 ],
261 "counts": [
262 178.0,
263 168.0,
264 5.0,
265 49.0
266 ],
267 "prices": [
268 0.0,
269 0.0,
270 0.0,
271 0.0
272 ],
273 "dual_gap": 0.041805267333984375,
274 "price_excess_corr": 0.0
275 },
276 {
277 "metric": 0.5037400126457214,
278 "history": [
279 1.6408466291427612,
280 1.4240649843215942,
281 1.3136035633087157,
282 1.229137110710144,
283 1.1278575944900513,
284 1.026499638557434,
285 0.9228047466278076,
286 0.8109598052501679,
287 0.7066186809539795,
288 0.6022843074798584,
289 0.5258250689506531,
290 0.49430953979492187,
291 0.4652757954597473,
292 0.44274723052978515,
293 0.4237631869316101,
294 0.39690431773662566,
295 0.39293508052825926,
296 0.38945595383644105
297 ],
298 "counts": [
299 0.0,
300 47.0,
301 353.0,
302 0.0
303 ],
304 "prices": [
305 0.0,
306 0.0,
307 0.0,
308 0.0
309 ],
310 "dual_gap": 59.741790771484375,
311 "price_excess_corr": 0.0
312 },
313 {
314 "metric": 0.5650720000267029,
315 "history": [
316 1.2339717054367065,
317 0.9556390428543091,
318 0.8510150122642517,
319 0.7696062850952149,
320 0.7309779691696167,
321 0.6998401021957398,
322 0.6714158976078033,
323 0.6412617421150207,
324 0.6248714089393616,
325 0.6193566513061524,
326 0.5911355257034302,
327 0.5783848667144775,
328 0.5490110605955124,
329 0.5357205724716186,
330 0.5112077903747558,
331 0.5009259700775146,
332 0.470411297082901,
333 0.4618472504615784
334 ],
335 "counts": [
336 142.0,
337 141.0,
338 105.0,
339 12.0
340 ],
341 "prices": [
342 0.0,
343 0.0,
344 0.0,
345 0.0
346 ],
347 "dual_gap": 14.618988037109375,
348 "price_excess_corr": 0.0
349 },
350 {
351 "metric": 0.5289599895477295,
352 "history": [
353 1.445556445121765,
354 1.2772163200378417,
355 1.2140379071235656,
356 1.1260969638824463,
357 1.0943745374679565,
358 1.0246259737014771,
359 0.9652957391738891,
360 0.8634569787979126,
361 0.7870700311660767,
362 0.7156670355796814,
363 0.6199389231204987,
364 0.5716872358322144,
365 0.5030802476406098,
366 0.4635558235645294,
367 0.41914095878601076,
368 0.38848516941070554,
369 0.36908327221870424,
370 0.34468125343322753
371 ],
372 "counts": [
373 22.0,
374 1.0,
375 355.0,
376 22.0
377 ],
378 "prices": [
379 0.0,
380 0.0,
381 0.0,
382 0.0
383 ],
384 "dual_gap": 85.62846374511719,
385 "price_excess_corr": 0.0
386 }
387 ]
388 }
389 },
390 "idea": {
391 "per_seed": [
392 0.6441643238067627,
393 0.5522328615188599,
394 0.6585541367530823,
395 0.6564445495605469,
396 0.738909900188446,
397 0.6927487850189209,
398 0.6766970753669739,
399 0.6938974857330322
400 ],
401 "mean": 0.6642061397433281,
402 "std": 0.05406981038533268,
403 "details": [
404 {
405 "metric": 0.6441643238067627,
406 "history": [
407 1.4031563186645508,
408 1.1396907997131347,
409 1.030085563659668,
410 0.9904113817214966,
411 0.9025879764556884,
412 0.8670305633544921,
413 0.8303156185150147,
414 0.7986777687072754,
415 0.7985850524902344,
416 0.7339662647247315,
417 0.698540813922882,
418 0.644290542602539,
419 0.6320906519889832,
420 0.5695153284072876,
421 0.5462218189239502,
422 0.5224982738494873,
423 0.4766631841659546,
424 0.4712458038330078
425 ],
426 "counts": [
427 81.0,
428 114.0,
429 107.0,
430 98.0
431 ],
432 "prices": [
433 0.001999999862164259,
434 0.1849999874830246,
435 0.3819999098777771,
436 0.11400001496076584
437 ],
438 "dual_gap": -0.86676025390625,
439 "price_excess_corr": 0.7315175533294678
440 },
441 {
442 "metric": 0.5522328615188599,
443 "history": [
444 1.5037132596969605,
445 1.1599038314819337,
446 1.046072883605957,
447 0.9561529111862183,
448 0.9125625324249268,
449 0.8491886901855469,
450 0.7865666329860688,
451 0.7643450164794922,
452 0.7428707802295684,
453 0.6975994086265564,
454 0.6477171063423157,
455 0.6253858184814454,
456 0.6086556792259217,
457 0.5550899171829223,
458 0.527807469367981,
459 0.5323403787612915,
460 0.5048370456695557,
461 0.4717504811286926
462 ],
463 "counts": [
464 92.0,
465 96.0,
466 100.0,
467 112.0
468 ],
469 "prices": [
470 0.07299996167421341,
471 0.005000000819563866,
472 0.01600000075995922,
473 0.05800000950694084
474 ],
475 "dual_gap": -0.9556369781494141,
476 "price_excess_corr": 0.10869929194450378
477 },
478 {
479 "metric": 0.6585541367530823,
480 "history": [
481 1.761208825111389,
482 1.1728641319274902,
483 1.1203809022903441,
484 0.9686005115509033,
485 0.9359130239486695,
486 0.8751321792602539,
487 0.8234063625335694,
488 0.8007956433296204,
489 0.7692797470092774,
490 0.7613382744789123,
491 0.7146817946434021,
492 0.699941680431366,
493 0.6649033844470977,
494 0.6361303853988648,
495 0.6227840328216553,
496 0.6006997179985046,
497 0.56922039270401,
498 0.5525556755065918
499 ],
500 "counts": [
501 95.0,
502 110.0,
503 79.0,
504 116.0
505 ],
506 "prices": [
507 0.008000001311302185,
508 0.08899999409914017,
509 0.003999999258667231,
510 0.2290000021457672
511 ],
512 "dual_gap": -2.62347412109375,
513 "price_excess_corr": 0.8477886915206909
514 },
515 {
516 "metric": 0.6564445495605469,
517 "history": [
518 1.7998930931091308,
519 1.2915659427642823,
520 1.1516711568832398,
521 1.1157651948928833,
522 1.0996507930755615,
523 1.0418196868896485,
524 0.9867689752578735,
525 0.9761204719543457,
526 0.9016803455352783,
527 0.87863596200943,
528 0.8343607997894287,
529 0.828931360244751,
530 0.765202465057373,
531 0.7308063387870789,
532 0.6686744666099549,
533 0.634844012260437,
534 0.5963686537742615,
535 0.5574372768402099
536 ],
537 "counts": [
538 86.0,
539 108.0,
540 99.0,
541 107.0
542 ],
543 "prices": [
544 0.053999993950128555,
545 0.13999998569488525,
546 2.3283064365386963e-10,
547 0.1860000044107437
548 ],
549 "dual_gap": -4.817523956298828,
550 "price_excess_corr": 0.6527703404426575
551 },
552 {
553 "metric": 0.738909900188446,
554 "history": [
555 1.5758513450622558,
556 1.1992204332351684,
557 1.0975655937194824,
558 1.106093807220459,
559 1.0530373549461365,
560 0.9710775756835938,
561 0.9341824054718018,
562 0.8848945617675781,
563 0.8314888620376587,
564 0.8328626918792724,
565 0.7714112186431885,
566 0.7077824592590332,
567 0.6948019361495972,
568 0.6471130168437957,
569 0.6198636960983276,
570 0.5824776190519333,
571 0.5528994226455688,
572 0.5487739419937134
573 ],
574 "counts": [
575 81.0,
576 92.0,
577 90.0,
578 137.0
579 ],
580 "prices": [
581 0.3009999692440033,
582 0.31300005316734314,
583 0.0,
584 0.09099999070167542
585 ],
586 "dual_gap": 0.3594779968261719,
587 "price_excess_corr": -0.414129376411438
588 },
589 {
590 "metric": 0.6927487850189209,
591 "history": [
592 1.578282914161682,
593 1.234548954963684,
594 1.0721720218658448,
595 1.0114439725875854,
596 0.9600631952285766,
597 0.9064272737503052,
598 0.8845085048675537,
599 0.8476518750190735,
600 0.7936444997787475,
601 0.7221487879753112,
602 0.675731315612793,
603 0.6356342124938965,
604 0.5966529107093811,
605 0.5677075898647308,
606 0.542996826171875,
607 0.5033763444423676,
608 0.47669029474258423,
609 0.46219895362854
610 ],
611 "counts": [
612 80.0,
613 120.0,
614 111.0,
615 89.0
616 ],
617 "prices": [
618 0.0,
619 0.40799999237060547,
620 0.6180000305175781,
621 0.08100000023841858
622 ],
623 "dual_gap": -11.450798034667969,
624 "price_excess_corr": 0.874034583568573
625 },
626 {
627 "metric": 0.6766970753669739,
628 "history": [
629 1.2320793676376343,
630 0.9413762331008911,
631 0.8169021797180176,
632 0.7195290040969848,
633 0.691906635761261,
634 0.6519558382034302,
635 0.6347930026054383,
636 0.6266996812820435,
637 0.6043409276008606,
638 0.6051623976230621,
639 0.5799094390869141,
640 0.5754379844665527,
641 0.5715164792537689,
642 0.5350594699382782,
643 0.522064962387085,
644 0.5143913173675537,
645 0.4961080849170685,
646 0.512918781042099
647 ],
648 "counts": [
649 131.0,
650 106.0,
651 83.0,
652 80.0
653 ],
654 "prices": [
655 0.2599999010562897,
656 0.3229999244213104,
657 0.3390001356601715,
658 0.0
659 ],
660 "dual_gap": 2.3051986694335938,
661 "price_excess_corr": 0.3783074617385864
662 },
663 {
664 "metric": 0.6938974857330322,
665 "history": [
666 1.3202440786361693,
667 1.056475853919983,
668 0.9645787239074707,
669 0.809308934211731,
670 0.7542427253723144,
671 0.694851930141449,
672 0.6566034412384033,
673 0.5901383829116821,
674 0.6083235359191894,
675 0.6138130974769592,
676 0.6434554958343506,
677 0.6685442996025085,
678 0.6232259523868561,
679 0.5891516327857971,
680 0.5718476819992065,
681 0.5396019458770752,
682 0.5289167666435242,
683 0.4906179118156433
684 ],
685 "counts": [
686 81.0,
687 95.0,
688 129.0,
689 95.0
690 ],
691 "prices": [
692 0.09799998253583908,
693 0.0,
694 0.5430000424385071,
695 0.1929999589920044
696 ],
697 "dual_gap": -0.504119873046875,
698 "price_excess_corr": 0.891348659992218
699 }
700 ],
701 "best_cfg": {
702 "lr": 0.01,
703 "rho": 0.001
704 }
705 },
706 "comparison": {
707 "delta_mean": 0.07675410062074661,
708 "idea_wins": 0,
709 "n_pairs": 8,
710 "per_seed_diffs": [
711 0.010402381420135498,
712 0.04728049039840698,
713 0.08048295974731445,
714 0.001484692096710205,
715 0.008810937404632568,
716 0.18900877237319946,
717 0.111625075340271,
718 0.16493749618530273
719 ],
720 "p_value": 0.0081,
721 "mde": 0.06092237367016021,
722 "mde_rel_pct": 10.37061234158858,
723 "verdict": "idea worse (significant)",
724 "system_worked": false
725 },
726 "mechanism_signature": {
727 "mechanism_signature": {
728 "prediction": "positive expert prices should correspond to positive realized load excess and reduce load imbalance",
729 "trained_price_excess_correlations": [
730 0.7315175533294678,
731 0.10869929194450378,
732 0.8477886915206909,
733 0.6527703404426575,
734 -0.414129376411438,
735 0.874034583568573,
736 0.3783074617385864,
737 0.891348659992218
738 ],
739 "trained_dual_gaps": [
740 -0.86676025390625,
741 -0.9556369781494141,
742 -2.62347412109375,
743 -4.817523956298828,
744 0.3594779968261719,
745 -11.450798034667969,
746 2.3051986694335938,
747 -0.504119873046875
748 ],
749 "mean_abs_dual_gap": 2.9853737354278564,
750 "confirmed": true
751 },
752 "idea_sweep": [
753 {
754 "cfg": {
755 "lr": 0.01,
756 "rho": 0.001
757 },
758 "result": {
759 "per_seed": [
760 0.6441643238067627,
761 0.5522328615188599,
762 0.6585541367530823,
763 0.6564445495605469,
764 0.738909900188446,
765 0.6927487850189209,
766 0.6766970753669739,
767 0.6938974857330322
768 ],
769 "mean": 0.6642061397433281,
770 "std": 0.05406981038533268,
771 "details": [
772 {
773 "metric": 0.6441643238067627,
774 "history": [
775 1.4031563186645508,
776 1.1396907997131347,
777 1.030085563659668,
778 0.9904113817214966,
779 0.9025879764556884,
780 0.8670305633544921,
781 0.8303156185150147,
782 0.7986777687072754,
783 0.7985850524902344,
784 0.7339662647247315,
785 0.698540813922882,
786 0.644290542602539,
787 0.6320906519889832,
788 0.5695153284072876,
789 0.5462218189239502,
790 0.5224982738494873,
791 0.4766631841659546,
792 0.4712458038330078
793 ],
794 "counts": [
795 81.0,
796 114.0,
797 107.0,
798 98.0
799 ],
800 "prices": [
801 0.001999999862164259,
802 0.1849999874830246,
803 0.3819999098777771,
804 0.11400001496076584
805 ],
806 "dual_gap": -0.86676025390625,
807 "price_excess_corr": 0.7315175533294678
808 },
809 {
810 "metric": 0.5522328615188599,
811 "history": [
812 1.5037132596969605,
813 1.1599038314819337,
814 1.046072883605957,
815 0.9561529111862183,
816 0.9125625324249268,
817 0.8491886901855469,
818 0.7865666329860688,
819 0.7643450164794922,
820 0.7428707802295684,
821 0.6975994086265564,
822 0.6477171063423157,
823 0.6253858184814454,
824 0.6086556792259217,
825 0.5550899171829223,
826 0.527807469367981,
827 0.5323403787612915,
828 0.5048370456695557,
829 0.4717504811286926
830 ],
831 "counts": [
832 92.0,
833 96.0,
834 100.0,
835 112.0
836 ],
837 "prices": [
838 0.07299996167421341,
839 0.005000000819563866,
840 0.01600000075995922,
841 0.05800000950694084
842 ],
843 "dual_gap": -0.9556369781494141,
844 "price_excess_corr": 0.10869929194450378
845 },
846 {
847 "metric": 0.6585541367530823,
848 "history": [
849 1.761208825111389,
850 1.1728641319274902,
851 1.1203809022903441,
852 0.9686005115509033,
853 0.9359130239486695,
854 0.8751321792602539,
855 0.8234063625335694,
856 0.8007956433296204,
857 0.7692797470092774,
858 0.7613382744789123,
859 0.7146817946434021,
860 0.699941680431366,
861 0.6649033844470977,
862 0.6361303853988648,
863 0.6227840328216553,
864 0.6006997179985046,
865 0.56922039270401,
866 0.5525556755065918
867 ],
868 "counts": [
869 95.0,
870 110.0,
871 79.0,
872 116.0
873 ],
874 "prices": [
875 0.008000001311302185,
876 0.08899999409914017,
877 0.003999999258667231,
878 0.2290000021457672
879 ],
880 "dual_gap": -2.62347412109375,
881 "price_excess_corr": 0.8477886915206909
882 },
883 {
884 "metric": 0.6564445495605469,
885 "history": [
886 1.7998930931091308,
887 1.2915659427642823,
888 1.1516711568832398,
889 1.1157651948928833,
890 1.0996507930755615,
891 1.0418196868896485,
892 0.9867689752578735,
893 0.9761204719543457,
894 0.9016803455352783,
895 0.87863596200943,
896 0.8343607997894287,
897 0.828931360244751,
898 0.765202465057373,
899 0.7308063387870789,
900 0.6686744666099549,
901 0.634844012260437,
902 0.5963686537742615,
903 0.5574372768402099
904 ],
905 "counts": [
906 86.0,
907 108.0,
908 99.0,
909 107.0
910 ],
911 "prices": [
912 0.053999993950128555,
913 0.13999998569488525,
914 2.3283064365386963e-10,
915 0.1860000044107437
916 ],
917 "dual_gap": -4.817523956298828,
918 "price_excess_corr": 0.6527703404426575
919 },
920 {
921 "metric": 0.738909900188446,
922 "history": [
923 1.5758513450622558,
924 1.1992204332351684,
925 1.0975655937194824,
926 1.106093807220459,
927 1.0530373549461365,
928 0.9710775756835938,
929 0.9341824054718018,
930 0.8848945617675781,
931 0.8314888620376587,
932 0.8328626918792724,
933 0.7714112186431885,
934 0.7077824592590332,
935 0.6948019361495972,
936 0.6471130168437957,
937 0.6198636960983276,
938 0.5824776190519333,
939 0.5528994226455688,
940 0.5487739419937134
941 ],
942 "counts": [
943 81.0,
944 92.0,
945 90.0,
946 137.0
947 ],
948 "prices": [
949 0.3009999692440033,
950 0.31300005316734314,
951 0.0,
952 0.09099999070167542
953 ],
954 "dual_gap": 0.3594779968261719,
955 "price_excess_corr": -0.414129376411438
956 },
957 {
958 "metric": 0.6927487850189209,
959 "history": [
960 1.578282914161682,
961 1.234548954963684,
962 1.0721720218658448,
963 1.0114439725875854,
964 0.9600631952285766,
965 0.9064272737503052,
966 0.8845085048675537,
967 0.8476518750190735,
968 0.7936444997787475,
969 0.7221487879753112,
970 0.675731315612793,
971 0.6356342124938965,
972 0.5966529107093811,
973 0.5677075898647308,
974 0.542996826171875,
975 0.5033763444423676,
976 0.47669029474258423,
977 0.46219895362854
978 ],
979 "counts": [
980 80.0,
981 120.0,
982 111.0,
983 89.0
984 ],
985 "prices": [
986 0.0,
987 0.40799999237060547,
988 0.6180000305175781,
989 0.08100000023841858
990 ],
991 "dual_gap": -11.450798034667969,
992 "price_excess_corr": 0.874034583568573
993 },
994 {
995 "metric": 0.6766970753669739,
996 "history": [
997 1.2320793676376343,
998 0.9413762331008911,
999 0.8169021797180176,
1000 0.7195290040969848,
1001 0.691906635761261,
1002 0.6519558382034302,
1003 0.6347930026054383,
1004 0.6266996812820435,
1005 0.6043409276008606,
1006 0.6051623976230621,
1007 0.5799094390869141,
1008 0.5754379844665527,
1009 0.5715164792537689,
1010 0.5350594699382782,
1011 0.522064962387085,
1012 0.5143913173675537,
1013 0.4961080849170685,
1014 0.512918781042099
1015 ],
1016 "counts": [
1017 131.0,
1018 106.0,
1019 83.0,
1020 80.0
1021 ],
1022 "prices": [
1023 0.2599999010562897,
1024 0.3229999244213104,
1025 0.3390001356601715,
1026 0.0
1027 ],
1028 "dual_gap": 2.3051986694335938,
1029 "price_excess_corr": 0.3783074617385864
1030 },
1031 {
1032 "metric": 0.6938974857330322,
1033 "history": [
1034 1.3202440786361693,
1035 1.056475853919983,
1036 0.9645787239074707,
1037 0.809308934211731,
1038 0.7542427253723144,
1039 0.694851930141449,
1040 0.6566034412384033,
1041 0.5901383829116821,
1042 0.6083235359191894,
1043 0.6138130974769592,
1044 0.6434554958343506,
1045 0.6685442996025085,
1046 0.6232259523868561,
1047 0.5891516327857971,
1048 0.5718476819992065,
1049 0.5396019458770752,
1050 0.5289167666435242,
1051 0.4906179118156433
1052 ],
1053 "counts": [
1054 81.0,
1055 95.0,
1056 129.0,
1057 95.0
1058 ],
1059 "prices": [
1060 0.09799998253583908,
1061 0.0,
1062 0.5430000424385071,
1063 0.1929999589920044
1064 ],
1065 "dual_gap": -0.504119873046875,
1066 "price_excess_corr": 0.891348659992218
1067 }
1068 ],
1069 "best_cfg": {
1070 "lr": 0.01,
1071 "rho": 0.001
1072 }
1073 }
1074 },
1075 {
1076 "cfg": {
1077 "lr": 0.01,
1078 "rho": 0.01
1079 },
1080 "result": {
1081 "per_seed": [
1082 0.6262717843055725,
1083 0.6611071825027466,
1084 0.7047536969184875,
1085 0.6719289422035217,
1086 0.7768282890319824,
1087 0.6908259391784668,
1088 0.6206275224685669,
1089 0.7099313735961914
1090 ],
1091 "mean": 0.682784341275692,
1092 "std": 0.050389091050960635,
1093 "details": [
1094 {
1095 "metric": 0.6262717843055725,
1096 "history": [
1097 1.3039574146270752,
1098 1.1375168132781983,
1099 0.9833345127105713,
1100 1.0023479652404785,
1101 0.9165901398658752,
1102 0.853428463935852,
1103 0.828366084098816,
1104 0.7824184989929199,
1105 0.730939371585846,
1106 0.7010899758338929,
1107 0.6970420432090759,
1108 0.619460997581482,
1109 0.6214600610733032,
1110 0.6054762434959412,
1111 0.5303078866004944,
1112 0.4873678207397461,
1113 0.4333805441856384,
1114 0.47833149433135985
1115 ],
1116 "counts": [
1117 87.0,
1118 120.0,
1119 103.0,
1120 90.0
1121 ],
1122 "prices": [
1123 0.04999999701976776,
1124 0.2099999338388443,
1125 0.45999982953071594,
1126 0.18000000715255737
1127 ],
1128 "dual_gap": 0.7860870361328125,
1129 "price_excess_corr": 0.4049873650074005
1130 },
1131 {
1132 "metric": 0.6611071825027466,
1133 "history": [
1134 1.525277624130249,
1135 1.2126392173767089,
1136 1.0078798913955689,
1137 1.0071892642974853,
1138 0.9479112243652343,
1139 0.9446576690673828,
1140 0.8098049712181091,
1141 0.9090130233764648,
1142 0.8573304033279419,
1143 0.7679864072799683,
1144 0.7141460466384888,
1145 0.6927071332931518,
1146 0.6666260385513305,
1147 0.6555062329769135,
1148 0.6064930319786072,
1149 0.6451085591316223,
1150 0.6082721376419067,
1151 0.6090406477451324
1152 ],
1153 "counts": [
1154 97.0,
1155 83.0,
1156 145.0,
1157 75.0
1158 ],
1159 "prices": [
1160 0.1499999314546585,
1161 0.09000000357627869,
1162 0.030000029131770134,
1163 0.16999995708465576
1164 ],
1165 "dual_gap": -1.5525569915771484,
1166 "price_excess_corr": -0.8204752206802368
1167 },
1168 {
1169 "metric": 0.7047536969184875,
1170 "history": [
1171 1.7032924365997315,
1172 1.1933565044403076,
1173 1.1769169855117798,
1174 0.9702837872505188,
1175 0.9021575140953064,
1176 0.9135065031051636,
1177 0.8400838613510132,
1178 0.8087026262283326,
1179 0.76284184217453,
1180 0.7864427018165588,
1181 0.7398368334770202,
1182 0.6987604725360871,
1183 0.7155058979988098,
1184 0.7891936492919922,
1185 0.6775259113311768,
1186 0.6355217003822327,
1187 0.6297998356819153,
1188 0.6280114138126374
1189 ],
1190 "counts": [
1191 104.0,
1192 113.0,
1193 77.0,
1194 106.0
1195 ],
1196 "prices": [
1197 0.0599999725818634,
1198 0.13999998569488525,
1199 0.07999999076128006,
1200 0.3200000822544098
1201 ],
1202 "dual_gap": -3.084991455078125,
1203 "price_excess_corr": 0.3812926113605499
1204 },
1205 {
1206 "metric": 0.6719289422035217,
1207 "history": [
1208 1.7835893440246582,
1209 1.3398815345764161,
1210 1.1745865392684935,
1211 1.1184628486633301,
1212 1.0979045367240905,
1213 1.087763638496399,
1214 1.085881280899048,
1215 1.0113541078567505,
1216 0.9435374069213868,
1217 0.938734905719757,
1218 0.922377724647522,
1219 0.8435751342773438,
1220 0.7949281311035157,
1221 0.7932442426681519,
1222 0.7761956405639648,
1223 0.7285706949234009,
1224 0.698066930770874,
1225 0.67836585521698
1226 ],
1227 "counts": [
1228 126.0,
1229 56.0,
1230 113.0,
1231 105.0
1232 ],
1233 "prices": [
1234 0.09999997168779373,
1235 0.24999991059303284,
1236 0.019999999552965164,
1237 0.24999995529651642
1238 ],
1239 "dual_gap": 0.5536575317382812,
1240 "price_excess_corr": -0.6561136245727539
1241 },
1242 {
1243 "metric": 0.7768282890319824,
1244 "history": [
1245 1.4601379203796387,
1246 1.1082266092300415,
1247 1.1362065076828003,
1248 1.0652369022369386,
1249 0.9823706912994384,
1250 0.9776483631134033,
1251 0.9667921781539917,
1252 0.9053830432891846,
1253 0.9509675192832947,
1254 0.8583141326904297,
1255 0.8127024602890015,
1256 0.8199803376197815,
1257 0.7261572551727294,
1258 0.7076331377029419,
1259 0.762781138420105,
1260 0.6751837837696075,
1261 0.6537091398239135,
1262 0.6659013748168945
1263 ],
1264 "counts": [
1265 117.0,
1266 95.0,
1267 86.0,
1268 102.0
1269 ],
1270 "prices": [
1271 0.3199998140335083,
1272 0.3899996280670166,
1273 0.0599999874830246,
1274 0.2099999189376831
1275 ],
1276 "dual_gap": -3.830963134765625,
1277 "price_excess_corr": 0.5433895587921143
1278 },
1279 {
1280 "metric": 0.6908259391784668,
1281 "history": [
1282 1.6264865016937256,
1283 1.1661974811553955,
1284 0.9919495415687561,
1285 0.9857214403152466,
1286 0.8792146921157837,
1287 0.9436957502365112,
1288 0.9131818580627441,
1289 0.7837913727760315,
1290 0.8391086673736572,
1291 0.7410305190086365,
1292 0.67766282081604,
1293 0.6405617022514343,
1294 0.654193127155304,
1295 0.5937393045425415,
1296 0.5505223798751832,
1297 0.48757129430770874,
1298 0.47665258169174196,
1299 0.5111252760887146
1300 ],
1301 "counts": [
1302 105.0,
1303 91.0,
1304 122.0,
1305 82.0
1306 ],
1307 "prices": [
1308 0.029999995604157448,
1309 0.5300000905990601,
1310 0.6700002551078796,
1311 0.1600000113248825
1312 ],
1313 "dual_gap": -4.6856689453125,
1314 "price_excess_corr": 0.4581056237220764
1315 },
1316 {
1317 "metric": 0.6206275224685669,
1318 "history": [
1319 1.2146726608276368,
1320 0.8417455530166627,
1321 0.8210588359832763,
1322 0.8491046547889709,
1323 0.7547090137004853,
1324 0.7639045429229736,
1325 0.705108392238617,
1326 0.693201768398285,
1327 0.6587955093383789,
1328 0.6621153998374939,
1329 0.6825254893302918,
1330 0.6306395292282104,
1331 0.6202736258506775,
1332 0.6084072184562683,
1333 0.5361759781837463,
1334 0.6167222261428833,
1335 0.5254430556297303,
1336 0.49973246693611145
1337 ],
1338 "counts": [
1339 98.0,
1340 82.0,
1341 92.0,
1342 128.0
1343 ],
1344 "prices": [
1345 0.37999996542930603,
1346 0.4399999976158142,
1347 0.41999995708465576,
1348 0.030000001192092896
1349 ],
1350 "dual_gap": 6.8090972900390625,
1351 "price_excess_corr": -0.9755735397338867
1352 },
1353 {
1354 "metric": 0.7099313735961914,
1355 "history": [
1356 1.273847484588623,
1357 0.9680283117294312,
1358 0.9676177501678467,
1359 0.7686456203460693,
1360 0.744644775390625,
1361 0.8059906673431396,
1362 0.7485140085220336,
1363 0.692188173532486,
1364 0.8139085674285889,
1365 0.8050192737579346,
1366 0.7376338243484497,
1367 0.6645472645759583,
1368 0.6088972043991089,
1369 0.5972435641288757,
1370 0.5730879020690918,
1371 0.6306612575054169,
1372 0.5813390743732453,
1373 0.47985951900482177
1374 ],
1375 "counts": [
1376 95.0,
1377 113.0,
1378 76.0,
1379 116.0
1380 ],
1381 "prices": [
1382 0.13999997079372406,
1383 0.039999984204769135,
1384 0.6700000762939453,
1385 0.23999977111816406
1386 ],
1387 "dual_gap": 11.904136657714844,
1388 "price_excess_corr": -0.8073238730430603
1389 }
1390 ]
1391 }
1392 },
1393 {
1394 "cfg": {
1395 "lr": 0.01,
1396 "rho": 0.1
1397 },
1398 "result": {
1399 "per_seed": [
1400 1.1683597564697266,
1401 0.9057276248931885,
1402 0.9819811582565308,
1403 1.0769124031066895,
1404 1.124353051185608,
1405 0.979142427444458,
1406 1.2250404357910156,
1407 1.028434157371521
1408 ],
1409 "mean": 1.0612438768148422,
1410 "std": 0.10744859492285688,
1411 "details": [
1412 {
1413 "metric": 1.1683597564697266,
1414 "history": [
1415 1.4490776538848877,
1416 1.5859220504760743,
1417 1.3607615947723388,
1418 1.4989393043518067,
1419 1.3760244274139404,
1420 1.3536517190933228,
1421 1.3585372638702393,
1422 1.3083824491500855,
1423 1.2400494956970214,
1424 1.3978913164138793,
1425 1.268368067741394,
1426 1.2482490491867066,
1427 1.2231985998153687,
1428 1.1448024535179138,
1429 1.1257777547836303,
1430 1.0693637561798095,
1431 0.9765922141075134,
1432 1.0164781403541565
1433 ],
1434 "counts": [
1435 0.0,
1436 0.0,
1437 321.0,
1438 79.0
1439 ],
1440 "prices": [
1441 5.100002765655518,
1442 4.300004005432129,
1443 1.3000054359436035,
1444 1.3000035285949707
1445 ],
1446 "dual_gap": 711.1893463134766,
1447 "price_excess_corr": -0.7495322823524475
1448 },
1449 {
1450 "metric": 0.9057276248931885,
1451 "history": [
1452 1.6662955570220948,
1453 1.6366234970092775,
1454 1.5101047563552856,
1455 1.5987392616271974,
1456 1.4745655584335327,
1457 1.4979643392562867,
1458 1.3766446256637572,
1459 1.3513256263732911,
1460 1.3477717399597169,
1461 1.3279736328125,
1462 1.2432228183746339,
1463 1.2315861248970033,
1464 1.1717516899108886,
1465 1.0781730556488036,
1466 1.0162011313438415,
1467 0.9924906349182129,
1468 0.8533278799057007,
1469 0.8498016858100891
1470 ],
1471 "counts": [
1472 92.0,
1473 0.0,
1474 0.0,
1475 308.0
1476 ],
1477 "prices": [
1478 2.300004005432129,
1479 3.900001049041748,
1480 4.60000467300415,
1481 2.100005626678467
1482 ],
1483 "dual_gap": 400.21055603027344,
1484 "price_excess_corr": -0.8114298582077026
1485 },
1486 {
1487 "metric": 0.9819811582565308,
1488 "history": [
1489 1.8330887985229491,
1490 1.7522169303894044,
1491 1.7377414941787719,
1492 1.614824047088623,
1493 1.5683921241760255,
1494 1.5938614463806153,
1495 1.5247077465057373,
1496 1.5466945004463195,
1497 1.404871940612793,
1498 1.451461374759674,
1499 1.3511654186248778,
1500 1.3450501358509064,
1501 1.2249249458312987,
1502 1.2138253259658813,
1503 1.1642186546325684,
1504 1.0807652282714844,
1505 0.9764101076126098,
1506 0.971161847114563
1507 ],
1508 "counts": [
1509 110.0,
1510 290.0,
1511 0.0,
1512 0.0
1513 ],
1514 "prices": [
1515 2.0000011920928955,
1516 1.8000009059906006,
1517 4.000002861022949,
1518 5.100006580352783
1519 ],
1520 "dual_gap": 533.8030090332031,
1521 "price_excess_corr": -0.8358966112136841
1522 },
1523 {
1524 "metric": 1.0769124031066895,
1525 "history": [
1526 1.7902608394622803,
1527 1.4834549045562744,
1528 1.5026324462890626,
1529 1.3873037981987,
1530 1.426707820892334,
1531 1.3190795135498048,
1532 1.2909542369842528,
1533 1.3072921562194824,
1534 1.1851996898651123,
1535 1.2980601835250853,
1536 1.1551457309722901,
1537 1.1571266746520996,
1538 1.0714560890197753,
1539 1.176503300666809,
1540 1.0852110576629639,
1541 1.0225802326202393,
1542 0.9678384566307068,
1543 0.9272657442092895
1544 ],
1545 "counts": [
1546 0.0,
1547 3.0,
1548 397.0,
1549 0.0
1550 ],
1551 "prices": [
1552 2.900002956390381,
1553 2.8000073432922363,
1554 1.5000014305114746,
1555 5.800005912780762
1556 ],
1557 "dual_gap": 567.5848979949951,
1558 "price_excess_corr": -0.6454119682312012
1559 },
1560 {
1561 "metric": 1.124353051185608,
1562 "history": [
1563 1.7129732704162597,
1564 1.7161536598205567,
1565 1.6164196014404297,
1566 1.5220919799804689,
1567 1.5200805282592773,
1568 1.4736564350128174,
1569 1.5009299993515015,
1570 1.4522486782073976,
1571 1.4154387760162352,
1572 1.405643253326416,
1573 1.3363902616500853,
1574 1.2636822080612182,
1575 1.3516197395324707,
1576 1.2489344120025634,
1577 1.1715259742736817,
1578 1.1471354389190673,
1579 1.05180935382843,
1580 1.0683834600448607
1581 ],
1582 "counts": [
1583 0.0,
1584 8.0,
1585 392.0,
1586 0.0
1587 ],
1588 "prices": [
1589 4.000000953674316,
1590 2.30000638961792,
1591 1.200000286102295,
1592 4.800002098083496
1593 ],
1594 "dual_gap": 618.683934211731,
1595 "price_excess_corr": -0.7796272039413452
1596 },
1597 {
1598 "metric": 0.979142427444458,
1599 "history": [
1600 1.7750358438491822,
1601 1.5844729232788086,
1602 1.6429149341583251,
1603 1.5185857009887695,
1604 1.502098972797394,
1605 1.544996886253357,
1606 1.422248454093933,
1607 1.4559745836257934,
1608 1.4031295108795165,
1609 1.3522978973388673,
1610 1.300237669944763,
1611 1.2396622562408448,
1612 1.217160804271698,
1613 1.1100136756896972,
1614 1.1584144115447998,
1615 0.981019310951233,
1616 1.005128688812256,
1617 0.8883059167861939
1618 ],
1619 "counts": [
1620 0.0,
1621 2.0,
1622 398.0,
1623 0.0
1624 ],
1625 "prices": [
1626 4.800004005432129,
1627 2.2000069618225098,
1628 1.9000087976455688,
1629 4.000002861022949
1630 ],
1631 "dual_gap": 597.8818855285645,
1632 "price_excess_corr": -0.6340560913085938
1633 },
1634 {
1635 "metric": 1.2250404357910156,
1636 "history": [
1637 1.3113063716888427,
1638 1.2002628803253175,
1639 1.4514206790924071,
1640 1.330623984336853,
1641 1.232826681137085,
1642 1.2343617129325866,
1643 1.2105679941177367,
1644 1.170863471031189,
1645 1.1571433877944945,
1646 1.1489333415031433,
1647 1.1103823566436768,
1648 1.1372274017333985,
1649 1.0873634719848633,
1650 1.0486375761032105,
1651 0.9828113007545471,
1652 1.0502234315872192,
1653 0.9858989048004151,
1654 0.9104355192184448
1655 ],
1656 "counts": [
1657 0.0,
1658 400.0,
1659 0.0,
1660 0.0
1661 ],
1662 "prices": [
1663 3.600005626678467,
1664 1.30000638961792,
1665 5.700007438659668,
1666 2.0000052452087402
1667 ],
1668 "dual_gap": 776.7353515625,
1669 "price_excess_corr": -0.6313034892082214
1670 },
1671 {
1672 "metric": 1.028434157371521,
1673 "history": [
1674 1.684798412322998,
1675 1.5274455451965332,
1676 1.4914986324310302,
1677 1.3847495126724243,
1678 1.5005753707885743,
1679 1.4229538440704346,
1680 1.40796156167984,
1681 1.3940145444869996,
1682 1.3152224588394166,
1683 1.2695887804031372,
1684 1.2292287492752074,
1685 1.0851865339279174,
1686 1.1410926818847655,
1687 1.0864189529418946,
1688 0.9851797819137573,
1689 0.9458699417114258,
1690 0.834864010810852,
1691 0.8689987945556641
1692 ],
1693 "counts": [
1694 0.0,
1695 400.0,
1696 0.0,
1697 0.0
1698 ],
1699 "prices": [
1700 1.9000014066696167,
1701 0.6000056266784668,
1702 4.200000286102295,
1703 6.000000953674316
1704 ],
1705 "dual_gap": 983.6130700111389,
1706 "price_excess_corr": -0.7151221632957458
1707 }
1708 ]
1709 }
1710 }
1711 ],
1712 "custom_track": {
1713 "name": "router_regime_regression",
1714 "file": "bench/custom_tracks/router_regime_regression.py",
1715 "domain": "moe-routing"
1716 }
1717 },
1718 "paired_raw_diffs": [
1719 0.010402381420135498,
1720 0.04728049039840698,
1721 0.08048295974731445,
1722 0.001484692096710205,
1723 0.008810937404632568,
1724 0.18900877237319946,
1725 0.111625075340271,
1726 0.16493749618530273
1727 ]
1728}