summaryrefslogtreecommitdiff
path: root/lib/parameters.py
blob: 4047c10678f3b36f714e103a1174da0e87925cf7 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
#!/usr/bin/env python3
import itertools
import logging
import numpy as np
import os
from collections import OrderedDict
from copy import deepcopy
from multiprocessing import Pool
import dfatool.functions as df
from .paramfit import ParamFit
from .utils import remove_indexes_from_tuple, is_numeric
from .utils import filter_aggregate_by_param, partition_by_param
from .utils import param_to_ndarray
from .utils import soft_cast_int, soft_cast_float

logger = logging.getLogger(__name__)

dfatool_fol_relevance_method = os.getenv("DFATOOL_FOL_RELEVANCE_METHOD", None)
dfatool_symreg_relevance_method = os.getenv("DFATOOL_SYMREG_RELEVANCE_METHOD", None)
dfatool_rmt_relevance_method = os.getenv("DFATOOL_RMT_RELEVANCE_METHOD", None)
dfatool_rmt_relevance_threshold = float(
    os.getenv("DFATOOL_RMT_RELEVANCE_THRESHOLD", "0.5")
)

if dfatool_rmt_relevance_method == "mi":
    import sklearn.feature_selection


def distinct_param_values(param_tuples):
    """
    Return the distinct values of each parameter in param_tuples.

    E.g. if param_tuples contains the distinct entries (1, 1), (1, 2), (1, 3), (0, 3),
    this function returns [[1, 0], [1, 2, 3]].

    Note that this function deliberately also consider None
    (uninitialized parameter with unknown value) as a distinct value. Benchmarks
    and drivers must ensure that a parameter is only None when its value is
    not important yet, e.g. a packet length parameter must only be None when
    write() or similar has not been called yet. Other parameters should always
    be initialized when leaving UNINITIALIZED.
    """
    if not len(param_tuples):
        logger.warning("distinct_param_values called with param_tuples=[]")
        return list()
    distinct_values = [OrderedDict() for i in range(len(param_tuples[0]))]
    for param_tuple in param_tuples:
        for i in range(len(param_tuple)):
            distinct_values[i][param_tuple[i]] = True

    # Convert sets to lists
    distinct_values = list(map(lambda x: list(x.keys()), distinct_values))
    return distinct_values


def _depends_on_param(corr_param, std_param, std_lut, threshold=0.5):
    if std_param == 0:
        # In general, std_param_lut < std_by_param. So, if std_by_param == 0, std_param_lut == 0 follows.
        # This means that the variation of param does not affect the model quality -> no influence
        # assert std_lut == 0
        return False
    return std_lut / std_param < threshold


def _mean_std_by_param(n_by_param, all_param_values, param_index):
    """
    Calculate the mean standard deviation for a static model where all parameters but `param_index` are constant.

    :param n_by_param: measurements of a specific model attribute partitioned by parameter values.
        Example: `{(0, 2): [2], (0, 4): [4], (0, 6): [6]}`
    :param all_param_values: distinct values of each parameter.
        E.g. for two parameters, the first being None, FOO, or BAR, and the second being 1, 2, 3, or 4, the argument is
        `[[None, 'FOO', 'BAR'], [1, 2, 3, 4]]`.
    :param param_index: index of variable parameter
    :returns: mean stddev
        *mean stddev* is the mean standard deviation of all measurements where parameter `param_index` is dynamic and all other parameters are fixed.
        E.g., if parameters are a, b, c ∈ {1,2,3} and 'index' corresponds to b, then
        this function returns the mean of the standard deviations of (a=1, b=*, c=1),
        (a=1, b=*, c=2), and so on.
    """
    return _mean_std_by_params(n_by_param, all_param_values, [param_index])


def _mean_std_by_params(n_by_param, all_param_values, param_indexes):
    """
    Calculate the mean standard deviation for a static model where all parameters but `param_indexes` are constant.

    :param n_by_param: measurements of a specific model attribute partitioned by parameter values.
        Example: `{(0, 2): [2], (0, 4): [4], (0, 6): [6]}`
    :param all_param_values: distinct values of each parameter.
        E.g. for two parameters, the first being None, FOO, or BAR, and the second being 1, 2, 3, or 4, the argument is
        `[[None, 'FOO', 'BAR'], [1, 2, 3, 4]]`.
    :param param_indexes: indexes of variable parameters
    :returns: mean stddev
        *mean stddev* is the mean standard deviation of all measurements where parameters `param_indexes` are dynamic and all other parameters are fixed.
        E.g., if parameters are a, b, c ∈ {1,2,3} and 'indexes' corresponds to b and c, then
        this function returns the mean of the standard deviations of (a=1, b=*, c=*),
        (a=2, b=*, c=*), and so on.
    """
    partition_by_tuple = dict()

    for k, v in n_by_param.items():
        tuple_key = remove_indexes_from_tuple(k, param_indexes)
        if not tuple_key in partition_by_tuple:
            partition_by_tuple[tuple_key] = list()
        partition_by_tuple[tuple_key].extend(v)

    if len(partition_by_tuple) == 0:
        return 0.0

    return np.mean([np.std(partition) for partition in partition_by_tuple.values()])


def _corr_by_param(attribute_data, param_values, param_index):
    """
    Return correlation coefficient (`np.corrcoef`) of `attribute_data` <-> `param_values[param_index]`

    A correlation coefficient close to 1 indicates that the attribute likely depends on the value of the parameter denoted by `param_index`, if it is nearly 0, it likely does not depend on it.

    If any value of `param_index` is not numeric (i.e., can not be parsed as float), this function returns 0.

    :param attribute_data: list or 1-D numpy array of measurements
    :param param_values: list of parameter values
    :param param_index: index of parameter in `by_name[*]['param']`
    """
    if _all_params_are_numeric(param_values, param_index):
        param_values = np.array(
            list((map(lambda x: float(x[param_index]), param_values)))
        )
        try:
            return np.corrcoef(attribute_data, param_values)[0, 1]
        except FloatingPointError:
            # Typically happens when all parameter values are identical.
            # Building a correlation coefficient is pointless in this case
            # -> assume no correlation
            return 0.0
        except (TypeError, ValueError) as e:
            logger.error(f"{e} in _corr_by_param(param_index={param_index})")
            logger.error(
                "while executing np.corrcoef({}, {}))".format(
                    attribute_data, param_values
                )
            )
            raise
    else:
        return 0.0


def _compute_param_statistics(
    data,
    param_names,
    param_tuples,
    arg_count=None,
    use_corrcoef=False,
    codependent_params=list(),
):
    """
    Compute standard deviation and correlation coefficient on parameterized data partitions.

    It is strongly recommended to vary all parameter values evenly.
    For instance, given two parameters, providing only the combinations
    (1, 1), (5, 1), (7, 1,) (10, 1), (1, 2), (1, 6) will lead to bogus results.
    It is better to provide (1, 1), (5, 1), (1, 2), (5, 2), ... (i.e. a cross product of all individual parameter values)

    arguments:
    data -- measurement data (ground truth). Must be a list or 1-D numpy array.
    param_names -- list of parameter names
    param_tuples -- list of parameter values corresponding to the order in param_names
    arg_count -- dict providing the number of functions args ("local parameters") for each function.
    use_corrcoef -- use correlation coefficient instead of stddev heuristic for parameter detection

    :returns: a dict with the following content:
    std_static -- static parameter-unaware model error: stddev of data
    std_param_lut -- static parameter-aware model error: mean stddev of data[*]
    std_by_param -- static parameter-aware model error ignoring a single parameter.
        dictionary with one key per parameter. The value is the mean stddev
        of measurements where all other parameters are fixed and the parameter
        in question is variable. E.g. std_by_param['X'] is the mean stddev of
        n_by_param[(X=*, Y=..., Z=...)].
    std_by_arg -- same, but ignoring a single function argument
        Only set if arg_count is non-zero, empty list otherwise.
    corr_by_param -- correlation coefficient
    corr_by_arg -- same, but ignoring a single function argument
        Only set if arg_count is non-zero, empty list otherwise.
    depends_on_param -- dict(parameter_name -> Bool). True if /attribute/ behaviour probably depends on /parameter_name/
    depends_on_arg -- list(bool). Same, but for function arguments, if any.
    """
    ret = dict()

    ret["by_param"] = by_param = partition_by_param(data, param_tuples)

    ret["use_corrcoef"] = use_corrcoef
    ret["_parameter_names"] = param_names

    ret["distinct_values_by_param_index"] = distinct_param_values(param_tuples)

    ret["distinct_values_by_param_name"] = dict()
    for i, param in enumerate(param_names):
        ret["distinct_values_by_param_name"][param] = ret[
            "distinct_values_by_param_index"
        ][i]

    ret["std_static"] = np.std(data)
    # TODO Gewichtung? Parameterkombinationen mit wenig verfügbaren Messdaten werden
    # genau so behandelt wie welchemit vielen verfügbaren Messdaten, in
    # std_static haben sie dagegen weniger Gewicht
    ret["std_param_lut"] = np.mean([np.std(v) for v in by_param.values()])

    ret["std_by_param"] = dict()

    ret["std_by_arg"] = list()

    ret["corr_by_param"] = dict()
    ret["corr_by_arg"] = list()

    ret["_depends_on_param"] = dict()
    ret["_depends_on_arg"] = list()

    np.seterr("raise")

    relevance_threshold = float(os.getenv("DFATOOL_PARAM_RELEVANCE_THRESHOLD", 0.5))

    for param_idx, param in enumerate(param_names):
        if param_idx < len(codependent_params) and codependent_params[param_idx]:
            by_param = partition_by_param(
                data, param_tuples, ignore_parameters=codependent_params[param_idx]
            )
            distinct_values = ret["distinct_values_by_param_index"].copy()
            for codependent_param_index in codependent_params[param_idx]:
                distinct_values[codependent_param_index] = [None]
        else:
            by_param = ret["by_param"]
            distinct_values = ret["distinct_values_by_param_index"]
        mean_std = _mean_std_by_param(by_param, distinct_values, param_idx)
        ret["std_by_param"][param] = mean_std
        ret["corr_by_param"][param] = _corr_by_param(data, param_tuples, param_idx)

        ret["_depends_on_param"][param] = _depends_on_param(
            ret["corr_by_param"][param],
            ret["std_by_param"][param],
            ret["std_param_lut"],
            relevance_threshold,
        )

    if arg_count:
        for arg_index in range(arg_count):
            param_idx = len(param_names) + arg_index
            if param_idx < len(codependent_params) and codependent_params[param_idx]:
                by_param = partition_by_param(
                    data, param_tuples, ignore_parameters=codependent_params[param_idx]
                )
                distinct_values = ret["distinct_values_by_param_index"].copy()
                for codependent_param_index in codependent_params[param_idx]:
                    distinct_values[codependent_param_index] = [None]
            else:
                by_param = ret["by_param"]
                distinct_values = ret["distinct_values_by_param_index"]
            mean_std = _mean_std_by_param(by_param, distinct_values, param_idx)
            ret["std_by_arg"].append(mean_std)
            ret["corr_by_arg"].append(_corr_by_param(data, param_tuples, param_idx))

            if False:
                ret["_depends_on_arg"].append(ret["corr_by_arg"][arg_index] > 0.1)
            elif ret["std_by_arg"][arg_index] == 0:
                # In general, std_param_lut < std_by_arg. So, if std_by_arg == 0, std_param_lut == 0 follows.
                # This means that the variation of arg does not affect the model quality -> no influence
                ret["_depends_on_arg"].append(False)
            else:
                ret["_depends_on_arg"].append(
                    ret["std_param_lut"] / ret["std_by_arg"][arg_index]
                    < relevance_threshold
                )

    return ret


def codependent_param_dict(param_values):
    """
    Detect pairs of codependent parameters in param_values.

    The parameter values are first normalized to integer values (e.g. 1, 7, 33 -> 0, 1, 2 and "foo", None, "Hello" -> 0, 1, 2).
    In essence, a pair of parameters (p1, p2) is codepenent if p2 changes only if p1 changes. This is calculated as follows:
    A pair of parameters (p1, p2) is codependent if, for each normalized value of p1, there is only one normalized value of p2 in the set of measurements with
    parameter 1 == p1. Essentially, this means that the mean standard deviation of parameter 2 values for each subset of measurements with a constant parameter
    1 value is zero.

    :param param_values: List of parameter values. Each list entry contains a list of parameter values for one measurement:
        ((param 1 value 1, param 2 value 1, ...), (param 1 value 2, param 2 value 2, ...), ...)
    :returns: dict of codependent parameter pairs. dict[(param 1 index, param 2 index)] is True iff param 1 and param 2 are codependent.
    """
    if not len(param_values):
        logger.warning("codependent_param_dict called with param_values=[]")
        return dict()
    if bool(int(os.getenv("DFATOOL_ULS_SKIP_CODEPENDENT_CHECK", 0))):
        return dict()
    lut = [dict() for i in param_values[0]]
    for param_index in range(len(param_values[0])):
        uniqs = set(map(lambda param_tuple: param_tuple[param_index], param_values))
        for uniq_index, uniq in enumerate(uniqs):
            lut[param_index][uniq] = uniq_index

    normed_param_values = list()
    for param_tuple in param_values:
        normed_param_values.append(
            tuple(map(lambda ipv: lut[ipv[0]][ipv[1]], enumerate(param_tuple)))
        )

    normed_param_values = np.array(normed_param_values)

    std_by_param = list()
    std_by_param_pair = dict()

    ret = dict()

    for param1_i in range(len(lut)):
        std_by_param.append(np.std(normed_param_values[:, param1_i]))
        for param2_i in range(param1_i + 1, len(lut)):
            stds = list()
            for param1_value in range(len(lut[param1_i])):
                tt = normed_param_values[:, param1_i] == param1_value
                values = normed_param_values[tt, param2_i]
                if len(values) <= 1:
                    stds.append(0.0)
                else:
                    stds.append(np.std(values))
            std_by_param_pair[(param1_i, param2_i)] = np.mean(stds)

    for param1_i in range(len(lut)):
        for param2_i in range(param1_i + 1, len(lut)):
            if std_by_param[param1_i] > 0 and std_by_param[param2_i] > 0:
                if std_by_param_pair[(param1_i, param2_i)] == 0:
                    ret[(param1_i, param2_i)] = True

    return ret


def _compute_param_statistics_parallel(arg):
    return {"key": arg["key"], "dict": _compute_param_statistics(*arg["args"])}


def _all_params_are_numeric(data, param_idx):
    """Check if all `data['param'][*][param_idx]` elements are numeric, as reported by `utils.is_numeric`."""
    param_values = list(map(lambda x: x[param_idx], data))
    return all(map(is_numeric, param_values))


class ParamType(dict):
    UNSET = 0
    USELESS = 1
    BOOLEAN = 2
    SCALAR = 3
    ENUM = 4

    def __init__(self, param_values, values_are_distinct=False):
        if values_are_distinct:
            distinct_values = param_values
        else:
            distinct_values = distinct_param_values(param_values)
        for param_index, param_values in enumerate(distinct_values):
            if None in param_values:
                none_adj = -1
            else:
                none_adj = 0
            value_count = len(param_values) + none_adj
            if value_count == 0:
                self[param_index] = self.UNSET
            elif value_count == 1:
                self[param_index] = self.USELESS
            elif value_count == 2:
                self[param_index] = self.BOOLEAN
            elif all(map(lambda n: n is None or is_numeric(n), param_values)):
                self[param_index] = self.SCALAR
            else:
                self[param_index] = self.ENUM


class ParallelParamStats:
    def __init__(self):
        self.queue = list()
        self.map = dict()

    def enqueue(self, key, attr):
        """
        Enqueue data series for statistics calculation.

        :param key: entry key used for retrieval. attr is stored in self.map[key]
            and extended with "by_param" and "stats" attributes once compute() has been called.
        :param attr: ModelAttribute instance. Edited in-place by compute()
        """
        self.queue.append(
            {
                "key": key,
                "args": [
                    attr.data,
                    attr.param_names,
                    attr.param_values,
                    attr.arg_count,
                    False,
                    attr.codependent_params,
                ],
            }
        )
        self.map[key] = attr

    def compute(self):
        """
        Compute statistics for previously enqueue ModelAttribute data.

        Statistics are computed in parallel with one process per core. Results are written to each ModelAttribute wich was passed via enqueue().
        """
        logger.debug("Computing param stats in parallel")
        with Pool() as pool:
            results = pool.map(_compute_param_statistics_parallel, self.queue)

        for result in results:
            self.map[result["key"]].by_param = result["dict"].pop("by_param")
            self.map[result["key"]].stats = ParamStats(result["dict"])


class ParamStats:
    """
    Statistics.

    :var std_static: static parameter-unaware model error (standard deviation of raw data)
    :var std_param_lut:  static parameter-aware model error (mean standard deviation of data partitioned by parameter)
    :var std_by_param: static parameter-aware model error ignoring a single parameter.
        dict mapping parameter name -> mean std of data partitioned so that all parameters but "parameter name" are constant.
    :var sty_by_arg: list, argument index -> mean std of data partition so that all parameters but "argument index" are constant
    :var _depends_on_param: dict, parameter name -> bool, True if the modeled behaviour likely depends on parameter name
    :var _depends_on_arg: list, argument index -> bool, True if the modeled behaviour likely depends on argument index
    """

    def __init__(self, data):
        self.__dict__.update(data)

    @classmethod
    def compute_for_attr(cls, attr, use_corrcoef=False):
        res = _compute_param_statistics(
            attr.data,
            attr.param_names,
            attr.param_values,
            arg_count=attr.arg_count,
            use_corrcoef=use_corrcoef,
            codependent_params=attr.codependent_params,
        )
        attr.by_param = res.pop("by_param")
        attr.stats = cls(res)

    def _generic_param_independence_ratio(self):
        """
        Return the heuristic ratio of parameter independence.

        This is not supported if the correlation coefficient is used.
        A value close to 1 means no influence, a value close to 0 means high probability of influence.
        """
        if self.use_corrcoef:
            # not supported
            raise ValueError
        if self.std_static == 0:
            return 0
        return self.std_param_lut / self.std_static

    def generic_param_dependence_ratio(self):
        """
        Return the heuristic ratio of parameter dependence.

        This is not supported if the correlation coefficient is used.
        A value close to 0 means no influence, a value close to 1 means high probability of influence.
        """
        return 1 - self._generic_param_independence_ratio()

    def _param_independence_ratio(self, param: str) -> float:
        """
        Return the heuristic ratio of parameter independence for param.

        A value close to 1 means no influence, a value close to 0 means high probability of influence.
        """
        if self.use_corrcoef:
            return 1 - np.abs(self.corr_by_param[param])
        if self.std_by_param[param] == 0:
            # if self.std_param_lut != 0:
            #    raise RuntimeError(f"wat: std_by_param[{param}]==0, but std_param_lut=={self.std_param_lut} ≠ 0")
            # In general, std_param_lut < std_by_param. So, if std_by_param == 0, std_param_lut == 0 follows.
            # This means that the variation of param does not affect the model quality -> no influence, return 1
            return 1.0

        return self.std_param_lut / self.std_by_param[param]

    def param_dependence_ratio(self, param: str) -> float:
        """
        Return the heuristic ratio of parameter dependence for param.

        A value close to 0 means no influence, a value close to 1 means high probability of influence.

        :param param: parameter name

        :returns: parameter dependence (float between 0 == no influence and 1 == high probability of influence)
        """
        return 1 - self._param_independence_ratio(param)

    def _arg_independence_ratio(self, arg_index):
        if self.use_corrcoef:
            return 1 - np.abs(self.corr_by_arg[arg_index])
        if self.std_by_arg[arg_index] == 0:
            if self.std_param_lut != 0:
                raise RuntimeError(
                    f"wat: std_by_arg[{arg_index}]==0, but std_param_lut=={self.std_param_lut} ≠ 0"
                )
            # In general, std_param_lut < std_by_arg. So, if std_by_arg == 0, std_param_lut == 0 follows.
            # This means that the variation of arg does not affect the model quality -> no influence, return 1
            return 1
        return self.std_param_lut / self.std_by_arg[arg_index]

    def arg_dependence_ratio(self, arg_index: int) -> float:
        return 1 - self._arg_independence_ratio(arg_index)

    # This heuristic is very similar to the "function is not much better than
    # median" checks in get_fitted. So far, doing it here as well is mostly
    # a performance and not an algorithm quality decision.
    # --df, 2018-04-18
    def depends_on_param(self, param):
        """Return whether attribute of state_or_trans depens on param."""
        return self._depends_on_param[param]

    # See notes on depends_on_param
    def depends_on_arg(self, arg_index):
        """Return whether attribute of state_or_trans depens on arg_index."""
        return self._depends_on_arg[arg_index]


class ModelAttribute:
    """
    A ModelAttribute instance handles a single model attribute, e.g. TX state power or something() function call duration, and corresponding benchmark data.

    It provides three models:
    - a static model (`mean`, `median`) as lower bound of model accuracy
    - a LUT model (`by_param`) as upper bound of model accuracy
    - a fitted model (`model_function`, a `ModelFunction` instance)
    """

    def __init__(
        self,
        name,
        attr,
        data,
        param_values,
        param_names,
        arg_count=0,
        codependent_param=dict(),
        param_type=dict(),
    ):
        # Data for model generation
        self.data = np.array(data)

        # Meta data
        self.name = name
        self.attr = attr
        self.param_values = param_values
        self.param_names = sorted(param_names)
        self.arg_count = arg_count

        self.log_param_names = self.param_names + list(
            map(lambda i: f"arg{i}", range(arg_count))
        )

        # dict: Parameter index -> Parameter type (UNSET, BOOLEAN, SCALAR, ...)
        self.param_type = param_type

        self.nonscalar_param_indexes = list(
            map(
                lambda kv: kv[0],
                filter(lambda kv: kv[1] != ParamType.SCALAR, self.param_type.items()),
            )
        )
        self.scalar_param_indexes = list(
            map(
                lambda kv: kv[0],
                filter(lambda kv: kv[1] == ParamType.SCALAR, self.param_type.items()),
            )
        )

        # Co-dependent parameters. If (param1_index, param2_index) in codependent_param, they are codependent.
        # In this case, only one of them must be used for parameter-dependent model attribute detection and modeling
        self.codependent_param_pair = codependent_param
        self.codependent_params = [list() for x in self.log_param_names]
        self.ignore_codependent_param = dict()

        # Static model used as lower bound of model accuracy
        if data is not None:
            self.mean = np.mean(data)
            self.median = np.median(data)
        else:
            self.mean = None
            self.median = None

        # LUT model used as upper bound of model accuracy
        self.by_param = None  # set via ParallelParamStats or get_by_param

        self.stats = None  # set via ParallelParamStats

        # param model override
        self.function_override = None

        # The best model we have. May be Static, Split, or Param (and later perhaps Substate)
        self.model_function = None

        self._check_codependent_param()

        # There must be at least 3 distinct data values (≠ None) if an analytic model
        # is to be fitted. For 2 (or fewer) values, decision trees are better.
        # Exceptions such as DFATOOL_RMT_SUBMODEL=fol (2 values sufficient)
        # can be handled via DFATOOL_ULS_MIN_DISTINCT_VALUES
        self.min_values_for_analytic_model = int(
            os.getenv("DFATOOL_ULS_MIN_DISTINCT_VALUES", "3")
        )

    def __repr__(self):
        mean = np.mean(self.data)
        return f"ModelAttribute<{self.name}, {self.attr}, mean={mean}>"

    def to_json(self, **kwargs):
        return {
            "paramNames": self.param_names,
            "argCount": self.arg_count,
            "modelFunction": self.model_function.to_json(**kwargs),
        }

    def to_dref(self, unit=None):
        ret = {"mean": (self.mean, unit), "median": (self.median, unit)}

        if issubclass(type(self.model_function), df.ModelFunction):
            ret["model/complexity"] = self.model_function.get_complexity_score()
        if self.by_param:
            ret["lut/complexity"] = len(self.by_param.keys()) + 1

        if type(self.model_function) in (
            df.SplitFunction,
            df.CARTFunction,
            df.XGBoostFunction,
            df.LMTFunction,
        ):
            ret["decision tree/nodes"] = self.model_function.get_number_of_nodes()
            ret["decision tree/leaves"] = self.model_function.get_number_of_leaves()
            ret["decision tree/inner nodes"] = (
                ret["decision tree/nodes"] - ret["decision tree/leaves"]
            )
            ret["decision tree/max depth"] = self.model_function.get_max_depth()
        elif type(self.model_function) in (df.StaticFunction, df.AnalyticFunction):
            ret["decision tree/nodes"] = 1
            ret["decision tree/leaves"] = 1
            ret["decision tree/inner nodes"] = 0
            ret["decision tree/max depth"] = 0

        return ret

    def to_dot(self):
        if type(self.model_function) in (
            df.SplitFunction,
            df.StaticFunction,
            df.AnalyticFunction,
            df.FOLFunction,
        ):
            import pydot

            graph = pydot.Dot("Regression Model Tree", graph_type="graph")
            self.model_function.to_dot(pydot, graph, self.param_names)
            return graph

        if type(self.model_function) == df.CARTFunction:
            import sklearn.tree

            return sklearn.tree.export_graphviz(
                self.model_function.regressor,
                out_file=None,
                feature_names=self.model_function.feature_names,
            )
        if type(self.model_function) == df.XGBoostFunction:
            import xgboost

            self.model_function.regressor.get_booster().feature_names = (
                self.model_function.feature_names
            )
            return [
                xgboost.to_graphviz(self.model_function.regressor, num_trees=i)
                for i in range(self.model_function.regressor.n_estimators)
            ]
        if type(self.model_function) == df.LMTFunction:
            return self.model_function.regressor.model_to_dot(
                feature_names=self.model_function.feature_names
            )
        return None

    def min(self):
        return np.min(self.data)

    def max(self):
        return np.max(self.data)

    def webconf_function_map(self):
        return self.model_function.webconf_function_map()

    @classmethod
    def from_json(cls, name, attr, data):
        param_names = data["paramNames"]
        arg_count = data["argCount"]

        self = cls(name, attr, None, None, param_names, arg_count)

        self.model_function = df.ModelFunction.from_json(data["modelFunction"])
        self.mean = self.model_function.value
        self.median = self.model_function.value

        return self

    def _check_codependent_param(self):
        for (
            (param1_index, param2_index),
            is_codependent,
        ) in self.codependent_param_pair.items():
            if not is_codependent:
                continue
            param1_values = map(lambda pv: pv[param1_index], self.param_values)
            param1_numeric_count = sum(map(is_numeric, param1_values))
            param2_values = map(lambda pv: pv[param2_index], self.param_values)
            param2_numeric_count = sum(map(is_numeric, param2_values))
            # If all occurences of (param1, param2) are either (None, None) or (not None, not None), removing one of them is sensible.
            # Otherwise, one parameter may decide whether the other one has an effect or not (or what kind of effect it has). This is important for
            # decision-tree models, so do not remove parameters in that case.
            params_are_pairwise_none = all(
                map(
                    lambda pv: not (
                        (pv[param1_index] is None) ^ (pv[param2_index] is None)
                    ),
                    self.param_values,
                )
            )
            if (
                param1_numeric_count >= param2_numeric_count
                and params_are_pairwise_none
            ):
                self.ignore_codependent_param[param2_index] = True
                self.codependent_params[param1_index].append(param2_index)
                logger.debug(
                    f"{self.name} {self.attr}: parameters ({self.log_param_names[param1_index]}, {self.log_param_names[param2_index]}) are codependent. Ignoring {self.log_param_names[param2_index]}"
                )
            elif (
                param2_numeric_count >= param1_numeric_count
                and params_are_pairwise_none
            ):
                self.ignore_codependent_param[param1_index] = True
                self.codependent_params[param2_index].append(param1_index)
                logger.debug(
                    f"{self.name} {self.attr}: parameters ({self.log_param_names[param1_index]}, {self.log_param_names[param2_index]}) are codependent. Ignoring {self.log_param_names[param1_index]}"
                )

    def get_static(self, use_mean=False):
        if use_mean:
            return self.mean
        return self.median

    def get_lut(self, param, use_mean=False):
        if use_mean:
            return np.mean(self.by_param[param])
        return np.median(self.by_param[param])

    def get_by_param(self):
        if self.by_param is None and self.param_values is not None:
            self.by_param = partition_by_param(self.data, self.param_values)
        return self.by_param

    def get_data_for_paramfit(self, safe_functions_enabled=False):
        ret = list()
        for param_index, param_name in enumerate(self.param_names):
            if (
                self.stats.depends_on_param(param_name)
                and not param_index in self.ignore_codependent_param
            ):
                by_param = self._by_param_for_index(param_index)
                ret.append(
                    (
                        (self.name, self.attr),
                        param_name,
                        (by_param, param_index, safe_functions_enabled),
                        dict(),
                    )
                )
        if self.arg_count:
            for arg_index in range(self.arg_count):
                param_index = len(self.param_names) + arg_index
                if (
                    self.stats.depends_on_arg(arg_index)
                    and not param_index in self.ignore_codependent_param
                ):
                    by_param = self._by_param_for_index(param_index)
                    ret.append(
                        (
                            (self.name, self.attr),
                            arg_index,
                            (by_param, param_index, safe_functions_enabled),
                            dict(),
                        )
                    )

        return ret

    def _by_param_for_index(self, param_index):
        if not self.codependent_params[param_index]:
            return self.by_param
        new_param_values = list()
        for param_tuple in self.param_values:
            new_param_tuple = param_tuple.copy()
            for i in self.codependent_params[param_index]:
                new_param_tuple[i] = None
            new_param_values.append(new_param_tuple)
        return partition_by_param(self.data, new_param_values)

    def depends_on_any_param(self):
        for param_index, param_name in enumerate(self.param_names):
            if (
                self.stats.depends_on_param(param_name)
                and not param_index in self.ignore_codependent_param
            ):
                return True
        return False

    def all_relevant_parameters_are_none_or_numeric(self):
        for param_index, param_name in enumerate(self.param_names):
            if (
                self.stats.depends_on_param(param_name)
                and not param_index in self.ignore_codependent_param
            ):
                param_values = list(map(lambda x: x[param_index], self.param_values))
                if not all(map(lambda n: n is None or is_numeric(n), param_values)):
                    return False
                distinct_values = self.stats.distinct_values_by_param_index[param_index]
                if (
                    None in distinct_values
                    and len(distinct_values) - 1 < self.min_values_for_analytic_model
                ) or len(distinct_values) < self.min_values_for_analytic_model:
                    return False
        return True

    def fit_override_function(self):
        function_str = self.function_override
        x = df.AnalyticFunction(
            self.median,
            function_str,
            self.param_names,
            self.arg_count,
            n_samples=self.data.shape[0],
            # fit_by_param=fit_result,
        )
        x.fit(self.by_param)
        if x.fit_success:
            self.model_function = x
        else:
            logger.warning(f"Fit of user-defined model function {function_str} failed.")

    def set_data_from_paramfit(self, paramfit, prefix=tuple()):
        fit_result = paramfit.get_result((self.name, self.attr) + prefix)
        if self.model_function is None:
            self.model_function = df.StaticFunction(
                self.median, n_samples=self.data.shape[0]
            )
        if os.getenv("DFATOOL_NO_PARAM"):
            pass
        elif len(fit_result.keys()):
            x = df.analytic.function_powerset(
                fit_result,
                self.param_names,
                self.arg_count,
                n_samples=self.data.shape[0],
            )
            x.value = self.median
            x.fit(self.by_param)

            if x.fit_success:
                self.model_function = x

    def build_cart(self):
        mf = df.CARTFunction(
            np.mean(self.data),
            n_samples=len(self.data),
            param_names=self.param_names,
            arg_count=self.arg_count,
        ).fit(
            self.param_values,
            self.data,
        )

        if mf.fit_success:
            self.model_function = mf
            return True
        else:
            logger.warning(f"{self.name}:{self.attr}:CART failed")
            self.model_function = df.StaticFunction(
                np.mean(self.data), n_samples=len(self.data)
            )
            return False

    def build_decart(self):
        mf = df.CARTFunction(
            np.mean(self.data),
            n_samples=len(self.data),
            param_names=self.param_names,
            arg_count=self.arg_count,
            decart=True,
        ).fit(
            self.param_values,
            self.data,
            scalar_param_indexes=self.scalar_param_indexes,
        )

        if mf.fit_success:
            self.model_function = mf
            return True
        else:
            logger.warning(f"{self.name}:{self.attr}:DECART failed")
            self.model_function = df.StaticFunction(
                np.mean(self.data), n_samples=len(self.data)
            )
            return False

    def build_fol(self):
        ignore_param_indexes = list()
        if dfatool_fol_relevance_method == "std_by_param":
            for param_index, param in enumerate(self.param_names):
                if not self.stats.depends_on_param(param):
                    ignore_param_indexes.append(param_index)
        if not self.stats:
            logger.warning(
                "build_fol_model called with ModelAttribute.stats unavailable -- overfitting likely"
            )
        else:
            for param_index, _ in enumerate(self.param_names):
                if len(self.stats.distinct_values_by_param_index[param_index]) < 2:
                    ignore_param_indexes.append(param_index)
        x = df.FOLFunction(
            np.mean(self.data),
            n_samples=self.data.shape[0],
            param_names=self.param_names,
            arg_count=self.arg_count,
        )
        x.fit(self.param_values, self.data, ignore_param_indexes=ignore_param_indexes)
        if x.fit_success:
            self.model_function = x
            return True
        else:
            logger.warning(f"{self.name}:{self.attr}:FOL failed")
            self.model_function = df.StaticFunction(
                np.mean(self.data), n_samples=len(self.data)
            )
            return False

    def build_lmt(self):
        mf = df.LMTFunction(
            np.mean(self.data),
            n_samples=len(self.data),
            param_names=self.param_names,
            arg_count=self.arg_count,
        ).fit(self.param_values, self.data)

        if mf.fit_success:
            self.model_function = mf
            return True
        else:
            logger.warning(f"{self.name}:{self.attr}:LMT failed")
            self.model_function = df.StaticFunction(
                np.mean(self.data), n_samples=len(self.data)
            )
            return False

    def build_symreg(self):
        ignore_param_indexes = list()
        if dfatool_symreg_relevance_method == "std_by_param":
            for param_index, param in enumerate(self.param_names):
                if not self.stats.depends_on_param(param):
                    ignore_param_indexes.append(param_index)
        x = df.SymbolicRegressionFunction(
            np.mean(self.data),
            n_samples=self.data.shape[0],
            param_names=self.param_names,
            arg_count=self.arg_count,
        ).fit(self.param_values, self.data, ignore_param_indexes=ignore_param_indexes)
        if x.fit_success:
            self.model_function = x
            return True
        else:
            logger.debug(
                f"Symbolic Regression model generation for {self.name} {self.attr} failed."
            )
            self.model_function = df.StaticFunction(
                np.mean(self.data), n_samples=len(self.data)
            )
            return False

    def build_lgbm(self):
        mf = df.LightGBMFunction(
            np.mean(self.data),
            n_samples=len(self.data),
            param_names=self.param_names,
            arg_count=self.arg_count,
        ).fit(self.param_values, self.data)

        if mf.fit_success:
            self.model_function = mf
            return True
        else:
            logger.warning(f"{self.name}:{self.attr}:LightGBM failed")
            self.model_function = df.StaticFunction(
                np.mean(self.data), n_samples=len(self.data)
            )
            return False

    def build_xgb(self):
        mf = df.XGBoostFunction(
            np.mean(self.data),
            n_samples=len(self.data),
            param_names=self.param_names,
            arg_count=self.arg_count,
        ).fit(self.param_values, self.data)

        if mf.fit_success:
            self.model_function = mf
            return True
        else:
            logger.warning(f"{self.name}:{self.attr}:XGBoost failed")
            self.model_function = df.StaticFunction(
                np.mean(self.data), n_samples=len(self.data)
            )
            return False

    def build_rmt(
        self,
        with_function_leaves=None,
        with_nonbinary_nodes=None,
        with_gplearn_symreg=None,
        loss_ignore_scalar=None,
        threshold=100,
    ):
        """
        Build a Decision Tree on `param` / `data` for kconfig models.

        :param parameters: parameter values for each measurement. [(data 1 param 1, data 1 param 2, ...), (data 2 param 1, data 2 param 2, ...), ...]
        :param data: Measurements. [data 1, data 2, data 3, ...]
        :param with_function_leaves: Use fitted function sets to generate function leaves for scalar parameters
        :param with_nonbinary_nodes: Allow non-binary nodes for enum and scalar parameters (i.e., nodes with more than two children)
        :param loss_ignore_scalar: Ignore scalar parameters when computing the loss for split candidates. Only sensible if with_function_leaves is enabled.
        :param threshold: Return a StaticFunction leaf node if std(data) < threshold. Default 100.

        :returns: SplitFunction or StaticFunction
        """

        if with_function_leaves is None:
            if os.getenv("DFATOOL_RMT_SUBMODEL", "uls") == "static":
                with_function_leaves = False
            else:
                with_function_leaves = True
        if with_nonbinary_nodes is None:
            with_nonbinary_nodes = bool(
                int(os.getenv("DFATOOL_RMT_NONBINARY_NODES", "1"))
            )
        if with_gplearn_symreg is None:
            with_gplearn_symreg = bool(int(os.getenv("DFATOOL_USE_SYMREG", "0")))
        if loss_ignore_scalar is None:
            loss_ignore_scalar = bool(
                int(os.getenv("DFATOOL_RMT_LOSS_IGNORE_SCALAR", "0"))
            )

        if loss_ignore_scalar and not with_function_leaves:
            logger.warning(
                "build_rmt {self.name} {self.attr} called with loss_ignore_scalar=True, with_function_leaves=False. This does not make sense."
            )

        max_depth = int(os.getenv("DFATOOL_RMT_MAX_DEPTH", "0"))

        logger.debug(f"build_rmt(threshold={threshold}, max_depth={max_depth})")

        self.model_function = self._build_rmt(
            self.param_values,
            self.data,
            with_function_leaves=with_function_leaves,
            with_nonbinary_nodes=with_nonbinary_nodes,
            loss_ignore_scalar=loss_ignore_scalar,
            submodel=os.getenv("DFATOOL_RMT_SUBMODEL", "uls"),
            max_depth=max_depth,
            threshold=threshold,
        )

    def _build_rmt(
        self,
        parameters,
        data,
        with_function_leaves=False,
        with_nonbinary_nodes=True,
        loss_ignore_scalar=False,
        submodel="uls",
        threshold=100,
        max_depth=0,
        level=0,
    ):
        """
        Build a Decision Tree on `param` / `data` for kconfig models.

        :param parameters: parameter values for each measurement. [(data 1 param 1, data 1 param 2, ...), (data 2 param 1, data 2 param 2, ...), ...]
        :param data: Measurements. [data 1, data 2, data 3, ...]
        :param with_function_leaves: Use fitted function sets to generate function leaves for scalar parameters
        :param with_nonbinary_nodes: Allow non-binary nodes for enum and scalar parameters (i.e., nodes with more than two children)
        :param loss_ignore_scalar: Ignore scalar parameters when computing the loss for split candidates. Only sensible if with_function_leaves is enabled.
        :param threshold: Return a StaticFunction leaf node if std(data) < threshold. Default 100.

        :returns: ModelFunction
        """

        nonarg_count = len(self.param_names)
        param_count = nonarg_count + self.arg_count
        # TODO eigentlich muss threshold hier auf Basis der aktuellen Messdatenpartition neu berechnet werden
        if param_count == 0 or np.std(data) <= threshold:
            return df.StaticFunction(np.mean(data), n_samples=len(data))
            # sf.value_error["std"] = np.std(data)

        loss = list()

        ffs_feasible = False
        if dfatool_rmt_relevance_method:
            irrelevant_params = list()
            if dfatool_rmt_relevance_method == "std_by_param":
                by_param = partition_by_param(data, parameters)
                distinct_values_by_param_index = distinct_param_values(parameters)
                std_lut = np.mean([np.std(v) for v in by_param.values()])
            elif dfatool_rmt_relevance_method == "mi":
                fit_parameters, _, ignore_index = param_to_ndarray(
                    parameters, with_nan=False, categorical_to_scalar=True
                )
                param_to_fit_param = dict()
                j = 0
                for i in range(param_count):
                    if not ignore_index[i]:
                        param_to_fit_param[i] = j
                        j += 1
                try:
                    mutual_information = (
                        sklearn.feature_selection.mutual_info_regression(
                            fit_parameters, data
                        )
                    )
                except ValueError as e:
                    logger.debug(f"mutual_info_regression failed: {e}")
                    mutual_information = [np.inf for _ in range(len(fit_parameters[0]))]

        if loss_ignore_scalar:
            ffs_eligible_params = list()
            ffs_unsuitable_params = list()
            for param_index in range(param_count):
                if param_index in self.ignore_codependent_param:
                    continue
                unique_values = list(set(map(lambda p: p[param_index], parameters)))
                if None in unique_values:
                    ffs_unsuitable_params.append(param_index)
                elif (
                    self.param_type[param_index] == ParamType.SCALAR
                    and len(unique_values) >= self.min_values_for_analytic_model
                ):
                    ffs_eligible_params.append(param_index)
                else:
                    ffs_unsuitable_params.append(param_index)

        for param_index in range(param_count):
            if (
                param_index >= nonarg_count
                and self.param_type[param_index] == ParamType.ENUM
            ):
                # do not split on non-numeric function arguments
                loss.append(np.inf)
                continue

            unique_values = list(set(map(lambda p: p[param_index], parameters)))

            if None in unique_values:
                # param is a choice and undefined in some configs. Do not split on it.
                loss.append(np.inf)
                continue

            if (
                with_function_leaves
                and self.param_type[param_index] == ParamType.SCALAR
                and len(unique_values) >= self.min_values_for_analytic_model
            ):
                # param can be modeled as a function. Do not split on it.
                loss.append(np.inf)
                ffs_feasible = True
                continue

            # if not with_nonbinary_nodes and sorted(unique_values) != [0, 1]:
            if not with_nonbinary_nodes and len(unique_values) > 2:
                # param cannot be handled with a binary split
                loss.append(np.inf)
                continue

            if dfatool_rmt_relevance_method == "std_by_param":
                std_by_param = _mean_std_by_params(
                    by_param,
                    distinct_values_by_param_index,
                    list(self.ignore_codependent_param.keys())
                    + irrelevant_params
                    + [param_index],
                )
                if not _depends_on_param(
                    None, std_by_param, std_lut, dfatool_rmt_relevance_threshold
                ):
                    irrelevant_params.append(param_index)
                    loss.append(np.inf)
                    continue
            elif dfatool_rmt_relevance_method == "mi":
                # The 'None in unique_values' check above ensures that param_to_fit_param[param_index] exists here
                if (
                    mutual_information[param_to_fit_param[param_index]]
                    < dfatool_rmt_relevance_threshold
                ):
                    loss.append(np.inf)
                    continue

            child_indexes = list()
            for value in unique_values:
                child_indexes.append(
                    list(
                        filter(
                            lambda i: parameters[i][param_index] == value,
                            range(len(parameters)),
                        )
                    )
                )
                assert len(child_indexes[-1]) > 0

            assert len(child_indexes) != 0

            if len(child_indexes) == 1:
                # this param only has a single value. there's no point in splitting.
                loss.append(np.inf)
                continue

            children = list()
            for child in child_indexes:
                child_data = list(map(lambda i: data[i], child))
                if loss_ignore_scalar and False:
                    child_param = list(map(lambda i: parameters[i], child))
                    child_data_by_scalar = partition_by_param(
                        child_data,
                        child_param,
                        ignore_parameters=list(self.ignore_codependent_param.keys())
                        + ffs_unsuitable_params,
                    )
                    logger.debug(f"got {len(child_data_by_scalar)} partitions")
                    for sub_data in child_data_by_scalar.values():
                        assert len(sub_data)
                        children.extend((np.array(sub_data) - np.mean(sub_data)) ** 2)
                else:
                    children.extend((np.array(child_data) - np.mean(child_data)) ** 2)

            assert not np.any(np.isnan(children))
            loss.append(np.sum(children))

        if (
            np.all(np.isinf(loss))
            or (max_depth and level >= max_depth)
            or np.min(loss) >= np.sum((np.array(data) - np.mean(data)) ** 2)
        ):
            if ffs_feasible:
                # try generating a function. if it fails, model_function is a StaticFunction.
                ma = ModelAttribute(
                    self.name + "_",
                    self.attr,
                    data,
                    parameters,
                    self.param_names,
                    arg_count=self.arg_count,
                    param_type=self.param_type,
                    codependent_param=codependent_param_dict(parameters),
                )
                if submodel == "cart":
                    if ma.build_cart():
                        return ma.model_function
                elif submodel == "symreg":
                    if ma.build_symreg():
                        return ma.model_function
                else:
                    ParamStats.compute_for_attr(ma)
                    paramfit = ParamFit(parallel=False)
                    for key, param, args, kwargs in ma.get_data_for_paramfit():
                        paramfit.enqueue(key, param, args, kwargs)
                    paramfit.fit()
                    ma.set_data_from_paramfit(paramfit)
                    return ma.model_function
            return df.StaticFunction(np.mean(data), n_samples=len(data))

        split_feasible = True
        if loss_ignore_scalar:
            data_by_scalar = partition_by_param(
                data,
                parameters,
                ignore_parameters=list(self.ignore_codependent_param.keys())
                + ffs_unsuitable_params,
            )
            if np.all(
                np.array([np.std(partition) for partition in data_by_scalar.values()])
                <= threshold
            ):
                # Varying non-scalar params in partitions with fixed scalar params does not affect system behaviour
                # -> further non-scalar splits are _probably_ not sensible
                # (_probably_ because this implicitly assumes that there are multiple scalar configurations for each non-scalar configuration.
                split_feasible = False

        if ffs_feasible and not split_feasible:
            # There is a _probably_ above: the heuristic assumes that there are multiple scalar configurations for each non-scalar configuration.
            # If there is just one it may recommend to stop splitting too early.
            # Hence, we will try generating an FFS leaf node here, but continue splitting if it turns out that it is no good.
            ma = ModelAttribute(
                self.name + "_",
                self.attr,
                data,
                parameters,
                self.param_names,
                arg_count=self.arg_count,
                param_type=self.param_type,
                codependent_param=codependent_param_dict(parameters),
            )
            if submodel == "cart":
                if ma.build_cart():
                    return ma.model_function
            elif submodel == "symreg":
                if ma.build_symreg():
                    return ma.model_function
            else:
                ParamStats.compute_for_attr(ma)
                paramfit = ParamFit(parallel=False)
                for key, param, args, kwargs in ma.get_data_for_paramfit():
                    paramfit.enqueue(key, param, args, kwargs)
                paramfit.fit()
                ma.set_data_from_paramfit(paramfit)
                if type(ma.model_function) == df.AnalyticFunction:
                    return ma.model_function

        symbol_index = np.argmin(loss)
        unique_values = list(set(map(lambda p: p[symbol_index], parameters)))

        child = dict()

        for value in unique_values:
            indexes = list(
                filter(
                    lambda i: parameters[i][symbol_index] == value,
                    range(len(parameters)),
                )
            )
            child_parameters = list(map(lambda i: parameters[i], indexes))
            child_data = list(map(lambda i: data[i], indexes))
            assert len(child_data)
            child[value] = self._build_rmt(
                child_parameters,
                child_data,
                with_function_leaves=with_function_leaves,
                with_nonbinary_nodes=with_nonbinary_nodes,
                loss_ignore_scalar=loss_ignore_scalar,
                submodel=submodel,
                threshold=threshold,
                max_depth=max_depth,
                level=level + 1,
            )

        assert len(child.values()) >= 2

        return df.SplitFunction(
            np.mean(data),
            symbol_index,
            self.log_param_names[symbol_index],
            child,
            n_samples=len(data),
        )