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 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639
|
\input texinfo
@c %**start of header
@setfilename optim.info
@settitle optim_doc
@c %**end of header
@c Nowadays the predined function index has entries for each @deftypefn
@c in additiont to each @findex. This would hopelessly clutter the index
@c here.
@defcodeindex mfn
@ifinfo
@dircategory Math
@direntry
* Octave-optim: (octave-optim). Non-linear optimization toolkit for Octave.
@end direntry
@end ifinfo
@copying
Additional documentation for the optim package for Octave.
Copyright @copyright{} @email{Olaf Till <i7tiol@@t-online.de>}
You can redistribute this documentation and/or modify it under the terms
of the GNU General Public License as published by the Free Software
Foundation; either version 3 of the License, or (at your option) any
later version.
This documentation is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General
Public License for more details.
You should have received a copy of the GNU General Public License along
with this documentation; if not, see <http://www.gnu.org/licenses/>.
@end copying
@include macros.texi
@macro mysee
@ifhtml
see
@end ifhtml
@end macro
@titlepage
@title Additional documentation for the optim package for Octave
@page
@vskip 0pt plus 1filll
@insertcopying
@end titlepage
@c No table of contents. The table would occupy most of the top node in
@c html and IMHO misleads the user to use the table instead of the menu
@c structure of the nodes, which would let some information unused.
@c
@c @contents
@c ------------------------------------------------------------------
@node Top
@top Additional documentation for the optim package for Octave
@ifhtml
The info version of this document is accessible, after package
installation, from the Octave commandline with @code{optim_doc()}.
@end ifhtml
This documentation applies to version @PACKAGEVERSION of the optim
package.
The optim package is a collection of additional functions related to
numerical optimization. For Octaves core optimization functions (not
contained in this package)
@mysee
@ref{Optimization,,,octave}.
@menu
Types of functions in the optim package
* Scalar optimization:: Functions for optimization of a scalar
objective function.
* Residual optimization:: Functions for optimization of a model
function returning an array.
* Zero finders:: Functions for finding the zero of a
scalar or array valued nonlinear user
function.
* Gradient functions:: Functions for numerical approximation of
gradients and Hessians.
* Helper functions:: Functions for algebraic tasks common to
optimization problems.
* Documentation:: Function optim_doc to view documentation.
* Compatibility wrappers:: Functions with traditional names and
arguments which work by calling a
different function.
Configuration
* Common frontend options:: Options common to all frontends.
* Common optimization options:: Options common to all optimization
frontends.
* Parameter structures:: Handling of structures of optimized
parameters.
* Additional parameters:: Passing additional parameters to user
functions.
Indices
* Function index:: Index of functions in optim.
* Concept index:: Concept index.
@end menu
@c ------------------------------------------------------------------
@node Scalar optimization
@chapter Functions for optimization of a scalar objective function
@cindex scalar optimization
@menu
Frontend
* nonlin_min:: Interface for scalar non-linear
optimization.
Backends
* lm_feasible:: L/M-like optimizer, constraints met
throughout optimization.
* octave_sqp:: A wrapper to core Octaves sqp function.
* siman:: Simulated annealing with constraints.
* samin:: Simulated annealing covering a range.
* d2_min:: Newton-like optimizer, no constraints.
Alternative frontend
* fmins:: Frontend for direct search
(gradient-less) algorithms.
Standalone functions
* nmsmax:: A Nelder-Mead simplex algorithm.
* mdsmax:: A multidirectional search algorithm.
* adsmax:: An alternating directions algorithm.
* nelder_mead_min:: Another Nelder-Mead algorithm.
* powell:: Direction-set (Powell's) method.
* bfgsmin:: Unconstrained BFGS algorithm.
* nrm:: Newton-Raphson algorithm.
* cg_min:: A conjugate gradient method.
* brent_line_min:: Linesearch, Brent method.
* line_min:: Linesearch (minimize a function along dx).
* de_min:: A differential evolution algorithm.
* battery:: Repeatedly call bfgsmin.
@end menu
@c ------------------------------------------------------------------
@node nonlin_min
@section Frontend nonlin_min for scalar non-linear optimization
@mfnindex nonlin_min
@DOCSTRING(nonlin_min)
@subheading Settings
The fields of the @var{settings} structure can be set with
@ref{XREFoptimset,,optimset,octave}.
For settings common to all frontends (including these for statistics)
@mysee
@ref{Common frontend options}.
For additional settings common to all optimization frontends
@mysee
@ref{Common optimization options}.
@subsubheading Specific defaults:
@multitable {@code{Algorithm}} {"lm_feasible"}
@item @code{Algorithm}:
@tab "lm_feasible"
@end multitable
@subsubheading Additional settings:
@table @code
@item objf_grad
Function computing the gradient of the objective function with respect
to the parameters. Default: real finite differences. Will be called
with the column vector of parameters and an informational structure as
arguments. If @code{dfdp} was specified by the user, the informational
structure has the fields @code{f}: value of objective function for
current parameters, @code{fixed}: logical vector indicating which
parameters are not optimized, so these partial derivatives need not be
computed and can be set to zero, @code{diffp}, @code{diff_onesided},
@code{lbound}, @code{ubound}: identical to the user settings of this
name, @code{plabels}: 1-dimensional cell-array of column-cell-arrays,
each column with labels for all parameters; the first column contains
the numerical indices of the parameters; the second and third columns,
present for structure based parameter handling,
@mysee
@ref{Parameter structures},
contain the names of the parameters and the subindices of
the parameters,
@mysee
@ref{Non-scalar parameters}, respectively. The default
gradient function will call the objective function with the second
argument set with fields @code{f}: as the @code{f} passed to the
gradient function, @code{plabels}: cell-array of 1x1 cell-arrays with
the entries of the column-cell-arrays of @code{plabels} as passed to the
jacobian function corresponding to current parameter, @code{side}:
@code{0} for one-sided interval, @code{1} or @code{2}, respectively, for
the sides of a two-sided interval, and @code{parallel}: logical scalar
indicating parallel computation of partial derivatives. This
information can be useful if the model function can omit some
computations depending on the currently computed partial derivative.
@item objf_hessian
Function computing the Hessian of the objective function with respect to
the parameters. The default is backend specific. Will be called with
the column vector of parameters as argument.
@item inverse_hessian
Logical scalar, indicating whether the Hessian function passed by the
user actually returns the inverse of the Hessian.
@item complex_step_derivative_objf
Logical scalar, default: @code{false}. Estimate gradient of objective
function with complex step derivative approximation. Use only if you
know that your objective function is suitable for this. No user
function for the gradient (@code{objf_grad}) must be specified.
@item save_state
String with path to a file which will be created for periodical saving
of the state of optimization. Useful for very long optimizations which
might get interrupted. The format of the saved state will be
backend-specific. Currently, only the @qcode{"siman"} backend honours
this option. Default: empty string, meaning no saving of state.
@item recover_state
String with path to a file created due to option @code{save_state},
which is used to recover a saved state before starting optimization.
Default: empty string, meaning no recovering of state.
@end table
@subheading Structure based parameter handling
Please
@mysee
@ref{Parameter structures}.
@subheading Backend information
Please
@mysee
@ref{Scalar optimization} and choose backend from menu.
@c ------------------------------------------------------------------
@node lm_feasible
@section Default backend lm_feasible of scalar optimization
@cindex lm_feasible
A Levenberg/Marquardt-like algorithm, attempting to honour constraints
throughout the course of optimization. This means that the initial
parameters must not violate constraints (to find an initial feasible set
of parameters, e.g. core Octaves @code{sqp} can be used
(
@mysee
@ref{octave_sqp}), by specifying an objective function which is
constant or which returns a norm of the distances to the initial
values). The Hessian is either supplied by the user or is approximated
by the BFGS algorithm. Core Octaves @code{sqp} performed better in some
tests with @emph{unconstrained} problems.
Returned value @var{cvg} will be @code{2} or @code{3} for success and
@code{0} or @code{-4} for failure (
@mysee
@ref{nonlin_min} for
meaning). Returned structure @var{outp} will have the fields
@code{niter}, @code{nobjf}, and @code{user_interaction}.
Backend-specific defaults are: @code{MaxIter}: 20, @code{fract_prec}:
@code{zeros (size (parameters))}, @code{max_fract_change}: @code{Inf}
for all parameters. The setting @code{TolX} is not honoured.
Interpretation of @code{Display}: if set to @qcode{"iter"}, currently
only information on applying @code{max_fract_change} is printed.
@c ------------------------------------------------------------------
@node octave_sqp
@section Backend wrapping core Octaves sqp function
@cindex octave_sqp
This backend calls the @code{sqp} function of core Octave, a sequential
quadratic programming algorithm with BFGS, so that it is usable with the
@code{nonlin_min} frontend of the optim package. @code{sqp} honours
constraints for the returned result, but not necessarily during the
course of optimization.
Compared to calling @code{sqp} directly (
@mysee
@ref{XREFsqp,,sqp,octave}),
@itemize
@item a different default (numerical) gradient function is used (that of
the frontend),
@item bounds are set to @code{+-Inf} by default, not to @code{+-realmax}.
@end itemize
The value of the additional setting @code{octave_sqp_tolerance}, a
tolerance to several termination criteria, is passed as the respective
argument to Octaves @code{sqp}, which provides the default @code{sqrt
(eps)}. The settings @code{TolFun}, @code{TolX}, @code{fract_prec},
@code{max_fract_change}, @code{Display}, and @code{user_interaction} are
not honoured.
Returned value @var{cvg} will be @code{1} for success and @code{0} or
@code{-4} for failure (
@mysee
@ref{nonlin_min} for meaning). Returned
structure @var{outp} will have the fields @code{niter}, @code{nobjf},
and @code{lambda}.
The default of @code{MaxIter} is that of Octaves @code{sqp} (100).
@c ------------------------------------------------------------------
@node siman
@section Simulated annealing backend siman of scalar optimization
@cindex siman
A simulated annealing (stochastic) optimizer, changing all parameters at
once in a single step, so being suitable for non-bound constraints.
No gradient or hessian of the objective function is used. The settings
@code{MaxIter}, @code{fract_prec}, @code{TolFun}, @code{TolX}, and
@code{max_fract_change} are not honoured.
Accepts the additional settings @code{T_init} (initial temperature,
default 0.01), @code{T_min} (final temperature, default 1.0e-5),
@code{mu_T} (factor of temperature decrease, default 1.005),
@code{iters_fixed_T} (iterations within one temperature step, default
10), @code{max_rand_step} (column vector or structure-based
configuration of maximum random steps for each parameter, default 0.005
* @var{pin}), @code{stoch_regain_constr} (if @code{true}, regain
constraints after a random step, otherwise take new random value until
constraints are met, default @code{false}), @code{trace_steps} (set
field @code{trace} of @var{outp} with a matrix with a row for each step,
first column iteration number, second column repeat number within
iteration, third column value of objective function, rest columns
parameter values, default @code{false}), and @code{siman_log} (set field
@code{log} of @var{outp} with a matrix with a row for each iteration,
first column temperature, second column value of objective function,
rest columns numbers of tries with decrease, no decrease but accepted,
and no decrease and rejected.
Steps with increase @code{diff} of objective function are accepted if
@code{rand (1) < exp (- diff / T)}, where @code{T} is the temperature of
the current iteration.
If regaining of constraints failed, optimization will be aborted and
returned value of @var{cvg} will be @code{0}. Otherwise, @var{cvg} will
be @code{1}. Returned structure @var{outp}, additionally to the possible
fields @code{trace} and @code{log} described above, will have the fields
@code{niter} and @code{user_interaction}.
Interpretation of @code{Display}: if set to @qcode{"iter"}, an
informational line is printed after each iteration.
If @code{parallel_local} is equivalent to @code{true}, the objective
function is evaluated for several parameter combinations in parallel. If
@code{parallel_local} is set to an integer @code{> 1}, this is the
maximal number of parallel processes; if it is @code{<= 1}, the maximal
number will be the number of available processor cores. The course of
optimization won't be changed by parallelization, provided the random
number generator starts with the same state. To achieve this, some of
the parallel results are discarded, causing the speedup to be smaller if
the rate of acceptance of results is high. Also, due to overhead, there
won't be any speedup, but even a slowdown, if the objective function is
not computationally extensive enough.
Honours options @code{save_state} and @code{recover_state}, described
for the frontend.
@c ------------------------------------------------------------------
@node samin
@section Simulated annealing backend samin of scalar optimization
@cindex samin
A simulated annealing (stochastic) optimizer, changing each parameter
separately, so not suitable for non-bound constraints. Requires
specification of lower and upper bounds for each parameter and attempts
to scan the whole region so defined.
No gradient or hessian of the objective function is used. The settings
@code{fract_prec} and @code{max_fract_change} are not honoured.
Maximal stepwidth is continuously adjusted according to the ratio of
accepted steps. Temperature is increased at the start until accepted
steps cover the whole parameter range, if this is not the case at the
initial temperature.
Accepts the additional settings @code{T_init} (initial temperature,
default 0.1), @code{mu_T} (factor of temperature decrease, default 1.2),
@code{iters_fixed_T} (subiterations within one temperature step, default
100), @code{iters_adjust_step} (number of subiterations after which
maximal stepwidth is adjusted, default 5), @code{niter_check_tolfun}
(number of iterations to compare with if checking the value of the
objective function for convergence, default 5), @code{trace_steps} (set
field @code{trace} of @var{outp} with a matrix with a row for each step,
first column iteration number, second column repeat number within
iteration, third column value of objective function, fourth column
temperature, rest columns parameter values, default @code{false}), and
@code{siman_log} (set field @code{log} of @var{outp} with a matrix with
a row for each iteration, first column temperature, second column value
of objective function, rest columns numbers of tries with decrease, no
decrease but accepted, no decrease and rejected, steps corrected back
into bounds, and steps yielding a new best value of the objective
function).
Steps with increase @code{diff} of objective function are accepted if
@code{rand (1) < exp (- diff / T)}, where @code{T} is the temperature of
the current iteration.
If converged, the returned value of @var{cvg} will be @code{1}, if
@code{MaxIters} is exceeded before convergence, @var{cvg} will be
@code{0}. The returned structure @var{outp}, additionally to the
possible fields @code{trace} and @code{log} described above, will have
the fields @code{niter} and @code{user_interaction}.
Interpretation of @code{Display}: if set to @qcode{"iter"},
informational text is printed after each iteration and at the end of
optimization. If set to @qcode{"final"}, informational text is printed
only at the end of optimization.
No parallel execution is performed in this backend.
Options @code{save_state} and @code{recover_state} are not honoured.
@c ------------------------------------------------------------------
@node d2_min
@section Unconstrained Newton-like optimization
@cindex d2_min
This backend features a Newton-like algorithm. The user has to supply a
Hessian function. No constraints are honoured. If the supplied Hessian
function actually returns the inverse of the Hessian, set
@code{inverse_hessian} to @code{true}. Supplying the inverse Hessian is
preferable, if possible.
Returned value @var{cvg} will be @code{2} or @code{3} for success and
@code{0} or @code{-1} for failure (
@mysee
@ref{nonlin_min} for
meaning). Returned structure @var{outp} will have the fields
@code{niter}, @code{nobjf}, and @code{user_interaction}.
Interpretation of @code{Display}: if set to @qcode{"iter"}, some
diagnostics are printed.
@c ------------------------------------------------------------------
@node fmins
@section Alternative frontend for gradient-less algorithms
@mfnindex fmins
@DOCSTRING(fmins)
@c ------------------------------------------------------------------
@node nmsmax
@section A Nelder-Mead simplex algorithm
@mfnindex nmsmax
This function is deprecated. It is available with a slightly different
interface in core Octave as `fminsearch'.
@subheading Helptext:
@DOCSTRINGVERBATIM(nmsmax)
@c ------------------------------------------------------------------
@node mdsmax
@section A multidirectional search algorithm
@mfnindex mdsmax
@subheading Helptext:
@DOCSTRINGVERBATIM(mdsmax)
@c ------------------------------------------------------------------
@node adsmax
@section An alternating directions algorithm
@mfnindex adsmax
@subheading Helptext:
@DOCSTRINGVERBATIM(adsmax)
@c ------------------------------------------------------------------
@node nelder_mead_min
@section Another Nelder-Mead algorithm
@mfnindex nelder_mead_min
This function does gradient-less minimization using the Nelder-Mead
algorithm. No constraints are honoured.
@subheading Helptext:
@DOCSTRINGVERBATIM(nelder_mead_min)
@c ------------------------------------------------------------------
@node powell
@section Direction-set (Powell's) method
@mfnindex powell
@DOCSTRING(powell)
@c ------------------------------------------------------------------
@node bfgsmin
@section Unconstrained BFGS algorithm
@mfnindex bfgsmin
BFGS or limited memory BFGS minimization of a function. No constraits
are honoured.
@subheading Helptext:
@DOCSTRINGVERBATIM(bfgsmin)
@c ------------------------------------------------------------------
@node nrm
@section Newton-Raphson algorithm
@mfnindex nrm
No constraints are honoured.
@DOCSTRING(nrm)
@c ------------------------------------------------------------------
@node cg_min
@section A conjugate gradient method
@mfnindex cg_min
@DOCSTRING(cg_min)
@c ------------------------------------------------------------------
@node brent_line_min
@section Linesearch, Brent method
@mfnindex brent_line_min
@DOCSTRING(brent_line_min)
@c ------------------------------------------------------------------
@node line_min
@section Linesearch, minimize a function along dx
@mfnindex line_min
@subheading Helptext:
@DOCSTRINGVERBATIM(line_min)
@c ------------------------------------------------------------------
@node de_min
@section A differential evolution (stochastic) optimizer
@mfnindex de_min
@subheading Helptext:
@DOCSTRINGVERBATIM(de_min)
@c ------------------------------------------------------------------
@node battery
@section Repeatedly call bfgsmin
@mfnindex battery
@subheading Helptext:
@DOCSTRINGVERBATIM(battery)
@c ------------------------------------------------------------------
@node Residual optimization
@chapter Functions for optimization of a model function returning an array
@cindex residual optimization
Model functions whose parameters are to be optimized may return a vector
or array of values. Either these or their differences to some constant
values (curve fitting) can be minimized in some sense, often, but not
necessarily, by minimizing the sum of their squares. It is usually
preferable to use optimizers designed for residual optimization for this
purpose. These can exploit information contained in the individual
elements of the returned array, which would not be possible if the user
calculated a norm (@abbr{e.g@.} sum of squares) of the elements and
performed a scalar optimization.
@menu
Optimization frontends
* nonlin_residmin:: The standard interface for non-linear
residual minimization.
* nonlin_curvefit:: A convenience interface, curve fitting.
Optimization backends
* lm_svd_feasible:: L/M algorithm with SVD, constraints met
throughout optimization.
Statistics frontends
* residmin_stat:: Statistics for residual minimization.
* curvefit_stat:: Statistics for curve fitting.
Statistics backends
* wls:: Statistics for weighted least squares.
Standalone functions
* lsqlin:: Linear least squares with linear
constraints.
* leasqr:: An older function for curve fitting.
* expfit:: Prony's method for non-linear exponential
fitting.
* polyfitinf:: Function polyfitinf for polynomial
fitting.
* wpolyfit:: Polynomial fitting suitable for polyconf.
* polyconf:: Confidence and prediction intervals for
polynomial fitting.
* LinearRegression:: Function LinearRegression.
* wsolve:: Another linear solver.
@end menu
@c ------------------------------------------------------------------
@node nonlin_residmin
@section Frontend nonlin_residmin for non-linear residual minimization
@mfnindex nonlin_residmin
@c include function help here
@DOCSTRING(nonlin_residmin)
@subheading Settings
The fields of the @var{settings} structure can be set with
@ref{XREFoptimset,,optimset,octave}.
For settings common to all frontends (including these for statistics)
@mysee
@ref{Common frontend options}.
For additional settings common to all optimization frontends
@mysee
@ref{Common optimization options}.
@subsubheading Specific defaults:
@multitable {@code{Algorithm}} {"lm_svd_feasible"}
@item @code{Algorithm}:
@tab "lm_svd_feasible"
@end multitable
@subsubheading Additional settings:
@table @code
@item weights
Array of weights for the residuals. Dimensions must match.
@anchor{XREFoptiondfdp}
@item dfdp
Function computing the Jacobian of the residuals with respect to the
parameters, assuming residuals are reshaped to a column vector.
Default: real finite differences. Will be called with the column vector
of parameters and an informational structure as arguments. If
@code{dfdp} was specified by the user, the informational structure has
the fields @code{f}: value of residuals for current parameters, reshaped
to a column vector, @code{fixed}: logical vector indicating which
parameters are not optimized, so these partial derivatives need not be
computed and can be set to zero, @code{diffp}, @code{diff_onesided},
@code{lbound}, @code{ubound}: identical to the user settings of this
name, @code{plabels}: 1-dimensional cell-array of column-cell-arrays,
each column with labels for all parameters; the first column contains
the numerical indices of the parameters; the second and third columns,
present for structure based parameter handling,
@mysee
@ref{Parameter structures},
contain the names of the parameters and the subindices of
the parameters,
@mysee
@ref{Non-scalar parameters}, respectively. The default
jacobian function will call the model function with the second argument
set with fields @code{f}: as the @code{f} passed to the jacobian
function, @code{plabels}: cell-array of 1x1 cell-arrays with the entries
of the column-cell-arrays of @code{plabels} as passed to the jacobian
function corresponding to current parameter, @code{side}: @code{0} for
one-sided interval, @code{1} or @code{2}, respectively, for the sides of
a two-sided interval, and @code{parallel}: logical scalar indicating
parallel computation of partial derivatives. This information can be
useful if the model function can omit some computations depending on the
currently computed partial derivative.
@item complex_step_derivative_f
Logical scalar, default: @code{false}. Estimate Jacobian of model
function with complex step derivative approximation. Use only if you
know that your model function is suitable for this. No user function for
the Jacobian (@code{dfdp}) must be specified.
@item plot_cmd
Function enabling backend to plot results or intermediate results. Will
be called with current computed residuals. Default: plot nothing. This
setting is deprecated and will disappear. Please use
@code{user_interaction} instead (
@mysee
@ref{Common optimization options}).
@end table
@subheading Structure based parameter handling
Please
@mysee
@ref{Parameter structures}.
@subheading Backend information
Please
@mysee
@ref{Residual optimization} and choose backend from menu under
`Optimization backends'.
@c ------------------------------------------------------------------
@node nonlin_curvefit
@section Function nonlin_curvefit() for curve fitting
@mfnindex nonlin_curvefit
In curve fitting, the model function computes values from a constant set
of `independents', and the intention is to minimize the differences of
these computed values to a constant set of `observations'. This can be
done with @code{nonlin_residmin}, but it is more convenient to use
@code{nonlin_curvefit}, which cares for passing the constant
`independents' to the model function and for calculating the differences
to the constant `observations'.
However, if in some optimization problem you notice that you end up with
passing dummy-values for the `independents' and zeros for the
`observations', you can more naturally use @code{nonlin_residmin}
instead of @code{nonlin_curvefit}.
@c include function help here
@DOCSTRING(nonlin_curvefit)
@c replace the cut out text
Also, if the setting @code{user_interaction} is given, additional
information is passed to these functions,
@mysee
@ref{Common optimization options}.
@c ------------------------------------------------------------------
@node lm_svd_feasible
@section Default backend lm_svd_feasible of residual minimization
@cindex lm_svd_feasible
Levenberg/Marquardt algorithm using singular value decomposition.
Constraints must be met by the initial parameters and are attempted to
be kept met throughout the optimization.
Returned value @var{cvg} will be @code{0}, @code{1}, or @code{2}.
Returned structure @var{outp} will have the fields @code{niter} and
@code{user_interaction}.
Backend-specific defaults are: @code{MaxIter}: 20, @code{fract_prec}:
@code{zeros (size (parameters))}, @code{max_fract_change}: @code{Inf}
for all parameters. The setting @code{TolX} is not honoured.
Interpretation of @code{Display}: if set to @qcode{"iter"}, currently
some diagnostics are printed.
Specific option: @code{lm_svd_feasible_alt_s}: if falling back to nearly
gradient descent, do it more like original Levenberg/Marquardt method,
with descent in each gradient component; for testing only.
@c ------------------------------------------------------------------
@node residmin_stat
@section Statistics for residual minimization
@mfnindex residmin_stat
@c include function help here
@DOCSTRING(residmin_stat)
@subheading Further settings
The fields of the @var{settings} structure can be set with
@ref{XREFoptimset,,optimset,octave}.
For settings common to all frontends
@mysee
@ref{Common frontend options}.
@subsubheading Additional settings:
@table @code
@item objf_type
Type of objective function of the optimization; must be specified in
many cases. This determines which backends to use. Currently, there are
only backends for the type "wls" (weighted least squares).
@item residuals
@item covd
Optional information on the result of optimization, residuals and
covariance matrix of data, respectively.
@item weights
Array of weights applied to the residuals in the previous
optimization. Dimensions must match those of the residuals.
@item dfdp
Can be set in the same way and has the same default as in
@code{nonlin_residmin} (
@mysee
@ref{nonlin_residmin}), but alternatively may
already contain the computed Jacobian of the model function at the final
parameters in matrix- or structure-form.
@item complex_step_derivative_f
Estimate Jacobian of model function with complex step derivative
approximation. Use only if you know that your model function is suitable
for this. No user function for the Jacobian (@code{dfdp}) must be
specified.
@end table
@subheading Structure based parameter handling
Please
@mysee
@ref{Parameter structures}.
@subheading Backend information
Please
@mysee
@ref{Residual optimization} and choose backend from menu under
`Statistics backends'.
@c ------------------------------------------------------------------
@node curvefit_stat
@section Statistics for curve fitting
@mfnindex curvefit_stat
As @code{nonlin_curvefit} can be used instead of @code{nonlin_residmin}
for curve fitting (
@mysee
@ref{nonlin_curvefit},
@mysee
@ref{nonlin_residmin}),
@code{curvefit_stat} can be used instead of @code{residmin_stat}
(
@mysee
@ref{residmin_stat}) for statistics on the results of curve fitting.
@c include function help here
@DOCSTRING(curvefit_stat)
@c ------------------------------------------------------------------
@node wls
@section Statistics for weighted least squares
@cindex wls
@cindex statistics for weighted least squares
The backends for @code{objf_type == "wls"} (currently the only supported
type of objective function) compute @code{covd} (due to user request or
as a prerequisite for @code{covp} and @code{corp}) as a diagonal matrix
by assuming that the variances of data points are proportional to the
reciprocal of the squared @code{weights} and guessing the factor of
proportionality from the residuals. If @code{covp} is not defined
(e.g. because the Jacobian has no full rank), an attempt is made to
still compute its uniquely defined elements, if any. In @code{corp},
interdependent parameters can cause elements of @code{1} or @code{-1},
which in this case are not the real coefficients of correlation, but
rather indicate the direction of parameter interdependence. To be
consistent with this, an attempt is made (often not successful) to
identify parameter interdependence and mark it with elements of @code{1}
or @code{-1} in @code{corp} even if the respective elements of
@code{covp} can not be computed.
@c ------------------------------------------------------------------
@node lsqlin
@section Linear least squares with linear constraints.
@mfnindex lsqlin
(This function does not fit well into this chapter because it is
actually a special case of quadratic programming).
@DOCSTRING(lsqlin)
@c ------------------------------------------------------------------
@node leasqr
@section An older function for curve fitting
@mfnindex leasqr
This was a popular function for curve fitting and has been enhanced to
honour constraints. @code{nonlin_curvefit} (
@mysee
@ref{nonlin_curvefit}) does
now the same job if used with the default backend, and should be
prefered due to its more powerful interface. The statistics returned by
@code{leasqr} can also (and partially better) be computed with
@code{curvefit_stat} (
@mysee
@ref{curvefit_stat}). There are currently two
things which still only @code{leasqr} does:
@itemize
@item internally providing a function for plotting fits during
optimization,
@item returning a pre-computed matrix for determining confidence
regions.
@end itemize
@DOCSTRING(leasqr)
@c ------------------------------------------------------------------
@node expfit
@section Prony's method for non-linear exponential fitting
@mfnindex expfit
@subheading Helptext:
@DOCSTRINGVERBATIM(expfit)
@c ------------------------------------------------------------------
@node polyfitinf
@section Function polyfitinf for polynomial fitting.
@mfnindex polyfitinf
@subheading Helptext:
@DOCSTRINGVERBATIM(polyfitinf)
@c ------------------------------------------------------------------
@node wpolyfit
@section Polynomial fitting suitable for polyconf
@mfnindex wpolyfit
@DOCSTRING(wpolyfit)
See also @ref{XREFpolyfit,,polyfit,octave}.
@c ------------------------------------------------------------------
@node polyconf
@section Confidence and prediction intervals for polynomial fitting
@mfnindex polyconf
@subheading Helptext:
@DOCSTRINGVERBATIM(polyconf)
@c ------------------------------------------------------------------
@node LinearRegression
@section Function LinearRegression
@mfnindex LinearRegression
@DOCSTRING(LinearRegression)
See also @ref{XREFregress,,regress,octave}, @ref{XREFleasqr,,leasqr},
@ref{XREFnonlin_curvefit,,nonlin_curvefit},
@ref{XREFpolyfit,,polyfit,octave}, @ref{XREFwpolyfit,,wpolyfit},
@ref{XREFexpfit,,expfit}.
@c ------------------------------------------------------------------
@node wsolve
@section Function wsolve, another linear solver
@mfnindex wsolve
@subheading Helptext:
@DOCSTRINGVERBATIM(wsolve)
@c ------------------------------------------------------------------
@node Zero finders
@chapter Functions for finding the zero of a nonlinear user function
@cindex zero finders
There is only one dedicated zero finder in the optim package, which is
just a vectorized version of Octaves fzero
(
@mysee
@ref{XREFfzero,,fzero,octave}).
@menu
* vfzero:: A vectorized version of fzero.
@end menu
@c ------------------------------------------------------------------
@node vfzero
@section A vectorized version of fzero
@mfnindex vfzero
@DOCSTRING(vfzero)
@c ------------------------------------------------------------------
@node Gradient functions
@chapter Functions for numerical approximation of gradients and Hessians
@cindex gradient functions
You should not usually need to use these functions directly or pass them
as arguments. They should be chosen and used by optimizer functions,
possibly subject to their configuration options.
@menu
* dfpdp:: Direct user interface to default numerical
gradient method of new frontends.
* deriv:: Higher order numerical derivatives.
* numgradient:: Another numerical gradient function.
* numhessian:: Numerical Hessian function.
* cdiff:: A string, yielding the numerical gradient if
evaluated.
* jacobs:: Complex step derivatives.
@end menu
@c ------------------------------------------------------------------
@node dfpdp
@section Direct user interface to default numerical gradient method of new frontends
@mfnindex dfpdp
@subheading Helptext:
@DOCSTRINGVERBATIM(dfpdp)
@c ------------------------------------------------------------------
@node deriv
@section Higher order numerical derivatives
@mfnindex deriv
@DOCSTRING(deriv)
@c ------------------------------------------------------------------
@node numgradient
@section Another numerical gradient function
@mfnindex numgradient
@subheading Helptext:
@DOCSTRINGVERBATIM(numgradient)
@c ------------------------------------------------------------------
@node numhessian
@section Numerical Hessian function
@mfnindex numhessian
@subheading Helptext:
@DOCSTRINGVERBATIM(numhessian)
@c ------------------------------------------------------------------
@node cdiff
@section A string, yielding the numerical gradient if evaluated
@mfnindex cdiff
@subheading Helptext:
@DOCSTRINGVERBATIM(cdiff)
@c ------------------------------------------------------------------
@node jacobs
@section Complex step derivatives
@mfnindex jacobs
@DOCSTRING(jacobs)
@c ------------------------------------------------------------------
@node Helper functions
@chapter Functions for algebraic tasks common to optimization problems
@cindex helper functions
@menu
* cpiv_bard:: A complementary pivoting algorithm.
* gjp:: Gauss-Jordan pivoting.
@end menu
@c ------------------------------------------------------------------
@node cpiv_bard
@section A complementary pivoting algorithm
@mfnindex cpiv_bard
@subheading Helptext:
@DOCSTRINGVERBATIM(cpiv_bard)
@c ------------------------------------------------------------------
@node gjp
@section Gauss-Jordan pivoting
@mfnindex gjp
@subheading Helptext:
@DOCSTRINGVERBATIM(gjp)
@c ------------------------------------------------------------------
@node Documentation
@chapter Function optim_doc to view documentation
@mfnindex optim_doc
@DOCSTRING(optim_doc)
@c ------------------------------------------------------------------
@node Compatibility wrappers
@chapter Traditional functions, working by calling a different function
@cindex compatibility wrappers
@menu
* linprog:: Linear programming.
* quadprog:: Quadratic programming.
* lsqnonlin:: Non-linear residual minimization.
* lsqcurvefit:: Curve fitting.
* nlinfit:: Non-linear regression.
@end menu
@c ------------------------------------------------------------------
@node linprog
@section Linear programming
@mfnindex linprog
This function works by calling @code{glpk} of core Octave.
@DOCSTRING(linprog)
See also @ref{XREFglpk,,glpk,octave}.
@c ------------------------------------------------------------------
@node quadprog
@section Quadratic programming
@mfnindex quadprog
This function is similar to @code{qp} of core Octave.
@DOCSTRING(quadprog)
@c ------------------------------------------------------------------
@node lsqnonlin
@section Non-linear residual minimization
@mfnindex lsqnonlin
This function is for Matlab compatibility. It attempts to work like
@code{lsqnonlin} by calling @ref{nonlin_residmin}.
@DOCSTRING(lsqnonlin)
@c ------------------------------------------------------------------
@node lsqcurvefit
@section Curve fitting
@mfnindex lsqcurvefit
This function is for Matlab compatibility. It attempts to work like
@code{lsqcurvefit} by calling @ref{nonlin_curvefit}.
@DOCSTRING(lsqcurvefit)
@c ------------------------------------------------------------------
@node nlinfit
@section Non-linear regression.
@mfnindex nlinfit
This function is for Matlab compatibility. It attempts to work like
@code{nlinfit} by calling @ref{nonlin_curvefit} and @ref{curvefit_stat}.
@DOCSTRING(nlinfit)
@c ------------------------------------------------------------------
@node Common frontend options
@chapter Options common to all frontends
@cindex common options
All frontends for optimization and for result statistics
(@ref{nonlin_min}, @ref{nonlin_residmin}, @ref{nonlin_curvefit},
@ref{residmin_stat}, @ref{curvefit_stat})accept the following options,
settable with @ref{XREFoptimset,,optimset,octave}.
These options are handled within the frontend.
@table @code
@item FinDiffRelStep
Column vector (or scalar, for all parameters) of fractional intervals
supposed to be used by gradient or Jacobian functions performing finite
differencing. Default: @code{.002 * ones (size (parameters))} for
central intervals and @code{.001 * ones (size (parameters))} for
one-sided intervals. The default function for finite differencing won't
let the absolute interval width get smaller than
@code{abs (FinDiffRelStep .* TypicalX} (see below).
@item diffp
Can be used alternatively to @code{FinDiffRelStep}, but for central
intervals twice the specified value will be used for backwards compatibility.
@item diff_onesided
Logical column vector (or scalar, for all parameters) indicating the
parameters for which one-sided intervals (instead of central intervals)
should be used by gradient or Jacobian functions performing finite
differencing. Default: @code{false (size (parameters))}.
@item FinDiffType
Can be used alternatively to @code{diff_onesided}, but always applies to
all parameters at once. Possible values: @code{"central"} (central
intervals) or @code{"forward"} (one-sided intervals).
@item TypicalX
Column vector (or scalar, for all parameters) whose absolute value
specifies minimal absolute parameter values for computation of intervals
in finite differencing by gradient or Jacobian functions (see
@code{FinDiffRelStep}). Default: 0.0001. Must not be zero.
@item cstep
Scalar step size for complex step derivative approximation of gradients
or Jacobians. Default: 1e-20.
@item parallel_local
Logical or numeric scalar, default: @code{false}. If the @code{parallel}
package, @code{version >= 2.0.5}, is loaded, estimate gradients of
objective function and Jacobians of model function and of constraints in
parallel processes. If @code{parallel_local} is set to an integer
@code{> 1}, this is number of parallel processes; if it is @code{<= 1},
the number of processes will be the number of available processor cores.
Works for default (real) finite differences and for complex step
derivatives. Due to overhead, a speed advantage can only be expected if
objective function, model function or constraint functions are time
consuming enough. Additionally, this setting is also passed to the
individual optimization backends, which may also consider this option
(see documentation of backends). If this option is equivalent to
@code{true}, a warning (ID: @code{optim:parallel_local}) will be issued
if no @code{parallel} package of a correct version is loaded.
@item parallel_net
Empty (default) or a parallel connections object, see function
@code{pconnect} of the @code{parallel} package. If not empty, estimate
gradients of objective function and Jacobians of model function and of
constraints using parallel processing in a network of machines. This
option currently only takes effect with patched versions of Octave. The
patch in @url{https://savannah.gnu.org/bugs/download.php?file_id=34902}
can be tried. The considerations regarding a speed advantage are similar
to those for option @code{parallel_local}.
@item fixed
Logical column vector indicating which parameters are not optimized, but
kept to their inital value.
@end table
@c ------------------------------------------------------------------
@node Common optimization options
@chapter Options common to all optimization frontends
@cindex common optimization options
All frontends for optimization (@ref{nonlin_min}, @ref{nonlin_residmin},
@ref{nonlin_curvefit}) accept the following options, settable with
@ref{XREFoptimset,,optimset,octave}.
@subheading Settings handled within the frontend
@table @code
@item Algorithm
String specifying the backend.
@item complex_step_derivative_inequc,
@item complex_step_derivative_equc
Logical scalars, default: @code{false}. Estimate Jacobian of general
inequality constraints and equality constraints, respectively, with
complex step derivative approximation. Use only if you know that your
function of general inequality constraints or function of general
equality constraints, respectively, is suitable for this. No user
function for the respective Jacobian must be specified.
@end table
@subheading Settings passed to the backend
Which of these options are actually honored is noted in the descriptions
of the individual backends.
@table @code
@item lbound,
@item ubound
Column vectors of lower and upper bounds for parameters. Default:
@code{-Inf} and @code{+Inf}, respectively. The bounds are non-strict,
i.e. parameters are allowed to be exactly equal to a bound. The default
function for gradients or Jacobians will respect bounds (but no further
inequality constraints) in finite differencing if the backend respects
bounds even during the course of optimization.
@item inequc
Further inequality constraints. Cell-array containing up to four
entries, two entries for linear inequality constraints and/or one or two
entries for general inequality constraints. Either linear or general
constraints may be the first entries, but the two entries for linear
constraints must be adjacent and, if two entries are given for general
constraints, they also must be adjacent. The two entries for linear
constraints are a matrix (say @code{m}) and a vector (say @code{v}),
specifying linear inequality constraints of the form @code{m.' *
parameters + v >= 0}. The first entry for general constraints must be a
differentiable column-vector valued function (say @code{h}), specifying
general inequality constraints of the form @code{h (p[, idx]) >= 0};
@code{p} is the column vector of optimized paraters and the optional
argument @code{idx} is a logical index. @code{h} has to return the
values of all constraints if @code{idx} is not given. It may choose to
return only the indexed constraints if @code{idx} is given (so
computation of the other constraints can be spared); in this case, the
additional setting @code{f_inequc_idx} has to be set to @code{true}. In
gradient determination, this function may be called with an
informational third argument, whose content depends on the function for
gradient determination. If a second entry for general inequality
constraints is given, it must be a function computing the jacobian of
the constraints with respect to the parameters. For this function, the
description of the setting @code{dfdp},
@mysee
@ref{XREFoptiondfdp,,dfdp},
applies, with 2 exceptions: 1) it is called with 3 arguments since it
has an additional argument @code{idx}, a logical index, at second
position, indicating which rows of the jacobian must be returned (if the
function chooses to return only indexed rows, the additional setting
@code{df_inequc_idx} has to be set to @code{true}). 2) the default
jacobian function calls @code{h} with 3 arguments, since the argument
@code{idx} is also supplied. Note that specifying linear constraints as
general constraints will generally waste performance, even if further,
non-linear, general constraints are also specified.
@item f_inequc_idx,
@item df_inequc_idx
Indicate that functions for general inequality constraints or their
jacobian, respectively, return only the values or derivatives for the
indexed parameters. See description of setting @code{inequc} above.
@item equc
Equality constraints. Specified the same way as inequality constraints
(see @code{inequc} above).
@item f_equc_idx,
@item df_equc_idx
As @code{f_inequc_idx} and @code{df_inequc_idx} above, but for equality
constraints.
@item cpiv
Function for complementary pivoting, usable in algorithms for
constraints. Default: @code{cpiv_bard}. Only the default function is
supplied with the package.
@item TolFun
Minimum fractional improvement in objective function (e.g. sum of
squares) in an iteration (termination criterium). Default: .0001.
@item TolX
Minimum fractional change in a norm of the parameters in an iteration
(termination criterium). Default: backend specific.
@item MaxIter
Maximum number of iterations (termination criterium). Default:
backend-specific.
@item fract_prec
Column Vector, minimum fractional changes of corresponding parameters in
an iteration (termination criterium if violated in two consecutive
iterations). Default: backend-specific.
@item max_fract_change
Column Vector, enforced maximum fractional changes in corresponding
parameters in an iteration. Default: backend-specific.
@item Display
String indicating the degree of verbosity. Default:
@qcode{"off"}. Possible values are currently @qcode{"off"} (no messages)
and @qcode{"iter"} (some messages after each iteration). Support of
this setting and its exact interpretation are backend-specific.
@item debug
Logical scalar, default: @code{false}. Will be passed to the backend,
which might print debugging information if @code{true}.
@item FunValCheck
If @qcode{"on"}, the output of user functions will be sanity-checked.
Default: @qcode{"off"}.
@item user_interaction
@c This setting has deliberately not been named as its Matlab equivalent
@c `OutputFcn' since it differs from the latter by requiring the
@c functions to return _two_ outputs. The rationale for the difference
@c is that information about the reason for a user-stop should be
@c possible to pass in the output. The second output can't be made
@c optional without possibly altering the warning state under which the
@c user-function runs.
Handle to a user function or cell-array with a number of these.
Functions must have this interface:
@example
[@var{stop}, @var{info}] = some_user_function (@var{p}, @var{vals},
@var{state});
@end example
If @var{stop} is @code{true}, the algorithm stops. In @var{info}
information about the reason for stopping can be returned in a free
format. @var{info} can be set to be empty, but it must be set. Note
that this is different from the otherwise similar Matlab setting
@code{OutputFcn}. The functions will be called by the algorithms at the
start with @var{state} set to @qcode{init}, after each iteration with
@var{state} set to @qcode{iter}, and at the end with @var{state} set to
@qcode{done}. @var{p} contains the current parameters, and @var{vals}
is a structure with other current values, the possible fields are
currently:
@table @code
@item iteration
number of the current iteration,
@item fval
value of objective function (for scalar optimization),
@item residual
residuals (for residual-based optimization),
@item model_y
in @code{nonlin_curvefit}, the output of the model function,
@item observations
in @code{nonlin_curvefit}, the constant observations,
@item model_x
in @code{nonlin_curvefit}, the constant argument @var{x}.
@end table
Information about the output of these functions when they were called
the last time (possibly causing a stop) will be contained in the output
@var{outp} of the frontend in field @code{user_interaction}. Subfield
@code{stop} is a vector containing the @var{stop} outputs of each
function, subfield @code{info} is a cell-array containing the output
@var{info} of each function. In the case of a stop, the output
@var{cvg} of the frontent will be @code{-1}.
@end table
@c ------------------------------------------------------------------
@node Parameter structures
@chapter Handling of structures of optimized parameters
@cindex parameter structures
It can be convenient not to handle the optimized parameters as elements
of a vector, but as named fields of a structure. The frontends
@code{nonlin_residmin}, @code{nonlin_curvefit}, @code{residmin_stat},
@code{curvefit_stat}, and @code{nonlin_min} can accept parameter
information in structure form, and can pass the parameters as a
structure to user functions, although the backends still handle the
parameters as vectors.
To use this feature, the initial parameters must be given in structure
form, or the setting @code{param_order} must be given, a cell-array with
names of the parameters. If both is done, only the parameters in
structure fields named in @code{param_order} will be optimized. If
there are still some non-structure-based configuration settings or user
functions, specifying @code{param_order} is mandatory even if the
initial parameters are given in structure form.
If the initial parameters are a structure, the parameters being the
optimization result will also be returned as a structure.
@menu
* Structure-based user functions:: Specify which user functions accept
parameter structures.
* Structure-based gradients and Hessians:: Format of returned values of
structure-based gradient and Hessian functions.
* Structure-based linear constraints:: Specify structure-based linear
constraints.
* Structure-based configuration settings:: Parameter-related
configuration settings in structure form.
* Non-scalar parameters:: Handling named parameter arrays.
@end menu
@c ------------------------------------------------------------------
@node Structure-based user functions
@section Specify which user functions accept parameter structures
@cindex structure-based user functions
The frontend must be told which user functions accept parameter
structures by setting corresponding settings to @code{true}. The names
of these settings depend on which user functions are applicable to the
respective frontend and are shown in the following table.
@multitable {inequality constraints} {@code{hessian_objf_pstruct}} {@code{nonlin_residmin},}
@headitem User function @tab Setting @tab Frontends
@item Objective function
@tab @code{objf_pstruct}
@tab @code{nonlin_min}
@item Gradient
@tab @code{grad_objf_pstruct}
@tab @code{nonlin_min}
@item Hessian
@tab @code{hessian_objf_pstruct}
@tab @code{nonlin_min}
@item Model function
@tab @code{f_pstruct}
@tab @code{nonlin_residmin}, @code{nonlin_curvefit}, @code{residmin_stat}, @code{curvefit_stat}
@item Jacobian
@tab @code{df_pstruct}
@tab @code{nonlin_residmin}, @code{nonlin_curvefit}, @code{residmin_stat}, @code{curvefit_stat}
@item General inequality constraints
@tab @code{f_inequc_pstruct}
@tab @code{nonlin_min}, @code{nonlin_residmin}, @code{nonlin_curvefit}
@item Jacobian of general inequality constraints
@tab @code{df_inequc_pstruct}
@tab @code{nonlin_min}, @code{nonlin_residmin}, @code{nonlin_curvefit}
@item General equality constraints
@tab @code{f_equc_pstruct}
@tab @code{nonlin_min}, @code{nonlin_residmin}, @code{nonlin_curvefit}
@item Jacobian of general equality constraints
@tab @code{df_equc_pstruct}
@tab @code{nonlin_min}, @code{nonlin_residmin}, @code{nonlin_curvefit}
@end multitable
@c ------------------------------------------------------------------
@node Structure-based gradients and Hessians
@section Format of returned values of structure-based gradient and Hessian functions
@cindex structure-based gradients
@cindex structure-based Hessians
Structure-based gradient or Jacobian functions, including Jacobians of
general constraints, must return the partial derivatives as fields of a
structure under the respective parameter names. For gradients, the
partial derivatives are scalar. For Jacobians, the partial derivatives
must be column vectors.
Structure-based Hessian functions must return the 2nd derivatives as
subfields in a two-level structure of parameter names. For example, if
the parameter names are @code{a} and @code{b}, the 2nd derivative with
respect to @code{a} and @code{b} must be in the field
@code{returned_structure.a.b} or @code{returned_structure.b.a} (there is
no need to specify both).
@c ------------------------------------------------------------------
@node Structure-based linear constraints
@section Specify structure-based linear constraints
@cindex structure-based linear constraints
Linear constraints, otherwise specified with a matrix and a vector@c
@c (@xref{})
, can be adapted to structure-based parameter handling by specifying,
instead of a matrix, a structure containing the rows of the matrix in
fields under the respective parameter names. In this case, rows
containing only zeros need not be given.
@c ------------------------------------------------------------------
@node Structure-based configuration settings
@section Parameter-related configuration settings in structure form
@cindex structure-based configuration settings
The vector-based settings @code{lbound}, @code{ubound}, @code{fixed},
@code{diffp}, @code{diff_onesided}, @code{fract_prec}, and
@code{max_fract_change} can be replaced by the setting
@code{param_config}. It is a structure that can contain fields
corresponding to parameter names. For each such field, there may be
subfields with the same names as the above vector-based settings, but
containing a scalar value for the respective parameter.
For example, if the parameters are named @code{a} and @code{b}, instead
of specifying
@example
settings = optimset ("lbound", [-Inf; 0],
"diff_onesided", [true; true]);
@end example
one can specify
@example
pconf.b.lbound = 0;
pconf.a.diff_onesided = true;
pconf.b.diff_onesided = true;
settings = optimset ("param_config", pconf);
@end example
If @code{param_config} is specified, none of the above vector-based
settings may be used.
@c ------------------------------------------------------------------
@node Non-scalar parameters
@section Handling named parameter arrays
@cindex non-scalar parameters
@cindex named parameter arrays
Parameters in named structure fields are allowed to be non-scalar real
arrays. In this case, their dimensions must be given by the setting
@code{param_dims}, a cell-array of dimension vectors, each containing at
least two dimensions; if not given, dimensions are taken from the
initial parameters, if these are given in a structure.
If there are any vector-based settings or not structure-based linear
constraints, they must correspond to an order of parameters defined as
follows:
All named parameter arrays are reshaped to vectors. Then, all
parameters, scalars and vectors, are concatenated in the order of
parameter names, given by the user.
Structure-based settings or structure-based initial parameters must
contain arrays with dimensions reshapable to those of the respective
parameters.
@c ------------------------------------------------------------------
@node Additional parameters
@chapter Passing additional parameters to user functions
@cindex additional parameters
Optimizers often require the user to supply functions (@abbr{e.g@.}
objective functions, model functions, constraint functions). The
interface of these functions --- arguments and returned values --- is
defined by the optimizer. Often, a user function needs additional
arguments, not covered by the defined interface, which are constant
throughout the optimization. These can be supplied by wrapping the user
function into an anonymous function. @xref{Anonymous
Functions,,,octave}, for further explanation and examples.
There are also some older optimizers in the optim package, written when
anonymous functions were not available in Octave. Some of these offer
an interface to user functions which is itself able to pass additional
constant variables in arbitrary argument positions. Newer optimizers
should not be written this way, since this is an unnecessary
complication.
Though it is possible to use global variables to pass additional data to
user functions, this is not recommended since it introduces the
possibility of name conflicts within the pool of global variables.
@c ------------------------------------------------------------------
@node Function index
@unnumbered Index of functions in optim
@printindex mfn
@c ------------------------------------------------------------------
@node Concept index
@unnumbered Concept index
@printindex cp
@bye
|