1 | /*************************************************************************
|
---|
2 | Copyright (c) 2009, Sergey Bochkanov (ALGLIB project).
|
---|
3 |
|
---|
4 | >>> SOURCE LICENSE >>>
|
---|
5 | This program is free software; you can redistribute it and/or modify
|
---|
6 | it under the terms of the GNU General Public License as published by
|
---|
7 | the Free Software Foundation (www.fsf.org); either version 2 of the
|
---|
8 | License, or (at your option) any later version.
|
---|
9 |
|
---|
10 | This program is distributed in the hope that it will be useful,
|
---|
11 | but WITHOUT ANY WARRANTY; without even the implied warranty of
|
---|
12 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
---|
13 | GNU General Public License for more details.
|
---|
14 |
|
---|
15 | A copy of the GNU General Public License is available at
|
---|
16 | http://www.fsf.org/licensing/licenses
|
---|
17 |
|
---|
18 | >>> END OF LICENSE >>>
|
---|
19 | *************************************************************************/
|
---|
20 |
|
---|
21 | using System;
|
---|
22 |
|
---|
23 | namespace alglib
|
---|
24 | {
|
---|
25 | public class minlm
|
---|
26 | {
|
---|
27 | public struct lmstate
|
---|
28 | {
|
---|
29 | public bool wrongparams;
|
---|
30 | public int n;
|
---|
31 | public int m;
|
---|
32 | public double epsf;
|
---|
33 | public double epsx;
|
---|
34 | public int maxits;
|
---|
35 | public int flags;
|
---|
36 | public int usermode;
|
---|
37 | public double[] x;
|
---|
38 | public double f;
|
---|
39 | public double[] fi;
|
---|
40 | public double[,] j;
|
---|
41 | public double[,] h;
|
---|
42 | public double[] g;
|
---|
43 | public bool needf;
|
---|
44 | public bool needfg;
|
---|
45 | public bool needfgh;
|
---|
46 | public bool needfij;
|
---|
47 | public bool xupdated;
|
---|
48 | public lbfgs.lbfgsstate internalstate;
|
---|
49 | public lbfgs.lbfgsreport internalrep;
|
---|
50 | public double[] xprec;
|
---|
51 | public double[] xbase;
|
---|
52 | public double[] xdir;
|
---|
53 | public double[] gbase;
|
---|
54 | public double[,] rawmodel;
|
---|
55 | public double[,] model;
|
---|
56 | public double[] work;
|
---|
57 | public AP.rcommstate rstate;
|
---|
58 | public int repiterationscount;
|
---|
59 | public int repterminationtype;
|
---|
60 | public int repnfunc;
|
---|
61 | public int repnjac;
|
---|
62 | public int repngrad;
|
---|
63 | public int repnhess;
|
---|
64 | public int repncholesky;
|
---|
65 | };
|
---|
66 |
|
---|
67 |
|
---|
68 | public struct lmreport
|
---|
69 | {
|
---|
70 | public int iterationscount;
|
---|
71 | public int terminationtype;
|
---|
72 | public int nfunc;
|
---|
73 | public int njac;
|
---|
74 | public int ngrad;
|
---|
75 | public int nhess;
|
---|
76 | public int ncholesky;
|
---|
77 | };
|
---|
78 |
|
---|
79 |
|
---|
80 |
|
---|
81 |
|
---|
82 | public const int lmmodefj = 0;
|
---|
83 | public const int lmmodefgj = 1;
|
---|
84 | public const int lmmodefgh = 2;
|
---|
85 | public const int lmflagnoprelbfgs = 1;
|
---|
86 | public const int lmflagnointlbfgs = 2;
|
---|
87 | public const int lmprelbfgsm = 5;
|
---|
88 | public const int lmintlbfgsits = 5;
|
---|
89 | public const int lbfgsnorealloc = 1;
|
---|
90 |
|
---|
91 |
|
---|
92 | /*************************************************************************
|
---|
93 | LEVENBERG-MARQUARDT-LIKE METHOD FOR NON-LINEAR OPTIMIZATION
|
---|
94 |
|
---|
95 | Optimization using function gradient and Hessian. Algorithm - Levenberg-
|
---|
96 | Marquardt modification with L-BFGS pre-optimization and internal
|
---|
97 | pre-conditioned L-BFGS optimization after each Levenberg-Marquardt step.
|
---|
98 |
|
---|
99 | Function F has general form (not "sum-of-squares"):
|
---|
100 |
|
---|
101 | F = F(x[0], ..., x[n-1])
|
---|
102 |
|
---|
103 | EXAMPLE
|
---|
104 |
|
---|
105 | See HTML-documentation.
|
---|
106 |
|
---|
107 | INPUT PARAMETERS:
|
---|
108 | N - dimension, N>1
|
---|
109 | X - initial solution, array[0..N-1]
|
---|
110 | EpsF - stopping criterion. Algorithm stops if
|
---|
111 | |F(k+1)-F(k)| <= EpsF*max{|F(k)|, |F(k+1)|, 1}
|
---|
112 | EpsX - stopping criterion. Algorithm stops if
|
---|
113 | |X(k+1)-X(k)| <= EpsX*(1+|X(k)|)
|
---|
114 | MaxIts - stopping criterion. Algorithm stops after MaxIts iterations.
|
---|
115 | MaxIts=0 means no stopping criterion.
|
---|
116 |
|
---|
117 | OUTPUT PARAMETERS:
|
---|
118 | State - structure which stores algorithm state between subsequent
|
---|
119 | calls of MinLMIteration. Used for reverse communication.
|
---|
120 | This structure should be passed to MinLMIteration subroutine.
|
---|
121 |
|
---|
122 | See also MinLMIteration, MinLMResults.
|
---|
123 |
|
---|
124 | NOTE
|
---|
125 |
|
---|
126 | Passing EpsF=0, EpsX=0 and MaxIts=0 (simultaneously) will lead to automatic
|
---|
127 | stopping criterion selection (small EpsX).
|
---|
128 |
|
---|
129 | -- ALGLIB --
|
---|
130 | Copyright 30.03.2009 by Bochkanov Sergey
|
---|
131 | *************************************************************************/
|
---|
132 | public static void minlmfgh(int n,
|
---|
133 | ref double[] x,
|
---|
134 | double epsf,
|
---|
135 | double epsx,
|
---|
136 | int maxits,
|
---|
137 | ref lmstate state)
|
---|
138 | {
|
---|
139 | int i_ = 0;
|
---|
140 |
|
---|
141 |
|
---|
142 | //
|
---|
143 | // Prepare RComm
|
---|
144 | //
|
---|
145 | state.rstate.ia = new int[3+1];
|
---|
146 | state.rstate.ba = new bool[0+1];
|
---|
147 | state.rstate.ra = new double[8+1];
|
---|
148 | state.rstate.stage = -1;
|
---|
149 |
|
---|
150 | //
|
---|
151 | // prepare internal structures
|
---|
152 | //
|
---|
153 | lmprepare(n, 0, true, ref state);
|
---|
154 |
|
---|
155 | //
|
---|
156 | // initialize, check parameters
|
---|
157 | //
|
---|
158 | state.xupdated = false;
|
---|
159 | state.n = n;
|
---|
160 | state.m = 0;
|
---|
161 | state.epsf = epsf;
|
---|
162 | state.epsx = epsx;
|
---|
163 | state.maxits = maxits;
|
---|
164 | state.flags = 0;
|
---|
165 | if( (double)(state.epsf)==(double)(0) & (double)(state.epsx)==(double)(0) & state.maxits==0 )
|
---|
166 | {
|
---|
167 | state.epsx = 1.0E-6;
|
---|
168 | }
|
---|
169 | state.usermode = lmmodefgh;
|
---|
170 | state.wrongparams = false;
|
---|
171 | if( n<1 | (double)(epsf)<(double)(0) | (double)(epsx)<(double)(0) | maxits<0 )
|
---|
172 | {
|
---|
173 | state.wrongparams = true;
|
---|
174 | return;
|
---|
175 | }
|
---|
176 | for(i_=0; i_<=n-1;i_++)
|
---|
177 | {
|
---|
178 | state.x[i_] = x[i_];
|
---|
179 | }
|
---|
180 | }
|
---|
181 |
|
---|
182 |
|
---|
183 | /*************************************************************************
|
---|
184 | LEVENBERG-MARQUARDT-LIKE METHOD FOR NON-LINEAR OPTIMIZATION
|
---|
185 |
|
---|
186 | Optimization using function gradient and Jacobian. Algorithm - Levenberg-
|
---|
187 | Marquardt modification with L-BFGS pre-optimization and internal
|
---|
188 | pre-conditioned L-BFGS optimization after each Levenberg-Marquardt step.
|
---|
189 |
|
---|
190 | Function F is represented as sum of squares:
|
---|
191 |
|
---|
192 | F = f[0]^2(x[0],...,x[n-1]) + ... + f[m-1]^2(x[0],...,x[n-1])
|
---|
193 |
|
---|
194 | EXAMPLE
|
---|
195 |
|
---|
196 | See HTML-documentation.
|
---|
197 |
|
---|
198 | INPUT PARAMETERS:
|
---|
199 | N - dimension, N>1
|
---|
200 | M - number of functions f[i]
|
---|
201 | X - initial solution, array[0..N-1]
|
---|
202 | EpsF - stopping criterion. Algorithm stops if
|
---|
203 | |F(k+1)-F(k)| <= EpsF*max{|F(k)|, |F(k+1)|, 1}
|
---|
204 | EpsX - stopping criterion. Algorithm stops if
|
---|
205 | |X(k+1)-X(k)| <= EpsX*(1+|X(k)|)
|
---|
206 | MaxIts - stopping criterion. Algorithm stops after MaxIts iterations.
|
---|
207 | MaxIts=0 means no stopping criterion.
|
---|
208 |
|
---|
209 | OUTPUT PARAMETERS:
|
---|
210 | State - structure which stores algorithm state between subsequent
|
---|
211 | calls of MinLMIteration. Used for reverse communication.
|
---|
212 | This structure should be passed to MinLMIteration subroutine.
|
---|
213 |
|
---|
214 | See also MinLMIteration, MinLMResults.
|
---|
215 |
|
---|
216 | NOTE
|
---|
217 |
|
---|
218 | Passing EpsF=0, EpsX=0 and MaxIts=0 (simultaneously) will lead to automatic
|
---|
219 | stopping criterion selection (small EpsX).
|
---|
220 |
|
---|
221 | -- ALGLIB --
|
---|
222 | Copyright 30.03.2009 by Bochkanov Sergey
|
---|
223 | *************************************************************************/
|
---|
224 | public static void minlmfgj(int n,
|
---|
225 | int m,
|
---|
226 | ref double[] x,
|
---|
227 | double epsf,
|
---|
228 | double epsx,
|
---|
229 | int maxits,
|
---|
230 | ref lmstate state)
|
---|
231 | {
|
---|
232 | int i_ = 0;
|
---|
233 |
|
---|
234 |
|
---|
235 | //
|
---|
236 | // Prepare RComm
|
---|
237 | //
|
---|
238 | state.rstate.ia = new int[3+1];
|
---|
239 | state.rstate.ba = new bool[0+1];
|
---|
240 | state.rstate.ra = new double[8+1];
|
---|
241 | state.rstate.stage = -1;
|
---|
242 |
|
---|
243 | //
|
---|
244 | // prepare internal structures
|
---|
245 | //
|
---|
246 | lmprepare(n, m, true, ref state);
|
---|
247 |
|
---|
248 | //
|
---|
249 | // initialize, check parameters
|
---|
250 | //
|
---|
251 | state.xupdated = false;
|
---|
252 | state.n = n;
|
---|
253 | state.m = m;
|
---|
254 | state.epsf = epsf;
|
---|
255 | state.epsx = epsx;
|
---|
256 | state.maxits = maxits;
|
---|
257 | state.flags = 0;
|
---|
258 | if( (double)(state.epsf)==(double)(0) & (double)(state.epsx)==(double)(0) & state.maxits==0 )
|
---|
259 | {
|
---|
260 | state.epsx = 1.0E-6;
|
---|
261 | }
|
---|
262 | state.usermode = lmmodefgj;
|
---|
263 | state.wrongparams = false;
|
---|
264 | if( n<1 | m<1 | (double)(epsf)<(double)(0) | (double)(epsx)<(double)(0) | maxits<0 )
|
---|
265 | {
|
---|
266 | state.wrongparams = true;
|
---|
267 | return;
|
---|
268 | }
|
---|
269 | for(i_=0; i_<=n-1;i_++)
|
---|
270 | {
|
---|
271 | state.x[i_] = x[i_];
|
---|
272 | }
|
---|
273 | }
|
---|
274 |
|
---|
275 |
|
---|
276 | /*************************************************************************
|
---|
277 | CLASSIC LEVENBERG-MARQUARDT METHOD FOR NON-LINEAR OPTIMIZATION
|
---|
278 |
|
---|
279 | Optimization using Jacobi matrix. Algorithm - classic Levenberg-Marquardt
|
---|
280 | method.
|
---|
281 |
|
---|
282 | Function F is represented as sum of squares:
|
---|
283 |
|
---|
284 | F = f[0]^2(x[0],...,x[n-1]) + ... + f[m-1]^2(x[0],...,x[n-1])
|
---|
285 |
|
---|
286 | EXAMPLE
|
---|
287 |
|
---|
288 | See HTML-documentation.
|
---|
289 |
|
---|
290 | INPUT PARAMETERS:
|
---|
291 | N - dimension, N>1
|
---|
292 | M - number of functions f[i]
|
---|
293 | X - initial solution, array[0..N-1]
|
---|
294 | EpsF - stopping criterion. Algorithm stops if
|
---|
295 | |F(k+1)-F(k)| <= EpsF*max{|F(k)|, |F(k+1)|, 1}
|
---|
296 | EpsX - stopping criterion. Algorithm stops if
|
---|
297 | |X(k+1)-X(k)| <= EpsX*(1+|X(k)|)
|
---|
298 | MaxIts - stopping criterion. Algorithm stops after MaxIts iterations.
|
---|
299 | MaxIts=0 means no stopping criterion.
|
---|
300 |
|
---|
301 | OUTPUT PARAMETERS:
|
---|
302 | State - structure which stores algorithm state between subsequent
|
---|
303 | calls of MinLMIteration. Used for reverse communication.
|
---|
304 | This structure should be passed to MinLMIteration subroutine.
|
---|
305 |
|
---|
306 | See also MinLMIteration, MinLMResults.
|
---|
307 |
|
---|
308 | NOTE
|
---|
309 |
|
---|
310 | Passing EpsF=0, EpsX=0 and MaxIts=0 (simultaneously) will lead to automatic
|
---|
311 | stopping criterion selection (small EpsX).
|
---|
312 |
|
---|
313 | -- ALGLIB --
|
---|
314 | Copyright 30.03.2009 by Bochkanov Sergey
|
---|
315 | *************************************************************************/
|
---|
316 | public static void minlmfj(int n,
|
---|
317 | int m,
|
---|
318 | ref double[] x,
|
---|
319 | double epsf,
|
---|
320 | double epsx,
|
---|
321 | int maxits,
|
---|
322 | ref lmstate state)
|
---|
323 | {
|
---|
324 | int i_ = 0;
|
---|
325 |
|
---|
326 |
|
---|
327 | //
|
---|
328 | // Prepare RComm
|
---|
329 | //
|
---|
330 | state.rstate.ia = new int[3+1];
|
---|
331 | state.rstate.ba = new bool[0+1];
|
---|
332 | state.rstate.ra = new double[8+1];
|
---|
333 | state.rstate.stage = -1;
|
---|
334 |
|
---|
335 | //
|
---|
336 | // prepare internal structures
|
---|
337 | //
|
---|
338 | lmprepare(n, m, true, ref state);
|
---|
339 |
|
---|
340 | //
|
---|
341 | // initialize, check parameters
|
---|
342 | //
|
---|
343 | state.xupdated = false;
|
---|
344 | state.n = n;
|
---|
345 | state.m = m;
|
---|
346 | state.epsf = epsf;
|
---|
347 | state.epsx = epsx;
|
---|
348 | state.maxits = maxits;
|
---|
349 | state.flags = 0;
|
---|
350 | if( (double)(state.epsf)==(double)(0) & (double)(state.epsx)==(double)(0) & state.maxits==0 )
|
---|
351 | {
|
---|
352 | state.epsx = 1.0E-6;
|
---|
353 | }
|
---|
354 | state.usermode = lmmodefj;
|
---|
355 | state.wrongparams = false;
|
---|
356 | if( n<1 | m<1 | (double)(epsf)<(double)(0) | (double)(epsx)<(double)(0) | maxits<0 )
|
---|
357 | {
|
---|
358 | state.wrongparams = true;
|
---|
359 | return;
|
---|
360 | }
|
---|
361 | for(i_=0; i_<=n-1;i_++)
|
---|
362 | {
|
---|
363 | state.x[i_] = x[i_];
|
---|
364 | }
|
---|
365 | }
|
---|
366 |
|
---|
367 |
|
---|
368 | /*************************************************************************
|
---|
369 | One Levenberg-Marquardt iteration.
|
---|
370 |
|
---|
371 | Called after inialization of State structure with MinLMXXX subroutine.
|
---|
372 | See HTML docs for examples.
|
---|
373 |
|
---|
374 | Input parameters:
|
---|
375 | State - structure which stores algorithm state between subsequent
|
---|
376 | calls and which is used for reverse communication. Must be
|
---|
377 | initialized with MinLMXXX call first.
|
---|
378 |
|
---|
379 | If subroutine returned False, iterative algorithm has converged.
|
---|
380 |
|
---|
381 | If subroutine returned True, then:
|
---|
382 | * if State.NeedF=True, - function value F at State.X[0..N-1]
|
---|
383 | is required
|
---|
384 | * if State.NeedFG=True - function value F and gradient G
|
---|
385 | are required
|
---|
386 | * if State.NeedFiJ=True - function vector f[i] and Jacobi matrix J
|
---|
387 | are required
|
---|
388 | * if State.NeedFGH=True - function value F, gradient G and Hesian H
|
---|
389 | are required
|
---|
390 |
|
---|
391 | One and only one of this fields can be set at time.
|
---|
392 |
|
---|
393 | Results are stored:
|
---|
394 | * function value - in LMState.F
|
---|
395 | * gradient - in LMState.G[0..N-1]
|
---|
396 | * Jacobi matrix - in LMState.J[0..M-1,0..N-1]
|
---|
397 | * Hessian - in LMState.H[0..N-1,0..N-1]
|
---|
398 |
|
---|
399 | -- ALGLIB --
|
---|
400 | Copyright 10.03.2009 by Bochkanov Sergey
|
---|
401 | *************************************************************************/
|
---|
402 | public static bool minlmiteration(ref lmstate state)
|
---|
403 | {
|
---|
404 | bool result = new bool();
|
---|
405 | int n = 0;
|
---|
406 | int m = 0;
|
---|
407 | int i = 0;
|
---|
408 | double xnorm = 0;
|
---|
409 | double stepnorm = 0;
|
---|
410 | bool spd = new bool();
|
---|
411 | double fbase = 0;
|
---|
412 | double fnew = 0;
|
---|
413 | double lambda = 0;
|
---|
414 | double nu = 0;
|
---|
415 | double lambdaup = 0;
|
---|
416 | double lambdadown = 0;
|
---|
417 | int lbfgsflags = 0;
|
---|
418 | double v = 0;
|
---|
419 | int i_ = 0;
|
---|
420 |
|
---|
421 |
|
---|
422 | //
|
---|
423 | // Reverse communication preparations
|
---|
424 | // I know it looks ugly, but it works the same way
|
---|
425 | // anywhere from C++ to Python.
|
---|
426 | //
|
---|
427 | // This code initializes locals by:
|
---|
428 | // * random values determined during code
|
---|
429 | // generation - on first subroutine call
|
---|
430 | // * values from previous call - on subsequent calls
|
---|
431 | //
|
---|
432 | if( state.rstate.stage>=0 )
|
---|
433 | {
|
---|
434 | n = state.rstate.ia[0];
|
---|
435 | m = state.rstate.ia[1];
|
---|
436 | i = state.rstate.ia[2];
|
---|
437 | lbfgsflags = state.rstate.ia[3];
|
---|
438 | spd = state.rstate.ba[0];
|
---|
439 | xnorm = state.rstate.ra[0];
|
---|
440 | stepnorm = state.rstate.ra[1];
|
---|
441 | fbase = state.rstate.ra[2];
|
---|
442 | fnew = state.rstate.ra[3];
|
---|
443 | lambda = state.rstate.ra[4];
|
---|
444 | nu = state.rstate.ra[5];
|
---|
445 | lambdaup = state.rstate.ra[6];
|
---|
446 | lambdadown = state.rstate.ra[7];
|
---|
447 | v = state.rstate.ra[8];
|
---|
448 | }
|
---|
449 | else
|
---|
450 | {
|
---|
451 | n = -983;
|
---|
452 | m = -989;
|
---|
453 | i = -834;
|
---|
454 | lbfgsflags = 900;
|
---|
455 | spd = true;
|
---|
456 | xnorm = 364;
|
---|
457 | stepnorm = 214;
|
---|
458 | fbase = -338;
|
---|
459 | fnew = -686;
|
---|
460 | lambda = 912;
|
---|
461 | nu = 585;
|
---|
462 | lambdaup = 497;
|
---|
463 | lambdadown = -271;
|
---|
464 | v = -581;
|
---|
465 | }
|
---|
466 | if( state.rstate.stage==0 )
|
---|
467 | {
|
---|
468 | goto lbl_0;
|
---|
469 | }
|
---|
470 | if( state.rstate.stage==1 )
|
---|
471 | {
|
---|
472 | goto lbl_1;
|
---|
473 | }
|
---|
474 | if( state.rstate.stage==2 )
|
---|
475 | {
|
---|
476 | goto lbl_2;
|
---|
477 | }
|
---|
478 | if( state.rstate.stage==3 )
|
---|
479 | {
|
---|
480 | goto lbl_3;
|
---|
481 | }
|
---|
482 | if( state.rstate.stage==4 )
|
---|
483 | {
|
---|
484 | goto lbl_4;
|
---|
485 | }
|
---|
486 | if( state.rstate.stage==5 )
|
---|
487 | {
|
---|
488 | goto lbl_5;
|
---|
489 | }
|
---|
490 | if( state.rstate.stage==6 )
|
---|
491 | {
|
---|
492 | goto lbl_6;
|
---|
493 | }
|
---|
494 |
|
---|
495 | //
|
---|
496 | // Routine body
|
---|
497 | //
|
---|
498 | System.Diagnostics.Debug.Assert(state.usermode==lmmodefj | state.usermode==lmmodefgj | state.usermode==lmmodefgh, "LM: internal error");
|
---|
499 | if( state.wrongparams )
|
---|
500 | {
|
---|
501 | state.repterminationtype = -1;
|
---|
502 | result = false;
|
---|
503 | return result;
|
---|
504 | }
|
---|
505 |
|
---|
506 | //
|
---|
507 | // prepare params
|
---|
508 | //
|
---|
509 | n = state.n;
|
---|
510 | m = state.m;
|
---|
511 | lambdaup = 10;
|
---|
512 | lambdadown = 0.3;
|
---|
513 | nu = 2;
|
---|
514 | lbfgsflags = 0;
|
---|
515 |
|
---|
516 | //
|
---|
517 | // if we have F and G
|
---|
518 | //
|
---|
519 | if( ! ((state.usermode==lmmodefgj | state.usermode==lmmodefgh) & state.flags/lmflagnoprelbfgs%2==0) )
|
---|
520 | {
|
---|
521 | goto lbl_7;
|
---|
522 | }
|
---|
523 |
|
---|
524 | //
|
---|
525 | // First stage of the hybrid algorithm: LBFGS
|
---|
526 | //
|
---|
527 | lbfgs.minlbfgs(n, Math.Min(n, lmprelbfgsm), ref state.x, 0.0, 0.0, 0.0, Math.Max(25, n), 0, ref state.internalstate);
|
---|
528 | lbl_9:
|
---|
529 | if( ! lbfgs.minlbfgsiteration(ref state.internalstate) )
|
---|
530 | {
|
---|
531 | goto lbl_10;
|
---|
532 | }
|
---|
533 |
|
---|
534 | //
|
---|
535 | // RComm
|
---|
536 | //
|
---|
537 | for(i_=0; i_<=n-1;i_++)
|
---|
538 | {
|
---|
539 | state.x[i_] = state.internalstate.x[i_];
|
---|
540 | }
|
---|
541 | lmclearrequestfields(ref state);
|
---|
542 | state.needfg = true;
|
---|
543 | state.rstate.stage = 0;
|
---|
544 | goto lbl_rcomm;
|
---|
545 | lbl_0:
|
---|
546 | state.repnfunc = state.repnfunc+1;
|
---|
547 | state.repngrad = state.repngrad+1;
|
---|
548 |
|
---|
549 | //
|
---|
550 | // Call LBFGS
|
---|
551 | //
|
---|
552 | state.internalstate.f = state.f;
|
---|
553 | for(i_=0; i_<=n-1;i_++)
|
---|
554 | {
|
---|
555 | state.internalstate.g[i_] = state.g[i_];
|
---|
556 | }
|
---|
557 | goto lbl_9;
|
---|
558 | lbl_10:
|
---|
559 | lbfgs.minlbfgsresults(ref state.internalstate, ref state.x, ref state.internalrep);
|
---|
560 | lbl_7:
|
---|
561 |
|
---|
562 | //
|
---|
563 | // Second stage of the hybrid algorithm: LM
|
---|
564 | // Initialize quadratic model.
|
---|
565 | //
|
---|
566 | if( state.usermode!=lmmodefgh )
|
---|
567 | {
|
---|
568 | goto lbl_11;
|
---|
569 | }
|
---|
570 |
|
---|
571 | //
|
---|
572 | // RComm
|
---|
573 | //
|
---|
574 | lmclearrequestfields(ref state);
|
---|
575 | state.needfgh = true;
|
---|
576 | state.rstate.stage = 1;
|
---|
577 | goto lbl_rcomm;
|
---|
578 | lbl_1:
|
---|
579 | state.repnfunc = state.repnfunc+1;
|
---|
580 | state.repngrad = state.repngrad+1;
|
---|
581 | state.repnhess = state.repnhess+1;
|
---|
582 |
|
---|
583 | //
|
---|
584 | // generate raw quadratic model
|
---|
585 | //
|
---|
586 | for(i=0; i<=n-1; i++)
|
---|
587 | {
|
---|
588 | for(i_=0; i_<=n-1;i_++)
|
---|
589 | {
|
---|
590 | state.rawmodel[i,i_] = state.h[i,i_];
|
---|
591 | }
|
---|
592 | }
|
---|
593 | for(i_=0; i_<=n-1;i_++)
|
---|
594 | {
|
---|
595 | state.gbase[i_] = state.g[i_];
|
---|
596 | }
|
---|
597 | fbase = state.f;
|
---|
598 | lbl_11:
|
---|
599 | if( ! (state.usermode==lmmodefgj | state.usermode==lmmodefj) )
|
---|
600 | {
|
---|
601 | goto lbl_13;
|
---|
602 | }
|
---|
603 |
|
---|
604 | //
|
---|
605 | // RComm
|
---|
606 | //
|
---|
607 | lmclearrequestfields(ref state);
|
---|
608 | state.needfij = true;
|
---|
609 | state.rstate.stage = 2;
|
---|
610 | goto lbl_rcomm;
|
---|
611 | lbl_2:
|
---|
612 | state.repnfunc = state.repnfunc+1;
|
---|
613 | state.repnjac = state.repnjac+1;
|
---|
614 |
|
---|
615 | //
|
---|
616 | // generate raw quadratic model
|
---|
617 | //
|
---|
618 | blas.matrixmatrixmultiply(ref state.j, 0, m-1, 0, n-1, true, ref state.j, 0, m-1, 0, n-1, false, 1.0, ref state.rawmodel, 0, n-1, 0, n-1, 0.0, ref state.work);
|
---|
619 | blas.matrixvectormultiply(ref state.j, 0, m-1, 0, n-1, true, ref state.fi, 0, m-1, 1.0, ref state.gbase, 0, n-1, 0.0);
|
---|
620 | fbase = 0.0;
|
---|
621 | for(i_=0; i_<=m-1;i_++)
|
---|
622 | {
|
---|
623 | fbase += state.fi[i_]*state.fi[i_];
|
---|
624 | }
|
---|
625 | lbl_13:
|
---|
626 | lambda = 0.001;
|
---|
627 | lbl_15:
|
---|
628 | if( false )
|
---|
629 | {
|
---|
630 | goto lbl_16;
|
---|
631 | }
|
---|
632 |
|
---|
633 | //
|
---|
634 | // 1. Model = RawModel+lambda*I
|
---|
635 | // 2. Try to solve (RawModel+Lambda*I)*dx = -g.
|
---|
636 | // Increase lambda if left part is not positive definite.
|
---|
637 | //
|
---|
638 | for(i=0; i<=n-1; i++)
|
---|
639 | {
|
---|
640 | for(i_=0; i_<=n-1;i_++)
|
---|
641 | {
|
---|
642 | state.model[i,i_] = state.rawmodel[i,i_];
|
---|
643 | }
|
---|
644 | state.model[i,i] = state.model[i,i]+lambda;
|
---|
645 | }
|
---|
646 | spd = cholesky.spdmatrixcholesky(ref state.model, n, true);
|
---|
647 | state.repncholesky = state.repncholesky+1;
|
---|
648 | if( !spd )
|
---|
649 | {
|
---|
650 | lambda = lambda*lambdaup*nu;
|
---|
651 | nu = nu*2;
|
---|
652 | goto lbl_15;
|
---|
653 | }
|
---|
654 | if( !spdsolve.spdmatrixcholeskysolve(ref state.model, state.gbase, n, true, ref state.xdir) )
|
---|
655 | {
|
---|
656 | lambda = lambda*lambdaup*nu;
|
---|
657 | nu = nu*2;
|
---|
658 | goto lbl_15;
|
---|
659 | }
|
---|
660 | for(i_=0; i_<=n-1;i_++)
|
---|
661 | {
|
---|
662 | state.xdir[i_] = -1*state.xdir[i_];
|
---|
663 | }
|
---|
664 |
|
---|
665 | //
|
---|
666 | // Candidate lambda found.
|
---|
667 | // 1. Save old w in WBase
|
---|
668 | // 1. Test some stopping criterions
|
---|
669 | // 2. If error(w+wdir)>error(w), increase lambda
|
---|
670 | //
|
---|
671 | for(i_=0; i_<=n-1;i_++)
|
---|
672 | {
|
---|
673 | state.xbase[i_] = state.x[i_];
|
---|
674 | }
|
---|
675 | for(i_=0; i_<=n-1;i_++)
|
---|
676 | {
|
---|
677 | state.x[i_] = state.x[i_] + state.xdir[i_];
|
---|
678 | }
|
---|
679 | xnorm = 0.0;
|
---|
680 | for(i_=0; i_<=n-1;i_++)
|
---|
681 | {
|
---|
682 | xnorm += state.xbase[i_]*state.xbase[i_];
|
---|
683 | }
|
---|
684 | stepnorm = 0.0;
|
---|
685 | for(i_=0; i_<=n-1;i_++)
|
---|
686 | {
|
---|
687 | stepnorm += state.xdir[i_]*state.xdir[i_];
|
---|
688 | }
|
---|
689 | xnorm = Math.Sqrt(xnorm);
|
---|
690 | stepnorm = Math.Sqrt(stepnorm);
|
---|
691 | if( (double)(stepnorm)<=(double)(state.epsx*(1+xnorm)) )
|
---|
692 | {
|
---|
693 |
|
---|
694 | //
|
---|
695 | // step size if small enough
|
---|
696 | //
|
---|
697 | state.repterminationtype = 2;
|
---|
698 | goto lbl_16;
|
---|
699 | }
|
---|
700 | lmclearrequestfields(ref state);
|
---|
701 | state.needf = true;
|
---|
702 | state.rstate.stage = 3;
|
---|
703 | goto lbl_rcomm;
|
---|
704 | lbl_3:
|
---|
705 | state.repnfunc = state.repnfunc+1;
|
---|
706 | fnew = state.f;
|
---|
707 | if( (double)(Math.Abs(fnew-fbase))<=(double)(state.epsf*Math.Max(1, Math.Max(Math.Abs(fbase), Math.Abs(fnew)))) )
|
---|
708 | {
|
---|
709 |
|
---|
710 | //
|
---|
711 | // function change is small enough
|
---|
712 | //
|
---|
713 | state.repterminationtype = 1;
|
---|
714 | goto lbl_16;
|
---|
715 | }
|
---|
716 | if( (double)(fnew)>(double)(fbase) )
|
---|
717 | {
|
---|
718 |
|
---|
719 | //
|
---|
720 | // restore state and continue out search for lambda
|
---|
721 | //
|
---|
722 | for(i_=0; i_<=n-1;i_++)
|
---|
723 | {
|
---|
724 | state.x[i_] = state.xbase[i_];
|
---|
725 | }
|
---|
726 | lambda = lambda*lambdaup*nu;
|
---|
727 | nu = nu*2;
|
---|
728 | goto lbl_15;
|
---|
729 | }
|
---|
730 | if( ! ((state.usermode==lmmodefgj | state.usermode==lmmodefgh) & state.flags/lmflagnointlbfgs%2==0) )
|
---|
731 | {
|
---|
732 | goto lbl_17;
|
---|
733 | }
|
---|
734 | System.Diagnostics.Debug.Assert(state.usermode==lmmodefgh);
|
---|
735 |
|
---|
736 | //
|
---|
737 | // Optimize using inv(cholesky(H)) as preconditioner
|
---|
738 | //
|
---|
739 | if( ! trinverse.rmatrixtrinverse(ref state.model, n, true, false) )
|
---|
740 | {
|
---|
741 | goto lbl_19;
|
---|
742 | }
|
---|
743 |
|
---|
744 | //
|
---|
745 | // if matrix can be inverted use it.
|
---|
746 | // just silently move to next iteration otherwise.
|
---|
747 | // (will be very, very rare, mostly for specially
|
---|
748 | // designed near-degenerate tasks)
|
---|
749 | //
|
---|
750 | for(i_=0; i_<=n-1;i_++)
|
---|
751 | {
|
---|
752 | state.xbase[i_] = state.x[i_];
|
---|
753 | }
|
---|
754 | for(i=0; i<=n-1; i++)
|
---|
755 | {
|
---|
756 | state.xprec[i] = 0;
|
---|
757 | }
|
---|
758 | lbfgs.minlbfgs(n, Math.Min(n, lmintlbfgsits), ref state.xprec, 0.0, 0.0, 0.0, lmintlbfgsits, lbfgsflags, ref state.internalstate);
|
---|
759 | lbl_21:
|
---|
760 | if( ! lbfgs.minlbfgsiteration(ref state.internalstate) )
|
---|
761 | {
|
---|
762 | goto lbl_22;
|
---|
763 | }
|
---|
764 |
|
---|
765 | //
|
---|
766 | // convert XPrec to unpreconditioned form, then call RComm.
|
---|
767 | //
|
---|
768 | for(i=0; i<=n-1; i++)
|
---|
769 | {
|
---|
770 | v = 0.0;
|
---|
771 | for(i_=i; i_<=n-1;i_++)
|
---|
772 | {
|
---|
773 | v += state.internalstate.x[i_]*state.model[i,i_];
|
---|
774 | }
|
---|
775 | state.x[i] = state.xbase[i]+v;
|
---|
776 | }
|
---|
777 | lmclearrequestfields(ref state);
|
---|
778 | state.needfg = true;
|
---|
779 | state.rstate.stage = 4;
|
---|
780 | goto lbl_rcomm;
|
---|
781 | lbl_4:
|
---|
782 | state.repnfunc = state.repnfunc+1;
|
---|
783 | state.repngrad = state.repngrad+1;
|
---|
784 |
|
---|
785 | //
|
---|
786 | // 1. pass State.F to State.InternalState.F
|
---|
787 | // 2. convert gradient back to preconditioned form
|
---|
788 | //
|
---|
789 | state.internalstate.f = state.f;
|
---|
790 | for(i=0; i<=n-1; i++)
|
---|
791 | {
|
---|
792 | state.internalstate.g[i] = 0;
|
---|
793 | }
|
---|
794 | for(i=0; i<=n-1; i++)
|
---|
795 | {
|
---|
796 | v = state.g[i];
|
---|
797 | for(i_=i; i_<=n-1;i_++)
|
---|
798 | {
|
---|
799 | state.internalstate.g[i_] = state.internalstate.g[i_] + v*state.model[i,i_];
|
---|
800 | }
|
---|
801 | }
|
---|
802 |
|
---|
803 | //
|
---|
804 | // next iteration
|
---|
805 | //
|
---|
806 | goto lbl_21;
|
---|
807 | lbl_22:
|
---|
808 |
|
---|
809 | //
|
---|
810 | // change LBFGS flags to NoRealloc.
|
---|
811 | // L-BFGS subroutine will use memory allocated from previous run.
|
---|
812 | // it is possible since all subsequent calls will be with same N/M.
|
---|
813 | //
|
---|
814 | lbfgsflags = lbfgsnorealloc;
|
---|
815 |
|
---|
816 | //
|
---|
817 | // back to unpreconditioned X
|
---|
818 | //
|
---|
819 | lbfgs.minlbfgsresults(ref state.internalstate, ref state.xprec, ref state.internalrep);
|
---|
820 | for(i=0; i<=n-1; i++)
|
---|
821 | {
|
---|
822 | v = 0.0;
|
---|
823 | for(i_=i; i_<=n-1;i_++)
|
---|
824 | {
|
---|
825 | v += state.xprec[i_]*state.model[i,i_];
|
---|
826 | }
|
---|
827 | state.x[i] = state.xbase[i]+v;
|
---|
828 | }
|
---|
829 | lbl_19:
|
---|
830 | lbl_17:
|
---|
831 |
|
---|
832 | //
|
---|
833 | // Accept new position.
|
---|
834 | // Calculate Hessian
|
---|
835 | //
|
---|
836 | if( state.usermode!=lmmodefgh )
|
---|
837 | {
|
---|
838 | goto lbl_23;
|
---|
839 | }
|
---|
840 |
|
---|
841 | //
|
---|
842 | // RComm
|
---|
843 | //
|
---|
844 | lmclearrequestfields(ref state);
|
---|
845 | state.needfgh = true;
|
---|
846 | state.rstate.stage = 5;
|
---|
847 | goto lbl_rcomm;
|
---|
848 | lbl_5:
|
---|
849 | state.repnfunc = state.repnfunc+1;
|
---|
850 | state.repngrad = state.repngrad+1;
|
---|
851 | state.repnhess = state.repnhess+1;
|
---|
852 |
|
---|
853 | //
|
---|
854 | // Update raw quadratic model
|
---|
855 | //
|
---|
856 | for(i=0; i<=n-1; i++)
|
---|
857 | {
|
---|
858 | for(i_=0; i_<=n-1;i_++)
|
---|
859 | {
|
---|
860 | state.rawmodel[i,i_] = state.h[i,i_];
|
---|
861 | }
|
---|
862 | }
|
---|
863 | for(i_=0; i_<=n-1;i_++)
|
---|
864 | {
|
---|
865 | state.gbase[i_] = state.g[i_];
|
---|
866 | }
|
---|
867 | fbase = state.f;
|
---|
868 | lbl_23:
|
---|
869 | if( ! (state.usermode==lmmodefgj | state.usermode==lmmodefj) )
|
---|
870 | {
|
---|
871 | goto lbl_25;
|
---|
872 | }
|
---|
873 |
|
---|
874 | //
|
---|
875 | // RComm
|
---|
876 | //
|
---|
877 | lmclearrequestfields(ref state);
|
---|
878 | state.needfij = true;
|
---|
879 | state.rstate.stage = 6;
|
---|
880 | goto lbl_rcomm;
|
---|
881 | lbl_6:
|
---|
882 | state.repnfunc = state.repnfunc+1;
|
---|
883 | state.repnjac = state.repnjac+1;
|
---|
884 |
|
---|
885 | //
|
---|
886 | // generate raw quadratic model
|
---|
887 | //
|
---|
888 | blas.matrixmatrixmultiply(ref state.j, 0, m-1, 0, n-1, true, ref state.j, 0, m-1, 0, n-1, false, 1.0, ref state.rawmodel, 0, n-1, 0, n-1, 0.0, ref state.work);
|
---|
889 | blas.matrixvectormultiply(ref state.j, 0, m-1, 0, n-1, true, ref state.fi, 0, m-1, 1.0, ref state.gbase, 0, n-1, 0.0);
|
---|
890 | fbase = 0.0;
|
---|
891 | for(i_=0; i_<=m-1;i_++)
|
---|
892 | {
|
---|
893 | fbase += state.fi[i_]*state.fi[i_];
|
---|
894 | }
|
---|
895 | lbl_25:
|
---|
896 | state.repiterationscount = state.repiterationscount+1;
|
---|
897 | if( state.repiterationscount>=state.maxits & state.maxits>0 )
|
---|
898 | {
|
---|
899 | state.repterminationtype = 5;
|
---|
900 | goto lbl_16;
|
---|
901 | }
|
---|
902 |
|
---|
903 | //
|
---|
904 | // Update lambda
|
---|
905 | //
|
---|
906 | lambda = lambda*lambdadown;
|
---|
907 | nu = 2;
|
---|
908 | goto lbl_15;
|
---|
909 | lbl_16:
|
---|
910 | result = false;
|
---|
911 | return result;
|
---|
912 |
|
---|
913 | //
|
---|
914 | // Saving state
|
---|
915 | //
|
---|
916 | lbl_rcomm:
|
---|
917 | result = true;
|
---|
918 | state.rstate.ia[0] = n;
|
---|
919 | state.rstate.ia[1] = m;
|
---|
920 | state.rstate.ia[2] = i;
|
---|
921 | state.rstate.ia[3] = lbfgsflags;
|
---|
922 | state.rstate.ba[0] = spd;
|
---|
923 | state.rstate.ra[0] = xnorm;
|
---|
924 | state.rstate.ra[1] = stepnorm;
|
---|
925 | state.rstate.ra[2] = fbase;
|
---|
926 | state.rstate.ra[3] = fnew;
|
---|
927 | state.rstate.ra[4] = lambda;
|
---|
928 | state.rstate.ra[5] = nu;
|
---|
929 | state.rstate.ra[6] = lambdaup;
|
---|
930 | state.rstate.ra[7] = lambdadown;
|
---|
931 | state.rstate.ra[8] = v;
|
---|
932 | return result;
|
---|
933 | }
|
---|
934 |
|
---|
935 |
|
---|
936 | /*************************************************************************
|
---|
937 | Levenberg-Marquardt algorithm results
|
---|
938 |
|
---|
939 | Called after MinLMIteration returned False.
|
---|
940 |
|
---|
941 | Input parameters:
|
---|
942 | State - algorithm state (used by MinLMIteration).
|
---|
943 |
|
---|
944 | Output parameters:
|
---|
945 | X - array[0..N-1], solution
|
---|
946 | Rep - optimization report:
|
---|
947 | * Rep.TerminationType completetion code:
|
---|
948 | * -1 incorrect parameters were specified
|
---|
949 | * 1 relative function improvement is no more than
|
---|
950 | EpsF.
|
---|
951 | * 2 relative step is no more than EpsX.
|
---|
952 | * 4 gradient norm is no more than EpsG
|
---|
953 | * 5 MaxIts steps was taken
|
---|
954 | * Rep.IterationsCount contains iterations count
|
---|
955 | * Rep.NFunc - number of function calculations
|
---|
956 | * Rep.NJac - number of Jacobi matrix calculations
|
---|
957 | * Rep.NGrad - number of gradient calculations
|
---|
958 | * Rep.NHess - number of Hessian calculations
|
---|
959 | * Rep.NCholesky - number of Cholesky decomposition calculations
|
---|
960 |
|
---|
961 | -- ALGLIB --
|
---|
962 | Copyright 10.03.2009 by Bochkanov Sergey
|
---|
963 | *************************************************************************/
|
---|
964 | public static void minlmresults(ref lmstate state,
|
---|
965 | ref double[] x,
|
---|
966 | ref lmreport rep)
|
---|
967 | {
|
---|
968 | int i_ = 0;
|
---|
969 |
|
---|
970 | x = new double[state.n-1+1];
|
---|
971 | for(i_=0; i_<=state.n-1;i_++)
|
---|
972 | {
|
---|
973 | x[i_] = state.x[i_];
|
---|
974 | }
|
---|
975 | rep.iterationscount = state.repiterationscount;
|
---|
976 | rep.terminationtype = state.repterminationtype;
|
---|
977 | rep.nfunc = state.repnfunc;
|
---|
978 | rep.njac = state.repnjac;
|
---|
979 | rep.ngrad = state.repngrad;
|
---|
980 | rep.nhess = state.repnhess;
|
---|
981 | rep.ncholesky = state.repncholesky;
|
---|
982 | }
|
---|
983 |
|
---|
984 |
|
---|
985 | /*************************************************************************
|
---|
986 | Prepare internal structures (except for RComm).
|
---|
987 |
|
---|
988 | Note: M must be zero for FGH mode, non-zero for FJ/FGJ mode.
|
---|
989 | *************************************************************************/
|
---|
990 | private static void lmprepare(int n,
|
---|
991 | int m,
|
---|
992 | bool havegrad,
|
---|
993 | ref lmstate state)
|
---|
994 | {
|
---|
995 | state.repiterationscount = 0;
|
---|
996 | state.repterminationtype = 0;
|
---|
997 | state.repnfunc = 0;
|
---|
998 | state.repnjac = 0;
|
---|
999 | state.repngrad = 0;
|
---|
1000 | state.repnhess = 0;
|
---|
1001 | state.repncholesky = 0;
|
---|
1002 | if( n<0 | m<0 )
|
---|
1003 | {
|
---|
1004 | return;
|
---|
1005 | }
|
---|
1006 | if( havegrad )
|
---|
1007 | {
|
---|
1008 | state.g = new double[n-1+1];
|
---|
1009 | }
|
---|
1010 | if( m!=0 )
|
---|
1011 | {
|
---|
1012 | state.j = new double[m-1+1, n-1+1];
|
---|
1013 | state.fi = new double[m-1+1];
|
---|
1014 | state.h = new double[0+1, 0+1];
|
---|
1015 | }
|
---|
1016 | else
|
---|
1017 | {
|
---|
1018 | state.j = new double[0+1, 0+1];
|
---|
1019 | state.fi = new double[0+1];
|
---|
1020 | state.h = new double[n-1+1, n-1+1];
|
---|
1021 | }
|
---|
1022 | state.x = new double[n-1+1];
|
---|
1023 | state.rawmodel = new double[n-1+1, n-1+1];
|
---|
1024 | state.model = new double[n-1+1, n-1+1];
|
---|
1025 | state.xbase = new double[n-1+1];
|
---|
1026 | state.xprec = new double[n-1+1];
|
---|
1027 | state.gbase = new double[n-1+1];
|
---|
1028 | state.xdir = new double[n-1+1];
|
---|
1029 | state.work = new double[Math.Max(n, m)+1];
|
---|
1030 | }
|
---|
1031 |
|
---|
1032 |
|
---|
1033 | /*************************************************************************
|
---|
1034 | Clears request fileds (to be sure that we don't forgot to clear something)
|
---|
1035 | *************************************************************************/
|
---|
1036 | private static void lmclearrequestfields(ref lmstate state)
|
---|
1037 | {
|
---|
1038 | state.needf = false;
|
---|
1039 | state.needfg = false;
|
---|
1040 | state.needfgh = false;
|
---|
1041 | state.needfij = false;
|
---|
1042 | }
|
---|
1043 | }
|
---|
1044 | }
|
---|