xref: /petsc/src/tao/unconstrained/impls/ntl/ntl.c (revision 117ef88edefbfc12e7c19efe87a19a2e1b0acd4f)
1 #include <../src/tao/unconstrained/impls/ntl/ntlimpl.h>
2 
3 #include <petscksp.h>
4 
5 #define NTL_INIT_CONSTANT         0
6 #define NTL_INIT_DIRECTION        1
7 #define NTL_INIT_INTERPOLATION    2
8 #define NTL_INIT_TYPES            3
9 
10 #define NTL_UPDATE_REDUCTION      0
11 #define NTL_UPDATE_INTERPOLATION  1
12 #define NTL_UPDATE_TYPES          2
13 
14 static const char *NTL_INIT[64] = {"constant","direction","interpolation"};
15 
16 static const char *NTL_UPDATE[64] = {"reduction","interpolation"};
17 
18 /* Implements Newton's Method with a trust-region, line-search approach for
19    solving unconstrained minimization problems.  A More'-Thuente line search
20    is used to guarantee that the bfgs preconditioner remains positive
21    definite. */
22 
23 #define NTL_NEWTON              0
24 #define NTL_BFGS                1
25 #define NTL_SCALED_GRADIENT     2
26 #define NTL_GRADIENT            3
27 
28 static PetscErrorCode TaoSolve_NTL(Tao tao)
29 {
30   TAO_NTL                      *tl = (TAO_NTL *)tao->data;
31   KSPType                      ksp_type;
32   PetscBool                    is_nash,is_stcg,is_gltr,is_bfgs,is_jacobi,is_symmetric,sym_set;
33   KSPConvergedReason           ksp_reason;
34   PC                           pc;
35   TaoLineSearchConvergedReason ls_reason;
36 
37   PetscReal                    fmin, ftrial, prered, actred, kappa, sigma;
38   PetscReal                    tau, tau_1, tau_2, tau_max, tau_min, max_radius;
39   PetscReal                    f, fold, gdx, gnorm;
40   PetscReal                    step = 1.0;
41 
42   PetscReal                    norm_d = 0.0;
43   PetscErrorCode               ierr;
44   PetscInt                     stepType;
45   PetscInt                     its;
46 
47   PetscInt                     bfgsUpdates = 0;
48   PetscInt                     needH;
49 
50   PetscInt                     i_max = 5;
51   PetscInt                     j_max = 1;
52   PetscInt                     i, j, n, N;
53 
54   PetscInt                     tr_reject;
55 
56   PetscFunctionBegin;
57   if (tao->XL || tao->XU || tao->ops->computebounds) {
58     ierr = PetscInfo(tao,"WARNING: Variable bounds have been set but will be ignored by ntl algorithm\n");CHKERRQ(ierr);
59   }
60 
61   ierr = KSPGetType(tao->ksp,&ksp_type);CHKERRQ(ierr);
62   ierr = PetscStrcmp(ksp_type,KSPNASH,&is_nash);CHKERRQ(ierr);
63   ierr = PetscStrcmp(ksp_type,KSPSTCG,&is_stcg);CHKERRQ(ierr);
64   ierr = PetscStrcmp(ksp_type,KSPGLTR,&is_gltr);CHKERRQ(ierr);
65   if (!is_nash && !is_stcg && !is_gltr) SETERRQ(PetscObjectComm((PetscObject)tao),PETSC_ERR_SUP,"TAO_NTR requires nash, stcg, or gltr for the KSP");
66 
67   /* Initialize the radius and modify if it is too large or small */
68   tao->trust = tao->trust0;
69   tao->trust = PetscMax(tao->trust, tl->min_radius);
70   tao->trust = PetscMin(tao->trust, tl->max_radius);
71 
72   /* Allocate the vectors needed for the BFGS approximation */
73   ierr = KSPGetPC(tao->ksp, &pc);CHKERRQ(ierr);
74   ierr = PetscObjectTypeCompare((PetscObject)pc, PCLMVM, &is_bfgs);CHKERRQ(ierr);
75   ierr = PetscObjectTypeCompare((PetscObject)pc, PCJACOBI, &is_jacobi);CHKERRQ(ierr);
76   if (is_bfgs) {
77     tl->bfgs_pre = pc;
78     ierr = PCLMVMGetMatLMVM(tl->bfgs_pre, &tl->M);CHKERRQ(ierr);
79     ierr = VecGetLocalSize(tao->solution, &n);CHKERRQ(ierr);
80     ierr = VecGetSize(tao->solution, &N);CHKERRQ(ierr);
81     ierr = MatSetSizes(tl->M, n, n, N, N);CHKERRQ(ierr);
82     ierr = MatLMVMAllocate(tl->M, tao->solution, tao->gradient);CHKERRQ(ierr);
83     ierr = MatIsSymmetricKnown(tl->M, &sym_set, &is_symmetric);CHKERRQ(ierr);
84     if (!sym_set || !is_symmetric) SETERRQ(PetscObjectComm((PetscObject)tao), PETSC_ERR_ARG_INCOMP, "LMVM matrix in the LMVM preconditioner must be symmetric.");
85   } else if (is_jacobi) {
86     ierr = PCJacobiSetUseAbs(pc,PETSC_TRUE);CHKERRQ(ierr);
87   }
88 
89   /* Check convergence criteria */
90   ierr = TaoComputeObjectiveAndGradient(tao, tao->solution, &f, tao->gradient);CHKERRQ(ierr);
91   ierr = VecNorm(tao->gradient, NORM_2, &gnorm);CHKERRQ(ierr);
92 if (PetscIsInfOrNanReal(f) || PetscIsInfOrNanReal(gnorm)) SETERRQ(PetscObjectComm((PetscObject)tao),PETSC_ERR_USER, "User provided compute function generated Inf or NaN");
93   needH = 1;
94 
95   tao->reason = TAO_CONTINUE_ITERATING;
96   ierr = TaoLogConvergenceHistory(tao,f,gnorm,0.0,tao->ksp_its);CHKERRQ(ierr);
97   ierr = TaoMonitor(tao,tao->niter,f,gnorm,0.0,step);CHKERRQ(ierr);
98   ierr = (*tao->ops->convergencetest)(tao,tao->cnvP);CHKERRQ(ierr);
99   if (tao->reason != TAO_CONTINUE_ITERATING) PetscFunctionReturn(0);
100 
101   /* Initialize trust-region radius */
102   switch(tl->init_type) {
103   case NTL_INIT_CONSTANT:
104     /* Use the initial radius specified */
105     break;
106 
107   case NTL_INIT_INTERPOLATION:
108     /* Use the initial radius specified */
109     max_radius = 0.0;
110 
111     for (j = 0; j < j_max; ++j) {
112       fmin = f;
113       sigma = 0.0;
114 
115       if (needH) {
116         ierr = TaoComputeHessian(tao,tao->solution,tao->hessian,tao->hessian_pre);CHKERRQ(ierr);
117         needH = 0;
118       }
119 
120       for (i = 0; i < i_max; ++i) {
121         ierr = VecCopy(tao->solution, tl->W);CHKERRQ(ierr);
122         ierr = VecAXPY(tl->W, -tao->trust/gnorm, tao->gradient);CHKERRQ(ierr);
123 
124         ierr = TaoComputeObjective(tao, tl->W, &ftrial);CHKERRQ(ierr);
125         if (PetscIsInfOrNanReal(ftrial)) {
126           tau = tl->gamma1_i;
127         } else {
128           if (ftrial < fmin) {
129             fmin = ftrial;
130             sigma = -tao->trust / gnorm;
131           }
132 
133           ierr = MatMult(tao->hessian, tao->gradient, tao->stepdirection);CHKERRQ(ierr);
134           ierr = VecDot(tao->gradient, tao->stepdirection, &prered);CHKERRQ(ierr);
135 
136           prered = tao->trust * (gnorm - 0.5 * tao->trust * prered / (gnorm * gnorm));
137           actred = f - ftrial;
138           if ((PetscAbsScalar(actred) <= tl->epsilon) && (PetscAbsScalar(prered) <= tl->epsilon)) {
139             kappa = 1.0;
140           } else {
141             kappa = actred / prered;
142           }
143 
144           tau_1 = tl->theta_i * gnorm * tao->trust / (tl->theta_i * gnorm * tao->trust + (1.0 - tl->theta_i) * prered - actred);
145           tau_2 = tl->theta_i * gnorm * tao->trust / (tl->theta_i * gnorm * tao->trust - (1.0 + tl->theta_i) * prered + actred);
146           tau_min = PetscMin(tau_1, tau_2);
147           tau_max = PetscMax(tau_1, tau_2);
148 
149           if (PetscAbsScalar(kappa - (PetscReal)1.0) <= tl->mu1_i) {
150             /* Great agreement */
151             max_radius = PetscMax(max_radius, tao->trust);
152 
153             if (tau_max < 1.0) {
154               tau = tl->gamma3_i;
155             } else if (tau_max > tl->gamma4_i) {
156               tau = tl->gamma4_i;
157             } else if (tau_1 >= 1.0 && tau_1 <= tl->gamma4_i && tau_2 < 1.0) {
158               tau = tau_1;
159             } else if (tau_2 >= 1.0 && tau_2 <= tl->gamma4_i && tau_1 < 1.0) {
160               tau = tau_2;
161             } else {
162               tau = tau_max;
163             }
164           } else if (PetscAbsScalar(kappa - (PetscReal)1.0) <= tl->mu2_i) {
165             /* Good agreement */
166             max_radius = PetscMax(max_radius, tao->trust);
167 
168             if (tau_max < tl->gamma2_i) {
169               tau = tl->gamma2_i;
170             } else if (tau_max > tl->gamma3_i) {
171               tau = tl->gamma3_i;
172             } else {
173               tau = tau_max;
174             }
175           } else {
176             /* Not good agreement */
177             if (tau_min > 1.0) {
178               tau = tl->gamma2_i;
179             } else if (tau_max < tl->gamma1_i) {
180               tau = tl->gamma1_i;
181             } else if ((tau_min < tl->gamma1_i) && (tau_max >= 1.0)) {
182               tau = tl->gamma1_i;
183             } else if ((tau_1 >= tl->gamma1_i) && (tau_1 < 1.0) &&  ((tau_2 < tl->gamma1_i) || (tau_2 >= 1.0))) {
184               tau = tau_1;
185             } else if ((tau_2 >= tl->gamma1_i) && (tau_2 < 1.0) &&  ((tau_1 < tl->gamma1_i) || (tau_2 >= 1.0))) {
186               tau = tau_2;
187             } else {
188               tau = tau_max;
189             }
190           }
191         }
192         tao->trust = tau * tao->trust;
193       }
194 
195       if (fmin < f) {
196         f = fmin;
197         ierr = VecAXPY(tao->solution, sigma, tao->gradient);CHKERRQ(ierr);
198         ierr = TaoComputeGradient(tao, tao->solution, tao->gradient);CHKERRQ(ierr);
199 
200         ierr = VecNorm(tao->gradient, NORM_2, &gnorm);CHKERRQ(ierr);
201         if (PetscIsInfOrNanReal(f) || PetscIsInfOrNanReal(gnorm)) SETERRQ(PetscObjectComm((PetscObject)tao),PETSC_ERR_USER, "User provided compute function generated Inf or NaN");
202         needH = 1;
203 
204         ierr = TaoLogConvergenceHistory(tao,f,gnorm,0.0,tao->ksp_its);CHKERRQ(ierr);
205         ierr = TaoMonitor(tao,tao->niter,f,gnorm,0.0,step);CHKERRQ(ierr);
206         ierr = (*tao->ops->convergencetest)(tao,tao->cnvP);CHKERRQ(ierr);
207         if (tao->reason != TAO_CONTINUE_ITERATING) PetscFunctionReturn(0);
208       }
209     }
210     tao->trust = PetscMax(tao->trust, max_radius);
211 
212     /* Modify the radius if it is too large or small */
213     tao->trust = PetscMax(tao->trust, tl->min_radius);
214     tao->trust = PetscMin(tao->trust, tl->max_radius);
215     break;
216 
217   default:
218     /* Norm of the first direction will initialize radius */
219     tao->trust = 0.0;
220     break;
221   }
222 
223   /* Set counter for gradient/reset steps */
224   tl->ntrust = 0;
225   tl->newt = 0;
226   tl->bfgs = 0;
227   tl->grad = 0;
228 
229   /* Have not converged; continue with Newton method */
230   while (tao->reason == TAO_CONTINUE_ITERATING) {
231     /* Call general purpose update function */
232     if (tao->ops->update) {
233       ierr = (*tao->ops->update)(tao, tao->niter, tao->user_update);CHKERRQ(ierr);
234     }
235     ++tao->niter;
236     tao->ksp_its=0;
237     /* Compute the Hessian */
238     if (needH) {
239       ierr = TaoComputeHessian(tao,tao->solution,tao->hessian,tao->hessian_pre);CHKERRQ(ierr);
240     }
241 
242     if (tl->bfgs_pre) {
243       /* Update the limited memory preconditioner */
244       ierr = MatLMVMUpdate(tl->M,tao->solution, tao->gradient);CHKERRQ(ierr);
245       ++bfgsUpdates;
246     }
247     ierr = KSPSetOperators(tao->ksp, tao->hessian, tao->hessian_pre);CHKERRQ(ierr);
248     /* Solve the Newton system of equations */
249     ierr = KSPCGSetRadius(tao->ksp,tl->max_radius);CHKERRQ(ierr);
250     ierr = KSPSolve(tao->ksp, tao->gradient, tao->stepdirection);CHKERRQ(ierr);
251     ierr = KSPGetIterationNumber(tao->ksp,&its);CHKERRQ(ierr);
252     tao->ksp_its+=its;
253     tao->ksp_tot_its+=its;
254     ierr = KSPCGGetNormD(tao->ksp, &norm_d);CHKERRQ(ierr);
255 
256     if (0.0 == tao->trust) {
257       /* Radius was uninitialized; use the norm of the direction */
258       if (norm_d > 0.0) {
259         tao->trust = norm_d;
260 
261         /* Modify the radius if it is too large or small */
262         tao->trust = PetscMax(tao->trust, tl->min_radius);
263         tao->trust = PetscMin(tao->trust, tl->max_radius);
264       } else {
265         /* The direction was bad; set radius to default value and re-solve
266            the trust-region subproblem to get a direction */
267         tao->trust = tao->trust0;
268 
269         /* Modify the radius if it is too large or small */
270         tao->trust = PetscMax(tao->trust, tl->min_radius);
271         tao->trust = PetscMin(tao->trust, tl->max_radius);
272 
273         ierr = KSPCGSetRadius(tao->ksp,tl->max_radius);CHKERRQ(ierr);
274         ierr = KSPSolve(tao->ksp, tao->gradient, tao->stepdirection);CHKERRQ(ierr);
275         ierr = KSPGetIterationNumber(tao->ksp,&its);CHKERRQ(ierr);
276         tao->ksp_its+=its;
277         tao->ksp_tot_its+=its;
278         ierr = KSPCGGetNormD(tao->ksp, &norm_d);CHKERRQ(ierr);
279 
280         if (norm_d == 0.0) SETERRQ(PetscObjectComm((PetscObject)tao),PETSC_ERR_PLIB, "Initial direction zero");
281       }
282     }
283 
284     ierr = VecScale(tao->stepdirection, -1.0);CHKERRQ(ierr);
285     ierr = KSPGetConvergedReason(tao->ksp, &ksp_reason);CHKERRQ(ierr);
286     if ((KSP_DIVERGED_INDEFINITE_PC == ksp_reason) && (tl->bfgs_pre)) {
287       /* Preconditioner is numerically indefinite; reset the
288          approximate if using BFGS preconditioning. */
289       ierr = MatLMVMReset(tl->M, PETSC_FALSE);CHKERRQ(ierr);
290       ierr = MatLMVMUpdate(tl->M, tao->solution, tao->gradient);CHKERRQ(ierr);
291       bfgsUpdates = 1;
292     }
293 
294     /* Check trust-region reduction conditions */
295     tr_reject = 0;
296     if (NTL_UPDATE_REDUCTION == tl->update_type) {
297       /* Get predicted reduction */
298       ierr = KSPCGGetObjFcn(tao->ksp,&prered);CHKERRQ(ierr);
299       if (prered >= 0.0) {
300         /* The predicted reduction has the wrong sign.  This cannot
301            happen in infinite precision arithmetic.  Step should
302            be rejected! */
303         tao->trust = tl->alpha1 * PetscMin(tao->trust, norm_d);
304         tr_reject = 1;
305       } else {
306         /* Compute trial step and function value */
307         ierr = VecCopy(tao->solution, tl->W);CHKERRQ(ierr);
308         ierr = VecAXPY(tl->W, 1.0, tao->stepdirection);CHKERRQ(ierr);
309         ierr = TaoComputeObjective(tao, tl->W, &ftrial);CHKERRQ(ierr);
310 
311         if (PetscIsInfOrNanReal(ftrial)) {
312           tao->trust = tl->alpha1 * PetscMin(tao->trust, norm_d);
313           tr_reject = 1;
314         } else {
315           /* Compute and actual reduction */
316           actred = f - ftrial;
317           prered = -prered;
318           if ((PetscAbsScalar(actred) <= tl->epsilon) &&
319               (PetscAbsScalar(prered) <= tl->epsilon)) {
320             kappa = 1.0;
321           } else {
322             kappa = actred / prered;
323           }
324 
325           /* Accept of reject the step and update radius */
326           if (kappa < tl->eta1) {
327             /* Reject the step */
328             tao->trust = tl->alpha1 * PetscMin(tao->trust, norm_d);
329             tr_reject = 1;
330           } else {
331             /* Accept the step */
332             if (kappa < tl->eta2) {
333               /* Marginal bad step */
334               tao->trust = tl->alpha2 * PetscMin(tao->trust, norm_d);
335             } else if (kappa < tl->eta3) {
336               /* Reasonable step */
337               tao->trust = tl->alpha3 * tao->trust;
338             } else if (kappa < tl->eta4) {
339               /* Good step */
340               tao->trust = PetscMax(tl->alpha4 * norm_d, tao->trust);
341             } else {
342               /* Very good step */
343               tao->trust = PetscMax(tl->alpha5 * norm_d, tao->trust);
344             }
345           }
346         }
347       }
348     } else {
349       /* Get predicted reduction */
350       ierr = KSPCGGetObjFcn(tao->ksp,&prered);CHKERRQ(ierr);
351       if (prered >= 0.0) {
352         /* The predicted reduction has the wrong sign.  This cannot
353            happen in infinite precision arithmetic.  Step should
354            be rejected! */
355         tao->trust = tl->gamma1 * PetscMin(tao->trust, norm_d);
356         tr_reject = 1;
357       } else {
358         ierr = VecCopy(tao->solution, tl->W);CHKERRQ(ierr);
359         ierr = VecAXPY(tl->W, 1.0, tao->stepdirection);CHKERRQ(ierr);
360         ierr = TaoComputeObjective(tao, tl->W, &ftrial);CHKERRQ(ierr);
361         if (PetscIsInfOrNanReal(ftrial)) {
362           tao->trust = tl->gamma1 * PetscMin(tao->trust, norm_d);
363           tr_reject = 1;
364         } else {
365           ierr = VecDot(tao->gradient, tao->stepdirection, &gdx);CHKERRQ(ierr);
366 
367           actred = f - ftrial;
368           prered = -prered;
369           if ((PetscAbsScalar(actred) <= tl->epsilon) &&
370               (PetscAbsScalar(prered) <= tl->epsilon)) {
371             kappa = 1.0;
372           } else {
373             kappa = actred / prered;
374           }
375 
376           tau_1 = tl->theta * gdx / (tl->theta * gdx - (1.0 - tl->theta) * prered + actred);
377           tau_2 = tl->theta * gdx / (tl->theta * gdx + (1.0 + tl->theta) * prered - actred);
378           tau_min = PetscMin(tau_1, tau_2);
379           tau_max = PetscMax(tau_1, tau_2);
380 
381           if (kappa >= 1.0 - tl->mu1) {
382             /* Great agreement; accept step and update radius */
383             if (tau_max < 1.0) {
384               tao->trust = PetscMax(tao->trust, tl->gamma3 * norm_d);
385             } else if (tau_max > tl->gamma4) {
386               tao->trust = PetscMax(tao->trust, tl->gamma4 * norm_d);
387             } else {
388               tao->trust = PetscMax(tao->trust, tau_max * norm_d);
389             }
390           } else if (kappa >= 1.0 - tl->mu2) {
391             /* Good agreement */
392 
393             if (tau_max < tl->gamma2) {
394               tao->trust = tl->gamma2 * PetscMin(tao->trust, norm_d);
395             } else if (tau_max > tl->gamma3) {
396               tao->trust = PetscMax(tao->trust, tl->gamma3 * norm_d);
397             } else if (tau_max < 1.0) {
398               tao->trust = tau_max * PetscMin(tao->trust, norm_d);
399             } else {
400               tao->trust = PetscMax(tao->trust, tau_max * norm_d);
401             }
402           } else {
403             /* Not good agreement */
404             if (tau_min > 1.0) {
405               tao->trust = tl->gamma2 * PetscMin(tao->trust, norm_d);
406             } else if (tau_max < tl->gamma1) {
407               tao->trust = tl->gamma1 * PetscMin(tao->trust, norm_d);
408             } else if ((tau_min < tl->gamma1) && (tau_max >= 1.0)) {
409               tao->trust = tl->gamma1 * PetscMin(tao->trust, norm_d);
410             } else if ((tau_1 >= tl->gamma1) && (tau_1 < 1.0) && ((tau_2 < tl->gamma1) || (tau_2 >= 1.0))) {
411               tao->trust = tau_1 * PetscMin(tao->trust, norm_d);
412             } else if ((tau_2 >= tl->gamma1) && (tau_2 < 1.0) && ((tau_1 < tl->gamma1) || (tau_2 >= 1.0))) {
413               tao->trust = tau_2 * PetscMin(tao->trust, norm_d);
414             } else {
415               tao->trust = tau_max * PetscMin(tao->trust, norm_d);
416             }
417             tr_reject = 1;
418           }
419         }
420       }
421     }
422 
423     if (tr_reject) {
424       /* The trust-region constraints rejected the step.  Apply a linesearch.
425          Check for descent direction. */
426       ierr = VecDot(tao->stepdirection, tao->gradient, &gdx);CHKERRQ(ierr);
427       if ((gdx >= 0.0) || PetscIsInfOrNanReal(gdx)) {
428         /* Newton step is not descent or direction produced Inf or NaN */
429 
430         if (!tl->bfgs_pre) {
431           /* We don't have the bfgs matrix around and updated
432              Must use gradient direction in this case */
433           ierr = VecCopy(tao->gradient, tao->stepdirection);CHKERRQ(ierr);
434           ierr = VecScale(tao->stepdirection, -1.0);CHKERRQ(ierr);
435           ++tl->grad;
436           stepType = NTL_GRADIENT;
437         } else {
438           /* Attempt to use the BFGS direction */
439           ierr = MatSolve(tl->M, tao->gradient, tao->stepdirection);CHKERRQ(ierr);
440           ierr = VecScale(tao->stepdirection, -1.0);CHKERRQ(ierr);
441 
442           /* Check for success (descent direction) */
443           ierr = VecDot(tao->stepdirection, tao->gradient, &gdx);CHKERRQ(ierr);
444           if ((gdx >= 0) || PetscIsInfOrNanReal(gdx)) {
445             /* BFGS direction is not descent or direction produced not a number
446                We can assert bfgsUpdates > 1 in this case because
447                the first solve produces the scaled gradient direction,
448                which is guaranteed to be descent */
449 
450             /* Use steepest descent direction (scaled) */
451             ierr = MatLMVMReset(tl->M, PETSC_FALSE);CHKERRQ(ierr);
452             ierr = MatLMVMUpdate(tl->M, tao->solution, tao->gradient);CHKERRQ(ierr);
453             ierr = MatSolve(tl->M, tao->gradient, tao->stepdirection);CHKERRQ(ierr);
454             ierr = VecScale(tao->stepdirection, -1.0);CHKERRQ(ierr);
455 
456             bfgsUpdates = 1;
457             ++tl->grad;
458             stepType = NTL_GRADIENT;
459           } else {
460             ierr = MatLMVMGetUpdateCount(tl->M, &bfgsUpdates);CHKERRQ(ierr);
461             if (1 == bfgsUpdates) {
462               /* The first BFGS direction is always the scaled gradient */
463               ++tl->grad;
464               stepType = NTL_GRADIENT;
465             } else {
466               ++tl->bfgs;
467               stepType = NTL_BFGS;
468             }
469           }
470         }
471       } else {
472         /* Computed Newton step is descent */
473         ++tl->newt;
474         stepType = NTL_NEWTON;
475       }
476 
477       /* Perform the linesearch */
478       fold = f;
479       ierr = VecCopy(tao->solution, tl->Xold);CHKERRQ(ierr);
480       ierr = VecCopy(tao->gradient, tl->Gold);CHKERRQ(ierr);
481 
482       step = 1.0;
483       ierr = TaoLineSearchApply(tao->linesearch, tao->solution, &f, tao->gradient, tao->stepdirection, &step, &ls_reason);CHKERRQ(ierr);
484       ierr = TaoAddLineSearchCounts(tao);CHKERRQ(ierr);
485 
486       while (ls_reason != TAOLINESEARCH_SUCCESS && ls_reason != TAOLINESEARCH_SUCCESS_USER && stepType != NTL_GRADIENT) {      /* Linesearch failed */
487         /* Linesearch failed */
488         f = fold;
489         ierr = VecCopy(tl->Xold, tao->solution);CHKERRQ(ierr);
490         ierr = VecCopy(tl->Gold, tao->gradient);CHKERRQ(ierr);
491 
492         switch(stepType) {
493         case NTL_NEWTON:
494           /* Failed to obtain acceptable iterate with Newton step */
495 
496           if (tl->bfgs_pre) {
497             /* We don't have the bfgs matrix around and being updated
498                Must use gradient direction in this case */
499             ierr = VecCopy(tao->gradient, tao->stepdirection);CHKERRQ(ierr);
500             ++tl->grad;
501             stepType = NTL_GRADIENT;
502           } else {
503             /* Attempt to use the BFGS direction */
504             ierr = MatSolve(tl->M, tao->gradient, tao->stepdirection);CHKERRQ(ierr);
505 
506 
507             /* Check for success (descent direction) */
508             ierr = VecDot(tao->stepdirection, tao->gradient, &gdx);CHKERRQ(ierr);
509             if ((gdx <= 0) || PetscIsInfOrNanReal(gdx)) {
510               /* BFGS direction is not descent or direction produced
511                  not a number.  We can assert bfgsUpdates > 1 in this case
512                  Use steepest descent direction (scaled) */
513               ierr = MatLMVMReset(tl->M, PETSC_FALSE);CHKERRQ(ierr);
514               ierr = MatLMVMUpdate(tl->M, tao->solution, tao->gradient);CHKERRQ(ierr);
515               ierr = MatSolve(tl->M, tao->gradient, tao->stepdirection);CHKERRQ(ierr);
516 
517               bfgsUpdates = 1;
518               ++tl->grad;
519               stepType = NTL_GRADIENT;
520             } else {
521               ierr = MatLMVMGetUpdateCount(tl->M, &bfgsUpdates);CHKERRQ(ierr);
522               if (1 == bfgsUpdates) {
523                 /* The first BFGS direction is always the scaled gradient */
524                 ++tl->grad;
525                 stepType = NTL_GRADIENT;
526               } else {
527                 ++tl->bfgs;
528                 stepType = NTL_BFGS;
529               }
530             }
531           }
532           break;
533 
534         case NTL_BFGS:
535           /* Can only enter if pc_type == NTL_PC_BFGS
536              Failed to obtain acceptable iterate with BFGS step
537              Attempt to use the scaled gradient direction */
538           ierr = MatLMVMReset(tl->M, PETSC_FALSE);CHKERRQ(ierr);
539           ierr = MatLMVMUpdate(tl->M, tao->solution, tao->gradient);CHKERRQ(ierr);
540           ierr = MatSolve(tl->M, tao->gradient, tao->stepdirection);CHKERRQ(ierr);
541 
542           bfgsUpdates = 1;
543           ++tl->grad;
544           stepType = NTL_GRADIENT;
545           break;
546         }
547         ierr = VecScale(tao->stepdirection, -1.0);CHKERRQ(ierr);
548 
549         /* This may be incorrect; linesearch has values for stepmax and stepmin
550            that should be reset. */
551         step = 1.0;
552         ierr = TaoLineSearchApply(tao->linesearch, tao->solution, &f, tao->gradient, tao->stepdirection, &step, &ls_reason);CHKERRQ(ierr);
553         ierr = TaoAddLineSearchCounts(tao);CHKERRQ(ierr);
554       }
555 
556       if (ls_reason != TAOLINESEARCH_SUCCESS && ls_reason != TAOLINESEARCH_SUCCESS_USER) {
557         /* Failed to find an improving point */
558         f = fold;
559         ierr = VecCopy(tl->Xold, tao->solution);CHKERRQ(ierr);
560         ierr = VecCopy(tl->Gold, tao->gradient);CHKERRQ(ierr);
561         tao->trust = 0.0;
562         step = 0.0;
563         tao->reason = TAO_DIVERGED_LS_FAILURE;
564         break;
565       } else if (stepType == NTL_NEWTON) {
566         if (step < tl->nu1) {
567           /* Very bad step taken; reduce radius */
568           tao->trust = tl->omega1 * PetscMin(norm_d, tao->trust);
569         } else if (step < tl->nu2) {
570           /* Reasonably bad step taken; reduce radius */
571           tao->trust = tl->omega2 * PetscMin(norm_d, tao->trust);
572         } else if (step < tl->nu3) {
573           /* Reasonable step was taken; leave radius alone */
574           if (tl->omega3 < 1.0) {
575             tao->trust = tl->omega3 * PetscMin(norm_d, tao->trust);
576           } else if (tl->omega3 > 1.0) {
577             tao->trust = PetscMax(tl->omega3 * norm_d, tao->trust);
578           }
579         } else if (step < tl->nu4) {
580           /* Full step taken; increase the radius */
581           tao->trust = PetscMax(tl->omega4 * norm_d, tao->trust);
582         } else {
583           /* More than full step taken; increase the radius */
584           tao->trust = PetscMax(tl->omega5 * norm_d, tao->trust);
585         }
586       } else {
587         /* Newton step was not good; reduce the radius */
588         tao->trust = tl->omega1 * PetscMin(norm_d, tao->trust);
589       }
590     } else {
591       /* Trust-region step is accepted */
592       ierr = VecCopy(tl->W, tao->solution);CHKERRQ(ierr);
593       f = ftrial;
594       ierr = TaoComputeGradient(tao, tao->solution, tao->gradient);CHKERRQ(ierr);
595       ++tl->ntrust;
596     }
597 
598     /* The radius may have been increased; modify if it is too large */
599     tao->trust = PetscMin(tao->trust, tl->max_radius);
600 
601     /* Check for converged */
602     ierr = VecNorm(tao->gradient, NORM_2, &gnorm);CHKERRQ(ierr);
603     if (PetscIsInfOrNanReal(f) || PetscIsInfOrNanReal(gnorm)) SETERRQ(PetscObjectComm((PetscObject)tao),PETSC_ERR_USER,"User provided compute function generated Not-a-Number");
604     needH = 1;
605 
606     ierr = TaoLogConvergenceHistory(tao,f,gnorm,0.0,tao->ksp_its);CHKERRQ(ierr);
607     ierr = TaoMonitor(tao,tao->niter,f,gnorm,0.0,step);CHKERRQ(ierr);
608     ierr = (*tao->ops->convergencetest)(tao,tao->cnvP);CHKERRQ(ierr);
609   }
610   PetscFunctionReturn(0);
611 }
612 
613 /* ---------------------------------------------------------- */
614 static PetscErrorCode TaoSetUp_NTL(Tao tao)
615 {
616   TAO_NTL        *tl = (TAO_NTL *)tao->data;
617   PetscErrorCode ierr;
618 
619   PetscFunctionBegin;
620   if (!tao->gradient) {ierr = VecDuplicate(tao->solution, &tao->gradient);CHKERRQ(ierr); }
621   if (!tao->stepdirection) {ierr = VecDuplicate(tao->solution, &tao->stepdirection);CHKERRQ(ierr);}
622   if (!tl->W) { ierr = VecDuplicate(tao->solution, &tl->W);CHKERRQ(ierr);}
623   if (!tl->Xold) { ierr = VecDuplicate(tao->solution, &tl->Xold);CHKERRQ(ierr);}
624   if (!tl->Gold) { ierr = VecDuplicate(tao->solution, &tl->Gold);CHKERRQ(ierr);}
625   tl->bfgs_pre = NULL;
626   tl->M = NULL;
627   PetscFunctionReturn(0);
628 }
629 
630 /*------------------------------------------------------------*/
631 static PetscErrorCode TaoDestroy_NTL(Tao tao)
632 {
633   TAO_NTL        *tl = (TAO_NTL *)tao->data;
634   PetscErrorCode ierr;
635 
636   PetscFunctionBegin;
637   if (tao->setupcalled) {
638     ierr = VecDestroy(&tl->W);CHKERRQ(ierr);
639     ierr = VecDestroy(&tl->Xold);CHKERRQ(ierr);
640     ierr = VecDestroy(&tl->Gold);CHKERRQ(ierr);
641   }
642   ierr = PetscFree(tao->data);CHKERRQ(ierr);
643   PetscFunctionReturn(0);
644 }
645 
646 /*------------------------------------------------------------*/
647 static PetscErrorCode TaoSetFromOptions_NTL(PetscOptionItems *PetscOptionsObject,Tao tao)
648 {
649   TAO_NTL        *tl = (TAO_NTL *)tao->data;
650   PetscErrorCode ierr;
651 
652   PetscFunctionBegin;
653   ierr = PetscOptionsHead(PetscOptionsObject,"Newton trust region with line search method for unconstrained optimization");CHKERRQ(ierr);
654   ierr = PetscOptionsEList("-tao_ntl_init_type", "radius initialization type", "", NTL_INIT, NTL_INIT_TYPES, NTL_INIT[tl->init_type], &tl->init_type,NULL);CHKERRQ(ierr);
655   ierr = PetscOptionsEList("-tao_ntl_update_type", "radius update type", "", NTL_UPDATE, NTL_UPDATE_TYPES, NTL_UPDATE[tl->update_type], &tl->update_type,NULL);CHKERRQ(ierr);
656   ierr = PetscOptionsReal("-tao_ntl_eta1", "poor steplength; reduce radius", "", tl->eta1, &tl->eta1,NULL);CHKERRQ(ierr);
657   ierr = PetscOptionsReal("-tao_ntl_eta2", "reasonable steplength; leave radius alone", "", tl->eta2, &tl->eta2,NULL);CHKERRQ(ierr);
658   ierr = PetscOptionsReal("-tao_ntl_eta3", "good steplength; increase radius", "", tl->eta3, &tl->eta3,NULL);CHKERRQ(ierr);
659   ierr = PetscOptionsReal("-tao_ntl_eta4", "excellent steplength; greatly increase radius", "", tl->eta4, &tl->eta4,NULL);CHKERRQ(ierr);
660   ierr = PetscOptionsReal("-tao_ntl_alpha1", "", "", tl->alpha1, &tl->alpha1,NULL);CHKERRQ(ierr);
661   ierr = PetscOptionsReal("-tao_ntl_alpha2", "", "", tl->alpha2, &tl->alpha2,NULL);CHKERRQ(ierr);
662   ierr = PetscOptionsReal("-tao_ntl_alpha3", "", "", tl->alpha3, &tl->alpha3,NULL);CHKERRQ(ierr);
663   ierr = PetscOptionsReal("-tao_ntl_alpha4", "", "", tl->alpha4, &tl->alpha4,NULL);CHKERRQ(ierr);
664   ierr = PetscOptionsReal("-tao_ntl_alpha5", "", "", tl->alpha5, &tl->alpha5,NULL);CHKERRQ(ierr);
665   ierr = PetscOptionsReal("-tao_ntl_nu1", "poor steplength; reduce radius", "", tl->nu1, &tl->nu1,NULL);CHKERRQ(ierr);
666   ierr = PetscOptionsReal("-tao_ntl_nu2", "reasonable steplength; leave radius alone", "", tl->nu2, &tl->nu2,NULL);CHKERRQ(ierr);
667   ierr = PetscOptionsReal("-tao_ntl_nu3", "good steplength; increase radius", "", tl->nu3, &tl->nu3,NULL);CHKERRQ(ierr);
668   ierr = PetscOptionsReal("-tao_ntl_nu4", "excellent steplength; greatly increase radius", "", tl->nu4, &tl->nu4,NULL);CHKERRQ(ierr);
669   ierr = PetscOptionsReal("-tao_ntl_omega1", "", "", tl->omega1, &tl->omega1,NULL);CHKERRQ(ierr);
670   ierr = PetscOptionsReal("-tao_ntl_omega2", "", "", tl->omega2, &tl->omega2,NULL);CHKERRQ(ierr);
671   ierr = PetscOptionsReal("-tao_ntl_omega3", "", "", tl->omega3, &tl->omega3,NULL);CHKERRQ(ierr);
672   ierr = PetscOptionsReal("-tao_ntl_omega4", "", "", tl->omega4, &tl->omega4,NULL);CHKERRQ(ierr);
673   ierr = PetscOptionsReal("-tao_ntl_omega5", "", "", tl->omega5, &tl->omega5,NULL);CHKERRQ(ierr);
674   ierr = PetscOptionsReal("-tao_ntl_mu1_i", "", "", tl->mu1_i, &tl->mu1_i,NULL);CHKERRQ(ierr);
675   ierr = PetscOptionsReal("-tao_ntl_mu2_i", "", "", tl->mu2_i, &tl->mu2_i,NULL);CHKERRQ(ierr);
676   ierr = PetscOptionsReal("-tao_ntl_gamma1_i", "", "", tl->gamma1_i, &tl->gamma1_i,NULL);CHKERRQ(ierr);
677   ierr = PetscOptionsReal("-tao_ntl_gamma2_i", "", "", tl->gamma2_i, &tl->gamma2_i,NULL);CHKERRQ(ierr);
678   ierr = PetscOptionsReal("-tao_ntl_gamma3_i", "", "", tl->gamma3_i, &tl->gamma3_i,NULL);CHKERRQ(ierr);
679   ierr = PetscOptionsReal("-tao_ntl_gamma4_i", "", "", tl->gamma4_i, &tl->gamma4_i,NULL);CHKERRQ(ierr);
680   ierr = PetscOptionsReal("-tao_ntl_theta_i", "", "", tl->theta_i, &tl->theta_i,NULL);CHKERRQ(ierr);
681   ierr = PetscOptionsReal("-tao_ntl_mu1", "", "", tl->mu1, &tl->mu1,NULL);CHKERRQ(ierr);
682   ierr = PetscOptionsReal("-tao_ntl_mu2", "", "", tl->mu2, &tl->mu2,NULL);CHKERRQ(ierr);
683   ierr = PetscOptionsReal("-tao_ntl_gamma1", "", "", tl->gamma1, &tl->gamma1,NULL);CHKERRQ(ierr);
684   ierr = PetscOptionsReal("-tao_ntl_gamma2", "", "", tl->gamma2, &tl->gamma2,NULL);CHKERRQ(ierr);
685   ierr = PetscOptionsReal("-tao_ntl_gamma3", "", "", tl->gamma3, &tl->gamma3,NULL);CHKERRQ(ierr);
686   ierr = PetscOptionsReal("-tao_ntl_gamma4", "", "", tl->gamma4, &tl->gamma4,NULL);CHKERRQ(ierr);
687   ierr = PetscOptionsReal("-tao_ntl_theta", "", "", tl->theta, &tl->theta,NULL);CHKERRQ(ierr);
688   ierr = PetscOptionsReal("-tao_ntl_min_radius", "lower bound on initial radius", "", tl->min_radius, &tl->min_radius,NULL);CHKERRQ(ierr);
689   ierr = PetscOptionsReal("-tao_ntl_max_radius", "upper bound on radius", "", tl->max_radius, &tl->max_radius,NULL);CHKERRQ(ierr);
690   ierr = PetscOptionsReal("-tao_ntl_epsilon", "tolerance used when computing actual and predicted reduction", "", tl->epsilon, &tl->epsilon,NULL);CHKERRQ(ierr);
691   ierr = PetscOptionsTail();CHKERRQ(ierr);
692   ierr = TaoLineSearchSetFromOptions(tao->linesearch);CHKERRQ(ierr);
693   ierr = KSPSetFromOptions(tao->ksp);CHKERRQ(ierr);
694   PetscFunctionReturn(0);
695 }
696 
697 /*------------------------------------------------------------*/
698 static PetscErrorCode TaoView_NTL(Tao tao, PetscViewer viewer)
699 {
700   TAO_NTL        *tl = (TAO_NTL *)tao->data;
701   PetscBool      isascii;
702   PetscErrorCode ierr;
703 
704   PetscFunctionBegin;
705   ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&isascii);CHKERRQ(ierr);
706   if (isascii) {
707     ierr = PetscViewerASCIIPushTab(viewer);CHKERRQ(ierr);
708     ierr = PetscViewerASCIIPrintf(viewer, "Trust-region steps: %D\n", tl->ntrust);CHKERRQ(ierr);
709     ierr = PetscViewerASCIIPrintf(viewer, "Newton search steps: %D\n", tl->newt);CHKERRQ(ierr);
710     ierr = PetscViewerASCIIPrintf(viewer, "BFGS search steps: %D\n", tl->bfgs);CHKERRQ(ierr);
711     ierr = PetscViewerASCIIPrintf(viewer, "Gradient search steps: %D\n", tl->grad);CHKERRQ(ierr);
712     ierr = PetscViewerASCIIPopTab(viewer);CHKERRQ(ierr);
713   }
714   PetscFunctionReturn(0);
715 }
716 
717 /* ---------------------------------------------------------- */
718 /*MC
719   TAONTL - Newton's method with trust region globalization and line search fallback.
720   At each iteration, the Newton trust region method solves the system for d
721   and performs a line search in the d direction:
722 
723             min_d  .5 dT Hk d + gkT d,  s.t.   ||d|| < Delta_k
724 
725   Options Database Keys:
726 + -tao_ntl_init_type - "constant","direction","interpolation"
727 . -tao_ntl_update_type - "reduction","interpolation"
728 . -tao_ntl_min_radius - lower bound on trust region radius
729 . -tao_ntl_max_radius - upper bound on trust region radius
730 . -tao_ntl_epsilon - tolerance for accepting actual / predicted reduction
731 . -tao_ntl_mu1_i - mu1 interpolation init factor
732 . -tao_ntl_mu2_i - mu2 interpolation init factor
733 . -tao_ntl_gamma1_i - gamma1 interpolation init factor
734 . -tao_ntl_gamma2_i - gamma2 interpolation init factor
735 . -tao_ntl_gamma3_i - gamma3 interpolation init factor
736 . -tao_ntl_gamma4_i - gamma4 interpolation init factor
737 . -tao_ntl_theta_i - theta1 interpolation init factor
738 . -tao_ntl_eta1 - eta1 reduction update factor
739 . -tao_ntl_eta2 - eta2 reduction update factor
740 . -tao_ntl_eta3 - eta3 reduction update factor
741 . -tao_ntl_eta4 - eta4 reduction update factor
742 . -tao_ntl_alpha1 - alpha1 reduction update factor
743 . -tao_ntl_alpha2 - alpha2 reduction update factor
744 . -tao_ntl_alpha3 - alpha3 reduction update factor
745 . -tao_ntl_alpha4 - alpha4 reduction update factor
746 . -tao_ntl_alpha4 - alpha4 reduction update factor
747 . -tao_ntl_mu1 - mu1 interpolation update
748 . -tao_ntl_mu2 - mu2 interpolation update
749 . -tao_ntl_gamma1 - gamma1 interpolcation update
750 . -tao_ntl_gamma2 - gamma2 interpolcation update
751 . -tao_ntl_gamma3 - gamma3 interpolcation update
752 . -tao_ntl_gamma4 - gamma4 interpolation update
753 - -tao_ntl_theta - theta1 interpolation update
754 
755   Level: beginner
756 M*/
757 PETSC_EXTERN PetscErrorCode TaoCreate_NTL(Tao tao)
758 {
759   TAO_NTL        *tl;
760   PetscErrorCode ierr;
761   const char     *morethuente_type = TAOLINESEARCHMT;
762 
763   PetscFunctionBegin;
764   ierr = PetscNewLog(tao,&tl);CHKERRQ(ierr);
765   tao->ops->setup = TaoSetUp_NTL;
766   tao->ops->solve = TaoSolve_NTL;
767   tao->ops->view = TaoView_NTL;
768   tao->ops->setfromoptions = TaoSetFromOptions_NTL;
769   tao->ops->destroy = TaoDestroy_NTL;
770 
771   /* Override default settings (unless already changed) */
772   if (!tao->max_it_changed) tao->max_it = 50;
773   if (!tao->trust0_changed) tao->trust0 = 100.0;
774 
775   tao->data = (void*)tl;
776 
777   /* Default values for trust-region radius update based on steplength */
778   tl->nu1 = 0.25;
779   tl->nu2 = 0.50;
780   tl->nu3 = 1.00;
781   tl->nu4 = 1.25;
782 
783   tl->omega1 = 0.25;
784   tl->omega2 = 0.50;
785   tl->omega3 = 1.00;
786   tl->omega4 = 2.00;
787   tl->omega5 = 4.00;
788 
789   /* Default values for trust-region radius update based on reduction */
790   tl->eta1 = 1.0e-4;
791   tl->eta2 = 0.25;
792   tl->eta3 = 0.50;
793   tl->eta4 = 0.90;
794 
795   tl->alpha1 = 0.25;
796   tl->alpha2 = 0.50;
797   tl->alpha3 = 1.00;
798   tl->alpha4 = 2.00;
799   tl->alpha5 = 4.00;
800 
801   /* Default values for trust-region radius update based on interpolation */
802   tl->mu1 = 0.10;
803   tl->mu2 = 0.50;
804 
805   tl->gamma1 = 0.25;
806   tl->gamma2 = 0.50;
807   tl->gamma3 = 2.00;
808   tl->gamma4 = 4.00;
809 
810   tl->theta = 0.05;
811 
812   /* Default values for trust region initialization based on interpolation */
813   tl->mu1_i = 0.35;
814   tl->mu2_i = 0.50;
815 
816   tl->gamma1_i = 0.0625;
817   tl->gamma2_i = 0.5;
818   tl->gamma3_i = 2.0;
819   tl->gamma4_i = 5.0;
820 
821   tl->theta_i = 0.25;
822 
823   /* Remaining parameters */
824   tl->min_radius = 1.0e-10;
825   tl->max_radius = 1.0e10;
826   tl->epsilon = 1.0e-6;
827 
828   tl->init_type       = NTL_INIT_INTERPOLATION;
829   tl->update_type     = NTL_UPDATE_REDUCTION;
830 
831   ierr = TaoLineSearchCreate(((PetscObject)tao)->comm, &tao->linesearch);CHKERRQ(ierr);
832   ierr = PetscObjectIncrementTabLevel((PetscObject)tao->linesearch,(PetscObject)tao,1);CHKERRQ(ierr);
833   ierr = TaoLineSearchSetType(tao->linesearch, morethuente_type);CHKERRQ(ierr);
834   ierr = TaoLineSearchUseTaoRoutines(tao->linesearch, tao);CHKERRQ(ierr);
835   ierr = TaoLineSearchSetOptionsPrefix(tao->linesearch,tao->hdr.prefix);CHKERRQ(ierr);
836   ierr = KSPCreate(((PetscObject)tao)->comm,&tao->ksp);CHKERRQ(ierr);
837   ierr = PetscObjectIncrementTabLevel((PetscObject)tao->ksp,(PetscObject)tao,1);CHKERRQ(ierr);
838   ierr = KSPSetOptionsPrefix(tao->ksp,tao->hdr.prefix);CHKERRQ(ierr);
839   ierr = KSPAppendOptionsPrefix(tao->ksp,"tao_ntl_");CHKERRQ(ierr);
840   ierr = KSPSetType(tao->ksp,KSPSTCG);CHKERRQ(ierr);
841   PetscFunctionReturn(0);
842 }
843