xref: /petsc/src/tao/bound/impls/bncg/bncg.c (revision b4e8f828033637a7b4f7fd14e8564d6efe661f39) 
1 #include <petsctaolinesearch.h>
2 #include <../src/tao/bound/impls/bncg/bncg.h>
3 
4 #define CG_FletcherReeves       0
5 #define CG_PolakRibiere         1
6 #define CG_PolakRibierePlus     2
7 #define CG_HestenesStiefel      3
8 #define CG_DaiYuan              4
9 #define CG_GradientDescent      5
10 #define CG_Types                6
11 
12 static const char *CG_Table[64] = {"fr", "pr", "prp", "hs", "dy", "gd"};
13 
14 #define CG_AS_NONE       0
15 #define CG_AS_BERTSEKAS  1
16 #define CG_AS_SIZE       2
17 
18 static const char *CG_AS_TYPE[64] = {"none", "bertsekas"};
19 
20 PetscErrorCode TaoBNCGSetRecycleFlag(Tao tao, PetscBool recycle)
21 {
22   TAO_BNCG                     *cg = (TAO_BNCG*)tao->data;
23 
24   PetscFunctionBegin;
25   cg->recycle = recycle;
26   PetscFunctionReturn(0);
27 }
28 
29 PetscErrorCode TaoBNCGEstimateActiveSet(Tao tao, PetscInt asType)
30 {
31   PetscErrorCode               ierr;
32   TAO_BNCG                     *cg = (TAO_BNCG *)tao->data;
33 
34   PetscFunctionBegin;
35   ierr = ISDestroy(&cg->inactive_old);CHKERRQ(ierr);
36   if (cg->inactive_idx) {
37     ierr = ISDuplicate(cg->inactive_idx, &cg->inactive_old);CHKERRQ(ierr);
38     ierr = ISCopy(cg->inactive_idx, cg->inactive_old);CHKERRQ(ierr);
39   }
40   switch (asType) {
41   case CG_AS_NONE:
42     ierr = ISDestroy(&cg->inactive_idx);CHKERRQ(ierr);
43     ierr = VecWhichInactive(tao->XL, tao->solution, cg->unprojected_gradient, tao->XU, PETSC_TRUE, &cg->inactive_idx);CHKERRQ(ierr);
44     ierr = ISDestroy(&cg->active_idx);CHKERRQ(ierr);
45     ierr = ISComplementVec(cg->inactive_idx, tao->solution, &cg->active_idx);CHKERRQ(ierr);
46     break;
47 
48   case CG_AS_BERTSEKAS:
49     /* Use gradient descent to estimate the active set */
50     ierr = VecCopy(cg->unprojected_gradient, cg->W);CHKERRQ(ierr);
51     ierr = VecScale(cg->W, -1.0);CHKERRQ(ierr);
52     ierr = TaoEstimateActiveBounds(tao->solution, tao->XL, tao->XU, cg->unprojected_gradient, cg->W, cg->work, cg->as_step, &cg->as_tol,
53                                    &cg->active_lower, &cg->active_upper, &cg->active_fixed, &cg->active_idx, &cg->inactive_idx);CHKERRQ(ierr);
54     break;
55 
56   default:
57     break;
58   }
59   PetscFunctionReturn(0);
60 }
61 
62 PetscErrorCode TaoBNCGBoundStep(Tao tao, PetscInt asType, Vec step)
63 {
64   PetscErrorCode               ierr;
65   TAO_BNCG                     *cg = (TAO_BNCG *)tao->data;
66 
67   PetscFunctionBegin;
68   switch (asType) {
69   case CG_AS_NONE:
70     ierr = VecISSet(step, cg->active_idx, 0.0);CHKERRQ(ierr);
71     break;
72 
73   case CG_AS_BERTSEKAS:
74     ierr = TaoBoundStep(tao->solution, tao->XL, tao->XU, cg->active_lower, cg->active_upper, cg->active_fixed, 1.0, step);CHKERRQ(ierr);
75     break;
76 
77   default:
78     break;
79   }
80   PetscFunctionReturn(0);
81 }
82 
83 static PetscErrorCode TaoSolve_BNCG(Tao tao)
84 {
85   TAO_BNCG                     *cg = (TAO_BNCG*)tao->data;
86   PetscErrorCode               ierr;
87   TaoLineSearchConvergedReason ls_status = TAOLINESEARCH_CONTINUE_ITERATING;
88   PetscReal                    step=1.0,gnorm,gnorm2,gd,ginner,beta,dnorm;
89   PetscReal                    gd_old,gnorm2_old,f_old,resnorm;
90   PetscBool                    cg_restart, gd_fallback = PETSC_FALSE;
91   PetscInt                     nDiff;
92 
93   PetscFunctionBegin;
94   /*   Project the current point onto the feasible set */
95   ierr = TaoComputeVariableBounds(tao);CHKERRQ(ierr);
96   ierr = TaoLineSearchSetVariableBounds(tao->linesearch,tao->XL,tao->XU);CHKERRQ(ierr);
97 
98   /* Project the initial point onto the feasible region */
99   ierr = TaoBoundSolution(tao->solution, tao->XL,tao->XU, 0.0, &nDiff, tao->solution);CHKERRQ(ierr);
100 
101   if (nDiff > 0 || !cg->recycle) {
102     /*  Solver is not being recycled so just compute the objective function and criteria */
103     ierr = TaoComputeObjectiveAndGradient(tao, tao->solution, &cg->f, cg->unprojected_gradient);CHKERRQ(ierr);
104   } else {
105     /* We are recycling, so we have to compute ||g_old||^2 for use in the CG step calculation */
106     ierr = VecDot(cg->G_old, cg->G_old, &gnorm2_old);CHKERRQ(ierr);
107   }
108   ierr = VecNorm(cg->unprojected_gradient,NORM_2,&gnorm);CHKERRQ(ierr);
109   if (PetscIsInfOrNanReal(cg->f) || PetscIsInfOrNanReal(gnorm)) SETERRQ(PETSC_COMM_SELF,1, "User provided compute function generated Inf or NaN");
110 
111   /* Estimate the active set and compute the projected gradient */
112   ierr = TaoBNCGEstimateActiveSet(tao, cg->as_type);CHKERRQ(ierr);
113 
114   /* Project the gradient and calculate the norm */
115   ierr = VecCopy(cg->unprojected_gradient, tao->gradient);CHKERRQ(ierr);
116   ierr = VecISSet(tao->gradient, cg->active_idx, 0.0);CHKERRQ(ierr);
117   ierr = VecNorm(tao->gradient,NORM_2,&gnorm);CHKERRQ(ierr);
118   gnorm2 = gnorm*gnorm;
119 
120   /* Convergence check */
121   tao->niter = 0;
122   tao->reason = TAO_CONTINUE_ITERATING;
123   ierr = TaoLogConvergenceHistory(tao, cg->f, gnorm, 0.0, tao->ksp_its);CHKERRQ(ierr);
124   ierr = TaoMonitor(tao, tao->niter, cg->f, gnorm, 0.0, step);CHKERRQ(ierr);
125   ierr = (*tao->ops->convergencetest)(tao,tao->cnvP);CHKERRQ(ierr);
126   if (tao->reason != TAO_CONTINUE_ITERATING) PetscFunctionReturn(0);
127 
128   /* Start optimization iterations */
129   cg->ls_fails = cg->broken_ortho = cg->descent_error = 0;
130   cg->resets = -1;
131   while (tao->reason == TAO_CONTINUE_ITERATING) {
132     ++tao->niter;
133 
134     /* Check restart conditions for using steepest descent */
135     cg_restart = PETSC_FALSE;
136     ierr = VecDot(tao->gradient, cg->G_old, &ginner);CHKERRQ(ierr);
137     ierr = VecNorm(tao->stepdirection, NORM_2, &dnorm);CHKERRQ(ierr);
138     if (tao->niter == 1 && !cg->recycle && dnorm != 0.0) {
139       /* 1) First iteration, with recycle disabled, and a non-zero previous step */
140       cg_restart = PETSC_TRUE;
141     } else if (PetscAbsScalar(ginner) >= cg->eta * gnorm2) {
142       /* 2) Gradients are far from orthogonal */
143       cg_restart = PETSC_TRUE;
144       ++cg->broken_ortho;
145     }
146 
147     /* Compute CG step */
148     if (cg_restart) {
149       beta = 0.0;
150       ++cg->resets;
151     } else {
152       switch (cg->cg_type) {
153       case CG_FletcherReeves:
154         beta = gnorm2 / gnorm2_old;
155         break;
156 
157       case CG_PolakRibiere:
158         beta = (gnorm2 - ginner) / gnorm2_old;
159         break;
160 
161       case CG_PolakRibierePlus:
162         beta = PetscMax((gnorm2-ginner)/gnorm2_old, 0.0);
163         break;
164 
165       case CG_HestenesStiefel:
166         ierr = VecDot(tao->gradient, tao->stepdirection, &gd);CHKERRQ(ierr);
167         ierr = VecDot(cg->G_old, tao->stepdirection, &gd_old);CHKERRQ(ierr);
168         beta = (gnorm2 - ginner) / (gd - gd_old);
169         break;
170 
171       case CG_DaiYuan:
172         ierr = VecDot(tao->gradient, tao->stepdirection, &gd);CHKERRQ(ierr);
173         ierr = VecDot(cg->G_old, tao->stepdirection, &gd_old);CHKERRQ(ierr);
174         beta = gnorm2 / (gd - gd_old);
175         break;
176 
177       case CG_GradientDescent:
178         beta = 0.0;
179         break;
180 
181       default:
182         beta = 0.0;
183         break;
184       }
185     }
186 
187     /*  Compute the direction d=-g + beta*d */
188     ierr = VecAXPBY(tao->stepdirection, -1.0, beta, tao->gradient);CHKERRQ(ierr);
189     ierr = TaoBNCGBoundStep(tao, cg->as_type, tao->stepdirection);CHKERRQ(ierr);
190 
191     if (cg->cg_type != CG_GradientDescent) {
192       /* Figure out which previously active variables became inactive this iteration */
193       ierr = ISDestroy(&cg->new_inactives);CHKERRQ(ierr);
194       if (cg->inactive_idx && cg->inactive_old) {
195         ierr = ISDifference(cg->inactive_idx, cg->inactive_old, &cg->new_inactives);CHKERRQ(ierr);
196       }
197 
198       /* Selectively reset the CG step those freshly inactive variables */
199       if (cg->new_inactives) {
200         ierr = VecGetSubVector(tao->stepdirection, cg->new_inactives, &cg->inactive_step);CHKERRQ(ierr);
201         ierr = VecGetSubVector(cg->unprojected_gradient, cg->new_inactives, &cg->inactive_grad);CHKERRQ(ierr);
202         ierr = VecCopy(cg->inactive_grad, cg->inactive_step);CHKERRQ(ierr);
203         ierr = VecScale(cg->inactive_step, -1.0);CHKERRQ(ierr);
204         ierr = VecRestoreSubVector(tao->stepdirection, cg->new_inactives, &cg->inactive_step);CHKERRQ(ierr);
205         ierr = VecRestoreSubVector(cg->unprojected_gradient, cg->new_inactives, &cg->inactive_grad);CHKERRQ(ierr);
206       }
207 
208       /* Verify that this is a descent direction */
209       ierr = VecDot(tao->gradient, tao->stepdirection, &gd);CHKERRQ(ierr);
210       ierr = VecNorm(tao->stepdirection, NORM_2, &dnorm);
211       if (gd > -cg->rho*PetscPowReal(dnorm, cg->pow)) {
212         /* Not a descent direction, so we reset back to projected gradient descent */
213         ierr = VecAXPBY(tao->stepdirection, -1.0, 0.0, tao->gradient);CHKERRQ(ierr);
214         ierr = TaoBNCGBoundStep(tao, cg->as_type, tao->stepdirection);CHKERRQ(ierr);
215         ++cg->resets;
216         ++cg->descent_error;
217         gd_fallback = PETSC_TRUE;
218       } else {
219         gd_fallback = PETSC_FALSE;
220       }
221     }
222 
223     /* Store solution and gradient info before it changes */
224     ierr = VecCopy(tao->solution, cg->X_old);CHKERRQ(ierr);
225     ierr = VecCopy(tao->gradient, cg->G_old);CHKERRQ(ierr);
226     ierr = VecCopy(cg->unprojected_gradient, cg->unprojected_gradient_old);CHKERRQ(ierr);
227     gnorm2_old = gnorm2;
228     f_old = cg->f;
229 
230     /* Perform bounded line search */
231     ierr = TaoLineSearchApply(tao->linesearch, tao->solution, &cg->f, cg->unprojected_gradient, tao->stepdirection, &step, &ls_status);CHKERRQ(ierr);
232     ierr = TaoAddLineSearchCounts(tao);CHKERRQ(ierr);
233 
234     /*  Check linesearch failure */
235     if (ls_status != TAOLINESEARCH_SUCCESS && ls_status != TAOLINESEARCH_SUCCESS_USER) {
236       ++cg->ls_fails;
237       /* Restore previous point */
238       gnorm2 = gnorm2_old;
239       cg->f = f_old;
240       ierr = VecCopy(cg->X_old, tao->solution);CHKERRQ(ierr);
241       ierr = VecCopy(cg->G_old, tao->gradient);CHKERRQ(ierr);
242       ierr = VecCopy(cg->unprojected_gradient_old, cg->unprojected_gradient);CHKERRQ(ierr);
243 
244       if (cg->cg_type == CG_GradientDescent || gd_fallback){
245         /* Nothing left to do but fail out of the optimization */
246         step = 0.0;
247         tao->reason = TAO_DIVERGED_LS_FAILURE;
248       } else {
249         /* Fall back on the unscaled gradient step */
250         ierr = VecCopy(tao->gradient, tao->stepdirection);CHKERRQ(ierr);
251         ierr = VecScale(tao->stepdirection, -1.0);CHKERRQ(ierr);
252         ierr = TaoBNCGBoundStep(tao, cg->as_type, tao->stepdirection);CHKERRQ(ierr);
253 
254         ierr = TaoLineSearchSetInitialStepLength(tao->linesearch, 1.0);CHKERRQ(ierr);
255         ierr = TaoLineSearchApply(tao->linesearch, tao->solution, &cg->f, cg->unprojected_gradient, tao->stepdirection, &step, &ls_status);CHKERRQ(ierr);
256         ierr = TaoAddLineSearchCounts(tao);CHKERRQ(ierr);
257 
258         if (ls_status != TAOLINESEARCH_SUCCESS && ls_status != TAOLINESEARCH_SUCCESS_USER){
259           cg->ls_fails++;
260           /* Restore previous point */
261           gnorm2 = gnorm2_old;
262           cg->f = f_old;
263           ierr = VecCopy(cg->X_old, tao->solution);CHKERRQ(ierr);
264           ierr = VecCopy(cg->G_old, tao->gradient);CHKERRQ(ierr);
265           ierr = VecCopy(cg->unprojected_gradient_old, cg->unprojected_gradient);CHKERRQ(ierr);
266 
267           /* Nothing left to do but fail out of the optimization */
268           step = 0.0;
269           tao->reason = TAO_DIVERGED_LS_FAILURE;
270         }
271       }
272     }
273 
274     if (tao->reason != TAO_DIVERGED_LS_FAILURE) {
275       /* Estimate the active set at the new solution */
276       ierr = TaoBNCGEstimateActiveSet(tao, cg->as_type);CHKERRQ(ierr);
277 
278       /* Compute the projected gradient and its norm */
279       ierr = VecCopy(cg->unprojected_gradient, tao->gradient);CHKERRQ(ierr);
280       ierr = VecISSet(tao->gradient, cg->active_idx, 0.0);CHKERRQ(ierr);
281       ierr = VecNorm(tao->gradient,NORM_2,&gnorm);CHKERRQ(ierr);
282       gnorm2 = gnorm*gnorm;
283     }
284 
285     /* Convergence test */
286     ierr = VecFischer(tao->solution, cg->unprojected_gradient, tao->XL, tao->XU, cg->W);CHKERRQ(ierr);
287     ierr = VecNorm(cg->W, NORM_2, &resnorm);CHKERRQ(ierr);
288     if (PetscIsInfOrNanReal(resnorm)) SETERRQ(PETSC_COMM_SELF,1, "User provided compute function generated Inf or NaN");
289     ierr = TaoLogConvergenceHistory(tao, cg->f, resnorm, 0.0, tao->ksp_its);CHKERRQ(ierr);
290     ierr = TaoMonitor(tao, tao->niter, cg->f, resnorm, 0.0, step);CHKERRQ(ierr);
291     ierr = (*tao->ops->convergencetest)(tao,tao->cnvP);CHKERRQ(ierr);
292   }
293   PetscFunctionReturn(0);
294 }
295 
296 static PetscErrorCode TaoSetUp_BNCG(Tao tao)
297 {
298   TAO_BNCG         *cg = (TAO_BNCG*)tao->data;
299   PetscErrorCode ierr;
300 
301   PetscFunctionBegin;
302   if (!tao->gradient) {
303     ierr = VecDuplicate(tao->solution,&tao->gradient);CHKERRQ(ierr);
304   }
305   if (!tao->stepdirection) {
306     ierr = VecDuplicate(tao->solution,&tao->stepdirection);CHKERRQ(ierr);
307   }
308   if (!cg->W) {
309     ierr = VecDuplicate(tao->solution,&cg->W);CHKERRQ(ierr);
310   }
311   if (!cg->work) {
312     ierr = VecDuplicate(tao->solution,&cg->work);CHKERRQ(ierr);
313   }
314   if (!cg->X_old) {
315     ierr = VecDuplicate(tao->solution,&cg->X_old);CHKERRQ(ierr);
316   }
317   if (!cg->G_old) {
318     ierr = VecDuplicate(tao->gradient,&cg->G_old);CHKERRQ(ierr);
319   }
320   if (!cg->unprojected_gradient) {
321     ierr = VecDuplicate(tao->gradient,&cg->unprojected_gradient);CHKERRQ(ierr);
322   }
323   if (!cg->unprojected_gradient_old) {
324     ierr = VecDuplicate(tao->gradient,&cg->unprojected_gradient_old);CHKERRQ(ierr);
325   }
326   PetscFunctionReturn(0);
327 }
328 
329 static PetscErrorCode TaoDestroy_BNCG(Tao tao)
330 {
331   TAO_BNCG       *cg = (TAO_BNCG*) tao->data;
332   PetscErrorCode ierr;
333 
334   PetscFunctionBegin;
335   if (tao->setupcalled) {
336     ierr = VecDestroy(&cg->W);CHKERRQ(ierr);
337     ierr = VecDestroy(&cg->work);CHKERRQ(ierr);
338     ierr = VecDestroy(&cg->X_old);CHKERRQ(ierr);
339     ierr = VecDestroy(&cg->G_old);CHKERRQ(ierr);
340     ierr = VecDestroy(&cg->unprojected_gradient);CHKERRQ(ierr);
341     ierr = VecDestroy(&cg->unprojected_gradient_old);CHKERRQ(ierr);
342   }
343   ierr = ISDestroy(&cg->active_lower);CHKERRQ(ierr);
344   ierr = ISDestroy(&cg->active_upper);CHKERRQ(ierr);
345   ierr = ISDestroy(&cg->active_fixed);CHKERRQ(ierr);
346   ierr = ISDestroy(&cg->active_idx);CHKERRQ(ierr);
347   ierr = ISDestroy(&cg->inactive_idx);CHKERRQ(ierr);
348   ierr = ISDestroy(&cg->inactive_old);CHKERRQ(ierr);
349   ierr = ISDestroy(&cg->new_inactives);CHKERRQ(ierr);
350   ierr = PetscFree(tao->data);CHKERRQ(ierr);
351   PetscFunctionReturn(0);
352 }
353 
354 static PetscErrorCode TaoSetFromOptions_BNCG(PetscOptionItems *PetscOptionsObject,Tao tao)
355  {
356     TAO_BNCG       *cg = (TAO_BNCG*)tao->data;
357     PetscErrorCode ierr;
358 
359     PetscFunctionBegin;
360     ierr = TaoLineSearchSetFromOptions(tao->linesearch);CHKERRQ(ierr);
361     ierr = PetscOptionsHead(PetscOptionsObject,"Nonlinear Conjugate Gradient method for unconstrained optimization");CHKERRQ(ierr);
362     ierr = PetscOptionsReal("-tao_bncg_eta","restart tolerance", "", cg->eta,&cg->eta,NULL);CHKERRQ(ierr);
363     ierr = PetscOptionsReal("-tao_bncg_rho","descent direction tolerance", "", cg->rho,&cg->rho,NULL);CHKERRQ(ierr);
364     ierr = PetscOptionsReal("-tao_bncg_pow","descent direction exponent", "", cg->pow,&cg->pow,NULL);CHKERRQ(ierr);
365     ierr = PetscOptionsEList("-tao_bncg_type","cg formula", "", CG_Table, CG_Types, CG_Table[cg->cg_type], &cg->cg_type,NULL);CHKERRQ(ierr);
366     ierr = PetscOptionsEList("-tao_bncg_as_type","active set estimation method", "", CG_AS_TYPE, CG_AS_SIZE, CG_AS_TYPE[cg->cg_type], &cg->cg_type,NULL);CHKERRQ(ierr);
367     ierr = PetscOptionsBool("-tao_bncg_recycle","enable recycling the existing solution and gradient at the start of a new solve","",cg->recycle,&cg->recycle,NULL);CHKERRQ(ierr);
368     ierr = PetscOptionsReal("-tao_bncg_as_tol", "initial tolerance used when estimating actively bounded variables","",cg->as_tol,&cg->as_tol,NULL);CHKERRQ(ierr);
369     ierr = PetscOptionsReal("-tao_bncg_as_step", "step length used when estimating actively bounded variables","",cg->as_step,&cg->as_step,NULL);CHKERRQ(ierr);
370    ierr = PetscOptionsTail();CHKERRQ(ierr);
371    PetscFunctionReturn(0);
372 }
373 
374 static PetscErrorCode TaoView_BNCG(Tao tao, PetscViewer viewer)
375 {
376   PetscBool      isascii;
377   TAO_BNCG       *cg = (TAO_BNCG*)tao->data;
378   PetscErrorCode ierr;
379 
380   PetscFunctionBegin;
381   ierr = PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii);CHKERRQ(ierr);
382   if (isascii) {
383     ierr = PetscViewerASCIIPushTab(viewer);CHKERRQ(ierr);
384     ierr = PetscViewerASCIIPrintf(viewer, "CG Type: %s\n", CG_Table[cg->cg_type]);CHKERRQ(ierr);
385     ierr = PetscViewerASCIIPrintf(viewer, "Resets: %i\n", cg->resets);CHKERRQ(ierr);
386     ierr = PetscViewerASCIIPrintf(viewer, "  Broken ortho: %i\n", cg->broken_ortho);CHKERRQ(ierr);
387     ierr = PetscViewerASCIIPrintf(viewer, "  Not a descent dir.: %i\n", cg->descent_error);CHKERRQ(ierr);
388     ierr = PetscViewerASCIIPrintf(viewer, "Line search fails: %i\n", cg->ls_fails);CHKERRQ(ierr);
389     ierr = PetscViewerASCIIPopTab(viewer);CHKERRQ(ierr);
390   }
391   PetscFunctionReturn(0);
392 }
393 
394 /*MC
395      TAOBNCG -   Bound-constrained Nonlinear Conjugate Gradient method.
396 
397    Options Database Keys:
398 +      -tao_bncg_recycle - enable recycling the latest calculated gradient vector in subsequent TaoSolve() calls
399 .      -tao_bncg_eta <r> - restart tolerance
400 .      -tao_bncg_type <taocg_type> - cg formula
401 .      -tao_bncg_as_type <none,bertsekas> - active set estimation method
402 .      -tao_bncg_as_tol <r> - tolerance used in Bertsekas active-set estimation
403 .      -tao_bncg_as_step <r> - trial step length used in Bertsekas active-set estimation
404 
405   Notes:
406      CG formulas are:
407          "fr" - Fletcher-Reeves
408          "pr" - Polak-Ribiere
409          "prp" - Polak-Ribiere-Plus
410          "hs" - Hestenes-Steifel
411          "dy" - Dai-Yuan
412          "gd" - Gradient Descent
413   Level: beginner
414 M*/
415 
416 
417 PETSC_EXTERN PetscErrorCode TaoCreate_BNCG(Tao tao)
418 {
419   TAO_BNCG       *cg;
420   const char     *morethuente_type = TAOLINESEARCHMT;
421   PetscErrorCode ierr;
422 
423   PetscFunctionBegin;
424   tao->ops->setup = TaoSetUp_BNCG;
425   tao->ops->solve = TaoSolve_BNCG;
426   tao->ops->view = TaoView_BNCG;
427   tao->ops->setfromoptions = TaoSetFromOptions_BNCG;
428   tao->ops->destroy = TaoDestroy_BNCG;
429 
430   /* Override default settings (unless already changed) */
431   if (!tao->max_it_changed) tao->max_it = 2000;
432   if (!tao->max_funcs_changed) tao->max_funcs = 4000;
433 
434   /*  Note: nondefault values should be used for nonlinear conjugate gradient  */
435   /*  method.  In particular, gtol should be less that 0.5; the value used in  */
436   /*  Nocedal and Wright is 0.10.  We use the default values for the  */
437   /*  linesearch because it seems to work better. */
438   ierr = TaoLineSearchCreate(((PetscObject)tao)->comm, &tao->linesearch);CHKERRQ(ierr);
439   ierr = PetscObjectIncrementTabLevel((PetscObject)tao->linesearch, (PetscObject)tao, 1);CHKERRQ(ierr);
440   ierr = TaoLineSearchSetType(tao->linesearch, morethuente_type);CHKERRQ(ierr);
441   ierr = TaoLineSearchUseTaoRoutines(tao->linesearch, tao);CHKERRQ(ierr);
442   ierr = TaoLineSearchSetOptionsPrefix(tao->linesearch,tao->hdr.prefix);CHKERRQ(ierr);
443 
444   ierr = PetscNewLog(tao,&cg);CHKERRQ(ierr);
445   tao->data = (void*)cg;
446   cg->rho = 1e-4;
447   cg->pow = 2.1;
448   cg->eta = 0.5;
449   cg->as_step = 0.001;
450   cg->as_tol = 0.001;
451   cg->as_type = CG_AS_BERTSEKAS;
452   cg->cg_type = CG_DaiYuan;
453   cg->recycle = PETSC_FALSE;
454   PetscFunctionReturn(0);
455 }
456