xref: /petsc/src/tao/unconstrained/impls/cg/taocg.c (revision ebead697dbf761eb322f829370bbe90b3bd93fa3) 
1 #include <petsctaolinesearch.h>
2 #include <../src/tao/unconstrained/impls/cg/taocg.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_Types                5
10 
11 static const char *CG_Table[64] = {"fr", "pr", "prp", "hs", "dy"};
12 
13 static PetscErrorCode TaoSolve_CG(Tao tao)
14 {
15   TAO_CG                       *cgP = (TAO_CG*)tao->data;
16   TaoLineSearchConvergedReason ls_status = TAOLINESEARCH_CONTINUE_ITERATING;
17   PetscReal                    step=1.0,f,gnorm,gnorm2,delta,gd,ginner,beta;
18   PetscReal                    gd_old,gnorm2_old,f_old;
19 
20   PetscFunctionBegin;
21   if (tao->XL || tao->XU || tao->ops->computebounds) {
22     PetscCall(PetscInfo(tao,"WARNING: Variable bounds have been set but will be ignored by cg algorithm\n"));
23   }
24 
25   /*  Check convergence criteria */
26   PetscCall(TaoComputeObjectiveAndGradient(tao, tao->solution, &f, tao->gradient));
27   PetscCall(VecNorm(tao->gradient,NORM_2,&gnorm));
28   PetscCheck(!PetscIsInfOrNanReal(f) && !PetscIsInfOrNanReal(gnorm),PetscObjectComm((PetscObject)tao),PETSC_ERR_USER, "User provided compute function generated Inf or NaN");
29 
30   tao->reason = TAO_CONTINUE_ITERATING;
31   PetscCall(TaoLogConvergenceHistory(tao,f,gnorm,0.0,tao->ksp_its));
32   PetscCall(TaoMonitor(tao,tao->niter,f,gnorm,0.0,step));
33   PetscCall((*tao->ops->convergencetest)(tao,tao->cnvP));
34   if (tao->reason != TAO_CONTINUE_ITERATING) PetscFunctionReturn(0);
35 
36   /*  Set initial direction to -gradient */
37   PetscCall(VecCopy(tao->gradient, tao->stepdirection));
38   PetscCall(VecScale(tao->stepdirection, -1.0));
39   gnorm2 = gnorm*gnorm;
40 
41   /*  Set initial scaling for the function */
42   if (f != 0.0) {
43     delta = 2.0*PetscAbsScalar(f) / gnorm2;
44     delta = PetscMax(delta,cgP->delta_min);
45     delta = PetscMin(delta,cgP->delta_max);
46   } else {
47     delta = 2.0 / gnorm2;
48     delta = PetscMax(delta,cgP->delta_min);
49     delta = PetscMin(delta,cgP->delta_max);
50   }
51   /*  Set counter for gradient and reset steps */
52   cgP->ngradsteps = 0;
53   cgP->nresetsteps = 0;
54 
55   while (1) {
56     /* Call general purpose update function */
57     if (tao->ops->update) PetscCall((*tao->ops->update)(tao, tao->niter, tao->user_update));
58 
59     /*  Save the current gradient information */
60     f_old = f;
61     gnorm2_old = gnorm2;
62     PetscCall(VecCopy(tao->solution, cgP->X_old));
63     PetscCall(VecCopy(tao->gradient, cgP->G_old));
64     PetscCall(VecDot(tao->gradient, tao->stepdirection, &gd));
65     if ((gd >= 0) || PetscIsInfOrNanReal(gd)) {
66       ++cgP->ngradsteps;
67       if (f != 0.0) {
68         delta = 2.0*PetscAbsScalar(f) / gnorm2;
69         delta = PetscMax(delta,cgP->delta_min);
70         delta = PetscMin(delta,cgP->delta_max);
71       } else {
72         delta = 2.0 / gnorm2;
73         delta = PetscMax(delta,cgP->delta_min);
74         delta = PetscMin(delta,cgP->delta_max);
75       }
76 
77       PetscCall(VecCopy(tao->gradient, tao->stepdirection));
78       PetscCall(VecScale(tao->stepdirection, -1.0));
79     }
80 
81     /*  Search direction for improving point */
82     PetscCall(TaoLineSearchSetInitialStepLength(tao->linesearch,delta));
83     PetscCall(TaoLineSearchApply(tao->linesearch, tao->solution, &f, tao->gradient, tao->stepdirection, &step, &ls_status));
84     PetscCall(TaoAddLineSearchCounts(tao));
85     if (ls_status != TAOLINESEARCH_SUCCESS && ls_status != TAOLINESEARCH_SUCCESS_USER) {
86       /*  Linesearch failed */
87       /*  Reset factors and use scaled gradient step */
88       ++cgP->nresetsteps;
89       f = f_old;
90       gnorm2 = gnorm2_old;
91       PetscCall(VecCopy(cgP->X_old, tao->solution));
92       PetscCall(VecCopy(cgP->G_old, tao->gradient));
93 
94       if (f != 0.0) {
95         delta = 2.0*PetscAbsScalar(f) / gnorm2;
96         delta = PetscMax(delta,cgP->delta_min);
97         delta = PetscMin(delta,cgP->delta_max);
98       } else {
99         delta = 2.0 / gnorm2;
100         delta = PetscMax(delta,cgP->delta_min);
101         delta = PetscMin(delta,cgP->delta_max);
102       }
103 
104       PetscCall(VecCopy(tao->gradient, tao->stepdirection));
105       PetscCall(VecScale(tao->stepdirection, -1.0));
106 
107       PetscCall(TaoLineSearchSetInitialStepLength(tao->linesearch,delta));
108       PetscCall(TaoLineSearchApply(tao->linesearch, tao->solution, &f, tao->gradient, tao->stepdirection, &step, &ls_status));
109       PetscCall(TaoAddLineSearchCounts(tao));
110 
111       if (ls_status != TAOLINESEARCH_SUCCESS && ls_status != TAOLINESEARCH_SUCCESS_USER) {
112         /*  Linesearch failed again */
113         /*  switch to unscaled gradient */
114         f = f_old;
115         PetscCall(VecCopy(cgP->X_old, tao->solution));
116         PetscCall(VecCopy(cgP->G_old, tao->gradient));
117         delta = 1.0;
118         PetscCall(VecCopy(tao->solution, tao->stepdirection));
119         PetscCall(VecScale(tao->stepdirection, -1.0));
120 
121         PetscCall(TaoLineSearchSetInitialStepLength(tao->linesearch,delta));
122         PetscCall(TaoLineSearchApply(tao->linesearch, tao->solution, &f, tao->gradient, tao->stepdirection, &step, &ls_status));
123         PetscCall(TaoAddLineSearchCounts(tao));
124         if (ls_status != TAOLINESEARCH_SUCCESS && ls_status != TAOLINESEARCH_SUCCESS_USER) {
125 
126           /*  Line search failed for last time -- give up */
127           f = f_old;
128           PetscCall(VecCopy(cgP->X_old, tao->solution));
129           PetscCall(VecCopy(cgP->G_old, tao->gradient));
130           step = 0.0;
131           tao->reason = TAO_DIVERGED_LS_FAILURE;
132         }
133       }
134     }
135 
136     /*  Check for bad value */
137     PetscCall(VecNorm(tao->gradient,NORM_2,&gnorm));
138     PetscCheck(!PetscIsInfOrNanReal(f) && !PetscIsInfOrNanReal(gnorm),PetscObjectComm((PetscObject)tao),PETSC_ERR_USER,"User-provided compute function generated Inf or NaN");
139 
140     /*  Check for termination */
141     gnorm2 =gnorm * gnorm;
142     tao->niter++;
143     PetscCall(TaoLogConvergenceHistory(tao,f,gnorm,0.0,tao->ksp_its));
144     PetscCall(TaoMonitor(tao,tao->niter,f,gnorm,0.0,step));
145     PetscCall((*tao->ops->convergencetest)(tao,tao->cnvP));
146     if (tao->reason != TAO_CONTINUE_ITERATING) {
147       break;
148     }
149 
150     /*  Check for restart condition */
151     PetscCall(VecDot(tao->gradient, cgP->G_old, &ginner));
152     if (PetscAbsScalar(ginner) >= cgP->eta * gnorm2) {
153       /*  Gradients far from orthogonal; use steepest descent direction */
154       beta = 0.0;
155     } else {
156       /*  Gradients close to orthogonal; use conjugate gradient formula */
157       switch (cgP->cg_type) {
158       case CG_FletcherReeves:
159         beta = gnorm2 / gnorm2_old;
160         break;
161 
162       case CG_PolakRibiere:
163         beta = (gnorm2 - ginner) / gnorm2_old;
164         break;
165 
166       case CG_PolakRibierePlus:
167         beta = PetscMax((gnorm2-ginner)/gnorm2_old, 0.0);
168         break;
169 
170       case CG_HestenesStiefel:
171         PetscCall(VecDot(tao->gradient, tao->stepdirection, &gd));
172         PetscCall(VecDot(cgP->G_old, tao->stepdirection, &gd_old));
173         beta = (gnorm2 - ginner) / (gd - gd_old);
174         break;
175 
176       case CG_DaiYuan:
177         PetscCall(VecDot(tao->gradient, tao->stepdirection, &gd));
178         PetscCall(VecDot(cgP->G_old, tao->stepdirection, &gd_old));
179         beta = gnorm2 / (gd - gd_old);
180         break;
181 
182       default:
183         beta = 0.0;
184         break;
185       }
186     }
187 
188     /*  Compute the direction d=-g + beta*d */
189     PetscCall(VecAXPBY(tao->stepdirection, -1.0, beta, tao->gradient));
190 
191     /*  update initial steplength choice */
192     delta = 1.0;
193     delta = PetscMax(delta, cgP->delta_min);
194     delta = PetscMin(delta, cgP->delta_max);
195   }
196   PetscFunctionReturn(0);
197 }
198 
199 static PetscErrorCode TaoSetUp_CG(Tao tao)
200 {
201   TAO_CG         *cgP = (TAO_CG*)tao->data;
202 
203   PetscFunctionBegin;
204   if (!tao->gradient) PetscCall(VecDuplicate(tao->solution,&tao->gradient));
205   if (!tao->stepdirection) PetscCall(VecDuplicate(tao->solution,&tao->stepdirection));
206   if (!cgP->X_old) PetscCall(VecDuplicate(tao->solution,&cgP->X_old));
207   if (!cgP->G_old) PetscCall(VecDuplicate(tao->gradient,&cgP->G_old));
208   PetscFunctionReturn(0);
209 }
210 
211 static PetscErrorCode TaoDestroy_CG(Tao tao)
212 {
213   TAO_CG         *cgP = (TAO_CG*) tao->data;
214 
215   PetscFunctionBegin;
216   if (tao->setupcalled) {
217     PetscCall(VecDestroy(&cgP->X_old));
218     PetscCall(VecDestroy(&cgP->G_old));
219   }
220   PetscCall(TaoLineSearchDestroy(&tao->linesearch));
221   PetscCall(PetscFree(tao->data));
222   PetscFunctionReturn(0);
223 }
224 
225 static PetscErrorCode TaoSetFromOptions_CG(PetscOptionItems *PetscOptionsObject,Tao tao)
226 {
227   TAO_CG         *cgP = (TAO_CG*)tao->data;
228 
229   PetscFunctionBegin;
230   PetscCall(TaoLineSearchSetFromOptions(tao->linesearch));
231   PetscOptionsHeadBegin(PetscOptionsObject,"Nonlinear Conjugate Gradient method for unconstrained optimization");
232   PetscCall(PetscOptionsReal("-tao_cg_eta","restart tolerance", "", cgP->eta,&cgP->eta,NULL));
233   PetscCall(PetscOptionsEList("-tao_cg_type","cg formula", "", CG_Table, CG_Types, CG_Table[cgP->cg_type], &cgP->cg_type,NULL));
234   PetscCall(PetscOptionsReal("-tao_cg_delta_min","minimum delta value", "", cgP->delta_min,&cgP->delta_min,NULL));
235   PetscCall(PetscOptionsReal("-tao_cg_delta_max","maximum delta value", "", cgP->delta_max,&cgP->delta_max,NULL));
236   PetscOptionsHeadEnd();
237   PetscFunctionReturn(0);
238 }
239 
240 static PetscErrorCode TaoView_CG(Tao tao, PetscViewer viewer)
241 {
242   PetscBool      isascii;
243   TAO_CG         *cgP = (TAO_CG*)tao->data;
244 
245   PetscFunctionBegin;
246   PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii));
247   if (isascii) {
248     PetscCall(PetscViewerASCIIPushTab(viewer));
249     PetscCall(PetscViewerASCIIPrintf(viewer, "CG Type: %s\n", CG_Table[cgP->cg_type]));
250     PetscCall(PetscViewerASCIIPrintf(viewer, "Gradient steps: %" PetscInt_FMT "\n", cgP->ngradsteps));
251     PetscCall(PetscViewerASCIIPrintf(viewer, "Reset steps: %" PetscInt_FMT "\n", cgP->nresetsteps));
252     PetscCall(PetscViewerASCIIPopTab(viewer));
253   }
254   PetscFunctionReturn(0);
255 }
256 
257 /*MC
258      TAOCG -   Nonlinear conjugate gradient method is an extension of the
259 nonlinear conjugate gradient solver for nonlinear optimization.
260 
261    Options Database Keys:
262 +      -tao_cg_eta <r> - restart tolerance
263 .      -tao_cg_type <taocg_type> - cg formula
264 .      -tao_cg_delta_min <r> - minimum delta value
265 -      -tao_cg_delta_max <r> - maximum delta value
266 
267   Notes:
268      CG formulas are:
269          "fr" - Fletcher-Reeves
270          "pr" - Polak-Ribiere
271          "prp" - Polak-Ribiere-Plus
272          "hs" - Hestenes-Steifel
273          "dy" - Dai-Yuan
274   Level: beginner
275 M*/
276 
277 PETSC_EXTERN PetscErrorCode TaoCreate_CG(Tao tao)
278 {
279   TAO_CG         *cgP;
280   const char     *morethuente_type = TAOLINESEARCHMT;
281 
282   PetscFunctionBegin;
283   tao->ops->setup = TaoSetUp_CG;
284   tao->ops->solve = TaoSolve_CG;
285   tao->ops->view = TaoView_CG;
286   tao->ops->setfromoptions = TaoSetFromOptions_CG;
287   tao->ops->destroy = TaoDestroy_CG;
288 
289   /* Override default settings (unless already changed) */
290   if (!tao->max_it_changed) tao->max_it = 2000;
291   if (!tao->max_funcs_changed) tao->max_funcs = 4000;
292 
293   /*  Note: nondefault values should be used for nonlinear conjugate gradient  */
294   /*  method.  In particular, gtol should be less that 0.5; the value used in  */
295   /*  Nocedal and Wright is 0.10.  We use the default values for the  */
296   /*  linesearch because it seems to work better. */
297   PetscCall(TaoLineSearchCreate(((PetscObject)tao)->comm, &tao->linesearch));
298   PetscCall(PetscObjectIncrementTabLevel((PetscObject)tao->linesearch, (PetscObject)tao, 1));
299   PetscCall(TaoLineSearchSetType(tao->linesearch, morethuente_type));
300   PetscCall(TaoLineSearchUseTaoRoutines(tao->linesearch, tao));
301   PetscCall(TaoLineSearchSetOptionsPrefix(tao->linesearch,tao->hdr.prefix));
302 
303   PetscCall(PetscNewLog(tao,&cgP));
304   tao->data = (void*)cgP;
305   cgP->eta = 0.1;
306   cgP->delta_min = 1e-7;
307   cgP->delta_max = 100;
308   cgP->cg_type = CG_PolakRibierePlus;
309   PetscFunctionReturn(0);
310 }
311