Actual source code: armijo.c
petsc-3.7.7 2017-09-25
1: #include <petsc/private/taolinesearchimpl.h>
2: #include <../src/tao/linesearch/impls/armijo/armijo.h>
4: #define REPLACE_FIFO 1
5: #define REPLACE_MRU 2
7: #define REFERENCE_MAX 1
8: #define REFERENCE_AVE 2
9: #define REFERENCE_MEAN 3
13: static PetscErrorCode TaoLineSearchDestroy_Armijo(TaoLineSearch ls)
14: {
15: TaoLineSearch_ARMIJO *armP = (TaoLineSearch_ARMIJO *)ls->data;
16: PetscErrorCode ierr;
19: PetscFree(armP->memory);
20: VecDestroy(&armP->x);
21: VecDestroy(&armP->work);
22: PetscFree(ls->data);
23: return(0);
24: }
28: static PetscErrorCode TaoLineSearchReset_Armijo(TaoLineSearch ls)
29: {
30: TaoLineSearch_ARMIJO *armP = (TaoLineSearch_ARMIJO *)ls->data;
31: PetscErrorCode ierr;
34: PetscFree(armP->memory);
35: armP->memorySetup = PETSC_FALSE;
36: return(0);
37: }
41: static PetscErrorCode TaoLineSearchSetFromOptions_Armijo(PetscOptionItems *PetscOptionsObject,TaoLineSearch ls)
42: {
43: TaoLineSearch_ARMIJO *armP = (TaoLineSearch_ARMIJO *)ls->data;
44: PetscErrorCode ierr;
47: PetscOptionsHead(PetscOptionsObject,"Armijo linesearch options");
48: PetscOptionsReal("-tao_ls_armijo_alpha", "initial reference constant", "", armP->alpha, &armP->alpha,NULL);
49: PetscOptionsReal("-tao_ls_armijo_beta_inf", "decrease constant one", "", armP->beta_inf, &armP->beta_inf,NULL);
50: PetscOptionsReal("-tao_ls_armijo_beta", "decrease constant", "", armP->beta, &armP->beta,NULL);
51: PetscOptionsReal("-tao_ls_armijo_sigma", "acceptance constant", "", armP->sigma, &armP->sigma,NULL);
52: PetscOptionsInt("-tao_ls_armijo_memory_size", "number of historical elements", "", armP->memorySize, &armP->memorySize,NULL);
53: PetscOptionsInt("-tao_ls_armijo_reference_policy", "policy for updating reference value", "", armP->referencePolicy, &armP->referencePolicy,NULL);
54: PetscOptionsInt("-tao_ls_armijo_replacement_policy", "policy for updating memory", "", armP->replacementPolicy, &armP->replacementPolicy,NULL);
55: PetscOptionsBool("-tao_ls_armijo_nondescending","Use nondescending armijo algorithm","",armP->nondescending,&armP->nondescending,NULL);
56: PetscOptionsTail();
57: return(0);
58: }
62: static PetscErrorCode TaoLineSearchView_Armijo(TaoLineSearch ls, PetscViewer pv)
63: {
64: TaoLineSearch_ARMIJO *armP = (TaoLineSearch_ARMIJO *)ls->data;
65: PetscBool isascii;
66: PetscErrorCode ierr;
69: PetscObjectTypeCompare((PetscObject)pv, PETSCVIEWERASCII, &isascii);
70: if (isascii) {
71: ierr=PetscViewerASCIIPrintf(pv," Armijo linesearch",armP->alpha);
72: if (armP->nondescending) {
73: PetscViewerASCIIPrintf(pv, " (nondescending)");
74: }
75: if (ls->bounded) {
76: PetscViewerASCIIPrintf(pv," (projected)");
77: }
78: ierr=PetscViewerASCIIPrintf(pv,": alpha=%g beta=%g ",(double)armP->alpha,(double)armP->beta);
79: ierr=PetscViewerASCIIPrintf(pv,"sigma=%g ",(double)armP->sigma);
80: ierr=PetscViewerASCIIPrintf(pv,"memsize=%D\n",armP->memorySize);
81: }
82: return(0);
83: }
87: /* @ TaoApply_Armijo - This routine performs a linesearch. It
88: backtracks until the (nonmonotone) Armijo conditions are satisfied.
90: Input Parameters:
91: + tao - Tao context
92: . X - current iterate (on output X contains new iterate, X + step*S)
93: . S - search direction
94: . f - merit function evaluated at X
95: . G - gradient of merit function evaluated at X
96: . W - work vector
97: - step - initial estimate of step length
99: Output parameters:
100: + f - merit function evaluated at new iterate, X + step*S
101: . G - gradient of merit function evaluated at new iterate, X + step*S
102: . X - new iterate
103: - step - final step length
105: @ */
106: static PetscErrorCode TaoLineSearchApply_Armijo(TaoLineSearch ls, Vec x, PetscReal *f, Vec g, Vec s)
107: {
108: TaoLineSearch_ARMIJO *armP = (TaoLineSearch_ARMIJO *)ls->data;
109: PetscErrorCode ierr;
110: PetscInt i;
111: PetscReal fact, ref, gdx;
112: PetscInt idx;
113: PetscBool g_computed=PETSC_FALSE; /* to prevent extra gradient computation */
117: ls->reason = TAOLINESEARCH_CONTINUE_ITERATING;
118: if (!armP->work) {
119: VecDuplicate(x,&armP->work);
120: armP->x = x;
121: PetscObjectReference((PetscObject)armP->x);
122: } else if (x != armP->x) {
123: /* If x has changed, then recreate work */
124: VecDestroy(&armP->work);
125: VecDuplicate(x,&armP->work);
126: PetscObjectDereference((PetscObject)armP->x);
127: armP->x = x;
128: PetscObjectReference((PetscObject)armP->x);
129: }
131: /* Check linesearch parameters */
132: if (armP->alpha < 1) {
133: PetscInfo1(ls,"Armijo line search error: alpha (%g) < 1\n", (double)armP->alpha);
134: ls->reason=TAOLINESEARCH_FAILED_BADPARAMETER;
135: } else if ((armP->beta <= 0) || (armP->beta >= 1)) {
136: PetscInfo1(ls,"Armijo line search error: beta (%g) invalid\n", (double)armP->beta);
137: ls->reason=TAOLINESEARCH_FAILED_BADPARAMETER;
138: } else if ((armP->beta_inf <= 0) || (armP->beta_inf >= 1)) {
139: PetscInfo1(ls,"Armijo line search error: beta_inf (%g) invalid\n", (double)armP->beta_inf);
140: ls->reason=TAOLINESEARCH_FAILED_BADPARAMETER;
141: } else if ((armP->sigma <= 0) || (armP->sigma >= 0.5)) {
142: PetscInfo1(ls,"Armijo line search error: sigma (%g) invalid\n", (double)armP->sigma);
143: ls->reason=TAOLINESEARCH_FAILED_BADPARAMETER;
144: } else if (armP->memorySize < 1) {
145: PetscInfo1(ls,"Armijo line search error: memory_size (%D) < 1\n", armP->memorySize);
146: ls->reason=TAOLINESEARCH_FAILED_BADPARAMETER;
147: } else if ((armP->referencePolicy != REFERENCE_MAX) && (armP->referencePolicy != REFERENCE_AVE) && (armP->referencePolicy != REFERENCE_MEAN)) {
148: PetscInfo(ls,"Armijo line search error: reference_policy invalid\n");
149: ls->reason=TAOLINESEARCH_FAILED_BADPARAMETER;
150: } else if ((armP->replacementPolicy != REPLACE_FIFO) && (armP->replacementPolicy != REPLACE_MRU)) {
151: PetscInfo(ls,"Armijo line search error: replacement_policy invalid\n");
152: ls->reason=TAOLINESEARCH_FAILED_BADPARAMETER;
153: } else if (PetscIsInfOrNanReal(*f)) {
154: PetscInfo(ls,"Armijo line search error: initial function inf or nan\n");
155: ls->reason=TAOLINESEARCH_FAILED_BADPARAMETER;
156: }
158: if (ls->reason != TAOLINESEARCH_CONTINUE_ITERATING) {
159: return(0);
160: }
162: /* Check to see of the memory has been allocated. If not, allocate
163: the historical array and populate it with the initial function
164: values. */
165: if (!armP->memory) {
166: PetscMalloc1(armP->memorySize, &armP->memory );
167: }
169: if (!armP->memorySetup) {
170: for (i = 0; i < armP->memorySize; i++) {
171: armP->memory[i] = armP->alpha*(*f);
172: }
174: armP->current = 0;
175: armP->lastReference = armP->memory[0];
176: armP->memorySetup=PETSC_TRUE;
177: }
179: /* Calculate reference value (MAX) */
180: ref = armP->memory[0];
181: idx = 0;
183: for (i = 1; i < armP->memorySize; i++) {
184: if (armP->memory[i] > ref) {
185: ref = armP->memory[i];
186: idx = i;
187: }
188: }
190: if (armP->referencePolicy == REFERENCE_AVE) {
191: ref = 0;
192: for (i = 0; i < armP->memorySize; i++) {
193: ref += armP->memory[i];
194: }
195: ref = ref / armP->memorySize;
196: ref = PetscMax(ref, armP->memory[armP->current]);
197: } else if (armP->referencePolicy == REFERENCE_MEAN) {
198: ref = PetscMin(ref, 0.5*(armP->lastReference + armP->memory[armP->current]));
199: }
200: VecDot(g,s,&gdx);
202: if (PetscIsInfOrNanReal(gdx)) {
203: PetscInfo1(ls,"Initial Line Search step * g is Inf or Nan (%g)\n",(double)gdx);
204: ls->reason=TAOLINESEARCH_FAILED_INFORNAN;
205: return(0);
206: }
207: if (gdx >= 0.0) {
208: PetscInfo1(ls,"Initial Line Search step is not descent direction (g's=%g)\n",(double)gdx);
209: ls->reason = TAOLINESEARCH_FAILED_ASCENT;
210: return(0);
211: }
213: if (armP->nondescending) {
214: fact = armP->sigma;
215: } else {
216: fact = armP->sigma * gdx;
217: }
218: ls->step = ls->initstep;
219: while (ls->step >= ls->stepmin && (ls->nfeval+ls->nfgeval) < ls->max_funcs) {
220: /* Calculate iterate */
221: VecCopy(x,armP->work);
222: VecAXPY(armP->work,ls->step,s);
223: if (ls->bounded) {
224: VecMedian(ls->lower,armP->work,ls->upper,armP->work);
225: }
227: /* Calculate function at new iterate */
228: if (ls->hasobjective) {
229: TaoLineSearchComputeObjective(ls,armP->work,f);
230: g_computed=PETSC_FALSE;
231: } else if (ls->usegts) {
232: TaoLineSearchComputeObjectiveAndGTS(ls,armP->work,f,&gdx);
233: g_computed=PETSC_FALSE;
234: } else {
235: TaoLineSearchComputeObjectiveAndGradient(ls,armP->work,f,g);
236: g_computed=PETSC_TRUE;
237: }
238: if (ls->step == ls->initstep) {
239: ls->f_fullstep = *f;
240: }
242: if (PetscIsInfOrNanReal(*f)) {
243: ls->step *= armP->beta_inf;
244: } else {
245: /* Check descent condition */
246: if (armP->nondescending && *f <= ref - ls->step*fact*ref)
247: break;
248: if (!armP->nondescending && *f <= ref + ls->step*fact) {
249: break;
250: }
252: ls->step *= armP->beta;
253: }
254: }
256: /* Check termination */
257: if (PetscIsInfOrNanReal(*f)) {
258: PetscInfo(ls, "Function is inf or nan.\n");
259: ls->reason = TAOLINESEARCH_FAILED_INFORNAN;
260: } else if (ls->step < ls->stepmin) {
261: PetscInfo(ls, "Step length is below tolerance.\n");
262: ls->reason = TAOLINESEARCH_HALTED_RTOL;
263: } else if ((ls->nfeval+ls->nfgeval) >= ls->max_funcs) {
264: PetscInfo2(ls, "Number of line search function evals (%D) > maximum allowed (%D)\n",ls->nfeval+ls->nfgeval, ls->max_funcs);
265: ls->reason = TAOLINESEARCH_HALTED_MAXFCN;
266: }
267: if (ls->reason) {
268: return(0);
269: }
271: /* Successful termination, update memory */
272: ls->reason = TAOLINESEARCH_SUCCESS;
273: armP->lastReference = ref;
274: if (armP->replacementPolicy == REPLACE_FIFO) {
275: armP->memory[armP->current++] = *f;
276: if (armP->current >= armP->memorySize) {
277: armP->current = 0;
278: }
279: } else {
280: armP->current = idx;
281: armP->memory[idx] = *f;
282: }
284: /* Update iterate and compute gradient */
285: VecCopy(armP->work,x);
286: if (!g_computed) {
287: TaoLineSearchComputeGradient(ls, x, g);
288: }
289: PetscInfo2(ls, "%D function evals in line search, step = %g\n",ls->nfeval, (double)ls->step);
290: return(0);
291: }
295: PETSC_EXTERN PetscErrorCode TaoLineSearchCreate_Armijo(TaoLineSearch ls)
296: {
297: TaoLineSearch_ARMIJO *armP;
298: PetscErrorCode ierr;
302: PetscNewLog(ls,&armP);
304: armP->memory = NULL;
305: armP->alpha = 1.0;
306: armP->beta = 0.5;
307: armP->beta_inf = 0.5;
308: armP->sigma = 1e-4;
309: armP->memorySize = 1;
310: armP->referencePolicy = REFERENCE_MAX;
311: armP->replacementPolicy = REPLACE_MRU;
312: armP->nondescending=PETSC_FALSE;
313: ls->data = (void*)armP;
314: ls->initstep=1.0;
315: ls->ops->setup=0;
316: ls->ops->apply=TaoLineSearchApply_Armijo;
317: ls->ops->view = TaoLineSearchView_Armijo;
318: ls->ops->destroy = TaoLineSearchDestroy_Armijo;
319: ls->ops->reset = TaoLineSearchReset_Armijo;
320: ls->ops->setfromoptions = TaoLineSearchSetFromOptions_Armijo;
321: return(0);
322: }