Actual source code: boundproj.c
2: #include "boundproj.h" /*I "tao_solver.h" I*/
6: static int TaoDestroy_LineSearch(TAO_SOLVER tao,void *linectx)
7: {
8: int info;
9: TAO_LINESEARCH2 *ctx = (TAO_LINESEARCH2 *)linectx;
11: TaoFunctionBegin;
12: if (ctx->setupcalled==1){
13: info=TaoVecDestroy(ctx->W2);CHKERRQ(info);
14: }
15: info = TaoFree(ctx);CHKERRQ(info);
16: TaoFunctionReturn(0);
17: }
21: static int TaoSetOptions_LineSearch(TAO_SOLVER tao,void *linectx)
22: {
23: TAO_LINESEARCH2 *ctx = (TAO_LINESEARCH2 *)linectx;
24: int info;
26: TaoFunctionBegin;
27: info = TaoOptionsHead("Projected line search options");CHKERRQ(info);
28: info = TaoOptionInt("-tao_ls_maxfev","max function evals in line search","",ctx->maxfev,&ctx->maxfev,0);CHKERRQ(info);
29: info = TaoOptionDouble("-tao_ls_ftol","tol for sufficient decrease","",ctx->ftol,&ctx->ftol,0);CHKERRQ(info);
30: info = TaoOptionDouble("-tao_ls_stepmin","lower bound for step","",ctx->stepmin,&ctx->stepmin,0);CHKERRQ(info);
31: info = TaoOptionDouble("-tao_ls_stepmax","upper bound for step","",ctx->stepmax,&ctx->stepmax,0);CHKERRQ(info);
32: info = TaoOptionsTail();CHKERRQ(info);
34: TaoFunctionReturn(0);
35: }
39: static int TaoView_LineSearch(TAO_SOLVER tao,void *ctx)
40: {
41: TAO_LINESEARCH2 *ls = (TAO_LINESEARCH2 *)ctx;
42: int info;
44: TaoFunctionBegin;
45: info = TaoPrintInt(tao," Line search: maxf=%d,",ls->maxfev);CHKERRQ(info);
46: info = TaoPrintDouble(tao," ftol=%g\n",ls->ftol);CHKERRQ(info);
47: TaoFunctionReturn(0);
48: }
52: /* @ TaoApply_LineSearch - This routine performs a line search. It
53: backtracks until the Armijo conditions are satisfied.
55: Input Parameters:
56: + tao - TAO_SOLVER context
57: . X - current iterate (on output X contains new iterate, X + step*S)
58: . S - search direction
59: . f - objective function evaluated at X
60: . G - gradient evaluated at X
61: . W - work vector
62: . gdx - inner product of gradient and the direction of the first linear manifold being searched
63: - step - initial estimate of step length
65: Output parameters:
66: + f - objective function evaluated at new iterate, X + step*S
67: . G - gradient evaluated at new iterate, X + step*S
68: . X - new iterate
69: - step - final step length
71: Info is set to one of:
72: . 0 - the line search succeeds; the sufficient decrease
73: condition and the directional derivative condition hold
75: negative number if an input parameter is invalid
76: . -1 - step < 0
77: . -2 - ftol < 0
78: - -7 - maxfev < 0
80: positive number > 1 if the line search otherwise terminates
81: + 2 - Relative width of the interval of uncertainty is
82: at most rtol.
83: . 3 - Maximum number of function evaluations (maxfev) has
84: been reached.
85: . 4 - Step is at the lower bound, stepmin.
86: . 5 - Step is at the upper bound, stepmax.
87: . 6 - Rounding errors may prevent further progress.
88: There may not be a step that satisfies the
89: sufficient decrease and curvature conditions.
90: Tolerances may be too small.
91: + 7 - Search direction is not a descent direction.
93: Notes:
94: This routine is used within the TAO_NLS method.
95: @ */
96: static int TaoApply_LineSearch(TAO_SOLVER tao,TaoVec* X,TaoVec* G,TaoVec* S,TaoVec* Gold,double *f,double *f_full,
97: double *step,int *info2,void*ctx)
98: {
99: TAO_LINESEARCH2 *neP = (TAO_LINESEARCH2 *) ctx;
100: int info, i;
101: double zero=0.0;
102: double finit,gdx,dginit,actred,prered,rho,d1=0,d2=0,d3=0;
103: TaoVec* XL, *XU, *Xold=neP->W2;
104: TaoTruth flag;
106: TaoFunctionBegin;
107: /* neP->stepmin - lower bound for step */
108: /* neP->stepmax - upper bound for step */
109: /* neP->rtol - relative tolerance for an acceptable step */
110: /* neP->ftol - tolerance for sufficient decrease condition */
111: /* neP->gtol - tolerance for curvature condition */
112: /* neP->nfev - number of function evaluations */
113: /* neP->maxfev - maximum number of function evaluations */
115: /* Check input parameters for errors */
117: *info2 = 0;
118: /* After 2 failures, Check that search direction is a descent direction */
119: /* This test is not really sufficient. */
120: if (neP->setupcalled){
121: info=X->Compatible(neP->W2,&flag); CHKERRQ(info);
122: if (flag==TAO_FALSE){
123: info=TaoVecDestroy(neP->W2); CHKERRQ(info);neP->W2=0;
124: neP->setupcalled=0;
125: }
126: }
127: if (neP->setupcalled==0){
128: info=X->Clone(&neP->W2); CHKERRQ(info);
129: Xold=neP->W2;
130: neP->setupcalled=1;
131: }
132: info = TaoGetVariableBounds(tao,&XL,&XU); CHKERRQ(info);
133: info = Gold->ScaleCopyFrom(-1.0,S); CHKERRQ(info);
134: info = Gold->BoundGradientProjection(S,XL,X,XU); CHKERRQ(info);
135: info = Gold->Dot(G,&dginit);CHKERRQ(info); /* dginit = G^T S */
136: dginit*=-1;
137: gdx=dginit;
138: if (dginit >= zero) {
139: info = G->CopyFrom(Gold); CHKERRQ(info);
140: info = X->CopyFrom(Xold); CHKERRQ(info);
141: info = PetscInfo(tao,"TaoApply_LineSearch2:Search direction not a descent direction\n"); CHKERRQ(info);
142: *info2 = 7; TaoFunctionReturn(0);
143: }
145: info = Xold->CopyFrom(X); CHKERRQ(info);
146: info = Gold->CopyFrom(G); CHKERRQ(info);
147: info=TaoGetVariableBounds(tao,&XL,&XU); CHKERRQ(info);
148: if (*step < zero) {
149: info = PetscInfo1(tao,"TaoApply_LineSearch:Line search error: step (%g) < 0\n",*step); CHKERRQ(info);
150: *info2 = -1; TaoFunctionReturn(0);
151: } else if (neP->ftol < zero) {
152: info = PetscInfo1(tao,"TaoApply_LineSearch:Line search error: ftol (%g) < 0\n",neP->ftol); CHKERRQ(info);
153: *info2 = -2; TaoFunctionReturn(0);
154: } else if (neP->maxfev < zero) {
155: info = PetscInfo1(tao,"TaoApply_LineSearch:Line search error: maxfev (%d) < 0\n",neP->maxfev); CHKERRQ(info);
156: *info2 = -7; TaoFunctionReturn(0);
157: }
159: /* Initialization */
160: neP->nfev = 0;
161: finit = *f;
162: for (i=0; i< neP->maxfev; i++) {
163:
164: /* Force the step to be within the bounds */
165: *step = TaoMax(*step,neP->stepmin);
166: *step = TaoMin(*step,neP->stepmax);
167:
168: if (0==i){
169: info = X->Axpy(*step,S);CHKERRQ(info);
170: info = X->Median(XL,X,XU);CHKERRQ(info);
171: info = TaoComputeMeritFunctionGradient(tao,X,f,G);
172: *f_full = *f;
174: neP->nfev++;
175: if (info==0 && *f==*f){ /* Successful function evaluation */
176: actred = *f - finit;
177: rho = actred/( (*step) * (-TaoAbsScalar(gdx)) );
178: if (actred < 0 && rho > neP->ftol){
179: break;
180: } else{
181: info=Xold->StepBoundInfo(XL,XU,S,&d3,&d2,&d1);CHKERRQ(info);
182: info = PetscInfo1(tao,"stepmax = %10f\n",d1); CHKERRQ(info);
184: *step = TaoMin(*step,d1);
185: info = Gold->Dot(X,&d1); CHKERRQ(info);
186: info = Gold->Dot(Xold,&prered); CHKERRQ(info);
187: prered=d1-prered;
188: rho = actred/(-TaoAbsScalar(prered));
189: if (actred < 0 && rho > neP->ftol){
190: break;
191: } else{
192: *step = (*step)/2;
193: }
194: }
195: } else { /* Function could not be evaluated at this point */
196: *step = (*step)*0.7;
197: }
199: } else {
200: info = X->Waxpby(*step,S,1.0,Xold);CHKERRQ(info);
201: info = X->Median(XL,X,XU);CHKERRQ(info);
202:
203: info = G->Waxpby(-1,Xold,1.0,X);CHKERRQ(info);
204: info = G->Dot(Gold,&prered); CHKERRQ(info);
205: info = TaoComputeMeritFunctionGradient(tao,X,f,G); CHKERRQ(info);
206: neP->nfev++;
207: if (info==0 && *f==*f){ /* Successful function evaluation */
208: actred = *f - finit;
209: rho = actred/(-TaoAbsScalar(prered));
210: /*
211: If sufficient progress has been obtained, accept the
212: point. Prered could be positive or negative.
213: Otherwise, backtrack.
214: */
216: if (actred < 0 && rho > neP->ftol){
217: info = PetscInfo4(tao,"TaoApply_LineSearch: steplength: %g, actual reduction: %g (hopefully positive), Predicted reduction: %g, rho: %g\n",*step,-actred,prered,rho); CHKERRQ(info);
218: break;
219: } else {
220: info = PetscInfo4(tao,"TaoApply_LineSearch: steplength: %g, actual reduction: %g (hopefully positive), Predicted reduction: %g, rho: %g\n",*step,-actred,prered,rho); CHKERRQ(info);
221: if (i<5){
222: *step = (*step)/2;
223: } else {
224: *step = (*step)/2;
225: }
226: }
227: } else { /* Function could not be evaluated at this point */
228: info = PetscInfo(tao,"TaoApply_LineSearch: Problem in function evaluation\n"); CHKERRQ(info);
229: *step = (*step)*0.7;
230: }
231: }
233: }
234: /* Convergence testing */
235:
236: if (*step <= neP->stepmin || *step >= neP->stepmax) {
237: *info2 = 6;
238: info = PetscInfo(tao,"TaoApply_LineSearch:Rounding errors may prevent further progress. May not be a step satisfying\n"); CHKERRQ(info);
239: info = PetscInfo(tao,"TaoApply_LineSearch:sufficient decrease and curvature conditions. Tolerances may be too small.\n"); CHKERRQ(info);
240: }
241: if (*step == neP->stepmax) {
242: info = PetscInfo1(tao,"TaoApply_LineSearch:Step is at the upper bound, stepmax (%g)\n",neP->stepmax); CHKERRQ(info);
243: *info2 = 5;
244: }
245: if (*step == neP->stepmin) {
246: info = PetscInfo1(tao,"TaoApply_LineSearch:Step is at the lower bound, stepmin (%g)\n",neP->stepmin); CHKERRQ(info);
247: *info2 = 4;
248: }
249: if (neP->nfev >= neP->maxfev) {
250: info = PetscInfo2(tao,"TaoApply_LineSearch:Number of line search function evals (%d) > maximum (%d)\n",neP->nfev,neP->maxfev); CHKERRQ(info);
251: *info2 = 3;
252: }
253: if ((neP->bracket) && (neP->stepmax - neP->stepmin <= neP->rtol*neP->stepmax)){
254: info = PetscInfo1(tao,"TaoApply_LineSearch:Relative width of interval of uncertainty is at most rtol (%g)\n",neP->rtol); CHKERRQ(info);
255: *info2 = 2;
256: }
257: /*
258: if ((*f <= ftest1) && (TaoAbsDouble(dg) <= neP->gtol*(-dginit))) {
259: info = PetscLogInfo((tao,"TaoApply_LineSearch:Line search success: Sufficient decrease and directional deriv conditions hold\n")); CHKERRQ(info);
260: *info2 = 1;
261: }
262: */
263:
264: /* Finish computations */
265: /*
266: info = PetscLogInfo((tao,"TaoApply_LineSearch:%d function evals in line search, step = %10.4f\n",neP->nfev,*step)); CHKERRQ(info);
267: */
268: TaoFunctionReturn(0);
269: }
273: /*@C
274: TaoCreateProjectedLineSearch - Create a line search
276: Input Parameters:
277: . tao - TAO_SOLVER context
279: Note:
280: This routine is used within the following TAO bound constrained
281: minimization solvers: TRON (tao_tron) and limited memory variable metric
282: (tao_blmvm). This line search projects points onto the feasible, bound
283: constrained region. It only enforces the Armijo descent condition.
285: Level: developer
287: .keywords: TAO_SOLVER, linesearch
288: @*/
289: int TaoCreateProjectedLineSearch(TAO_SOLVER tao)
290: {
291: int info;
292: TAO_LINESEARCH2 *neP;
294: TaoFunctionBegin;
296: info = TaoNew(TAO_LINESEARCH2,&neP); CHKERRQ(info);
297: info = PetscLogObjectMemory(tao,sizeof(TAO_LINESEARCH2)); CHKERRQ(info);
299: neP->ftol = 0.001;
300: neP->rtol = 0.0;
301: neP->gtol = 0.0;
302: neP->stepmin = 1.0e-30;
303: neP->stepmax = 1.0e+20;
304: neP->nfev = 0;
305: neP->bracket = 0;
306: neP->infoc = 1;
307: neP->maxfev = 30;
308: neP->W2 = TAO_NULL;
309: neP->setupcalled = 0;
311: info = TaoSetLineSearch(tao,0,
312: TaoSetOptions_LineSearch,
313: TaoApply_LineSearch,
314: TaoView_LineSearch,
315: TaoDestroy_LineSearch,
316: (void *) neP);CHKERRQ(info);
318: TaoFunctionReturn(0);
319: }