xref: /petsc/src/tao/unconstrained/impls/lmvm/lmvm.c (revision f89ca46fb01025fa5f21ef09d10cb4723982ea5b)
1 #include "taolinesearch.h"
2 #include "../src/tao/matrix/lmvmmat.h"
3 #include "lmvm.h"
4 
5 #define LMVM_BFGS                0
6 #define LMVM_SCALED_GRADIENT     1
7 #define LMVM_GRADIENT            2
8 
9 #undef __FUNCT__
10 #define __FUNCT__ "TaoSolve_LMVM"
11 static PetscErrorCode TaoSolve_LMVM(TaoSolver tao)
12 {
13 
14   TAO_LMVM *lmP = (TAO_LMVM *)tao->data;
15 
16   PetscReal f, fold, gdx, gnorm;
17   PetscReal step = 1.0;
18 
19   PetscReal delta;
20 
21   PetscErrorCode ierr;
22   PetscInt stepType;
23   PetscInt iter = 0;
24   TaoSolverTerminationReason reason = TAO_CONTINUE_ITERATING;
25   TaoLineSearchTerminationReason ls_status = TAOLINESEARCH_CONTINUE_ITERATING;
26 
27   PetscFunctionBegin;
28 
29   if (tao->XL || tao->XU || tao->ops->computebounds) {
30     ierr = PetscPrintf(((PetscObject)tao)->comm,"WARNING: Variable bounds have been set but will be ignored by lmvm algorithm\n"); CHKERRQ(ierr);
31   }
32 
33   /*  Check convergence criteria */
34   ierr = TaoComputeObjectiveAndGradient(tao, tao->solution, &f, tao->gradient); CHKERRQ(ierr);
35   ierr = VecNorm(tao->gradient,NORM_2,&gnorm); CHKERRQ(ierr);
36   if (PetscIsInfOrNanReal(f) || PetscIsInfOrNanReal(gnorm)) {
37     SETERRQ(PETSC_COMM_SELF,1, "User provided compute function generated Inf or NaN");
38   }
39 
40   ierr = TaoMonitor(tao, iter, f, gnorm, 0.0, step, &reason); CHKERRQ(ierr);
41   if (reason != TAO_CONTINUE_ITERATING) {
42     PetscFunctionReturn(0);
43   }
44 
45   /*  Set initial scaling for the function */
46   if (f != 0.0) {
47     delta = 2.0 * PetscAbsScalar(f) / (gnorm*gnorm);
48   }
49   else {
50     delta = 2.0 / (gnorm*gnorm);
51   }
52   ierr = MatLMVMSetDelta(lmP->M,delta); CHKERRQ(ierr);
53 
54   /*  Set counter for gradient/reset steps */
55   lmP->bfgs = 0;
56   lmP->sgrad = 0;
57   lmP->grad = 0;
58 
59   /*  Have not converged; continue with Newton method */
60   while (reason == TAO_CONTINUE_ITERATING) {
61     /*  Compute direction */
62     ierr = MatLMVMUpdate(lmP->M,tao->solution,tao->gradient); CHKERRQ(ierr);
63     ierr = MatLMVMSolve(lmP->M, tao->gradient, lmP->D); CHKERRQ(ierr);
64     ++lmP->bfgs;
65 
66     /*  Check for success (descent direction) */
67     ierr = VecDot(lmP->D, tao->gradient, &gdx); CHKERRQ(ierr);
68     if ((gdx <= 0.0) || PetscIsInfOrNanReal(gdx)) {
69       /* Step is not descent or direction produced not a number
70          We can assert bfgsUpdates > 1 in this case because
71          the first solve produces the scaled gradient direction,
72          which is guaranteed to be descent
73 
74          Use steepest descent direction (scaled)
75       */
76 
77       ++lmP->grad;
78 
79       if (f != 0.0) {
80         delta = 2.0 * PetscAbsScalar(f) / (gnorm*gnorm);
81       }
82       else {
83         delta = 2.0 / (gnorm*gnorm);
84       }
85       ierr = MatLMVMSetDelta(lmP->M, delta); CHKERRQ(ierr);
86       ierr = MatLMVMReset(lmP->M); CHKERRQ(ierr);
87       ierr = MatLMVMUpdate(lmP->M, tao->solution, tao->gradient); CHKERRQ(ierr);
88       ierr = MatLMVMSolve(lmP->M,tao->gradient, lmP->D); CHKERRQ(ierr);
89 
90       /* On a reset, the direction cannot be not a number; it is a
91 	 scaled gradient step.  No need to check for this condition. */
92 
93       lmP->bfgs = 1;
94       ++lmP->sgrad;
95       stepType = LMVM_SCALED_GRADIENT;
96     }
97     else {
98       if (1 == lmP->bfgs) {
99         /*  The first BFGS direction is always the scaled gradient */
100         ++lmP->sgrad;
101         stepType = LMVM_SCALED_GRADIENT;
102       }
103       else {
104         ++lmP->bfgs;
105         stepType = LMVM_BFGS;
106       }
107     }
108     ierr = VecScale(lmP->D, -1.0); CHKERRQ(ierr);
109 
110     /*  Perform the linesearch */
111     fold = f;
112     ierr = VecCopy(tao->solution, lmP->Xold); CHKERRQ(ierr);
113     ierr = VecCopy(tao->gradient, lmP->Gold); CHKERRQ(ierr);
114 
115     ierr = TaoLineSearchApply(tao->linesearch, tao->solution, &f, tao->gradient, lmP->D, &step,&ls_status); CHKERRQ(ierr);
116     ierr = TaoAddLineSearchCounts(tao); CHKERRQ(ierr);
117 
118 
119     while (ls_status != TAOLINESEARCH_SUCCESS &&
120 	   ls_status != TAOLINESEARCH_SUCCESS_USER
121 	   && (stepType != LMVM_GRADIENT)) {
122       /*  Linesearch failed */
123       /*  Reset factors and use scaled gradient step */
124       f = fold;
125       ierr = VecCopy(lmP->Xold, tao->solution); CHKERRQ(ierr);
126       ierr = VecCopy(lmP->Gold, tao->gradient); CHKERRQ(ierr);
127 
128       switch(stepType) {
129       case LMVM_BFGS:
130         /*  Failed to obtain acceptable iterate with BFGS step */
131         /*  Attempt to use the scaled gradient direction */
132 
133         if (f != 0.0) {
134           delta = 2.0 * PetscAbsScalar(f) / (gnorm*gnorm);
135         }
136         else {
137           delta = 2.0 / (gnorm*gnorm);
138         }
139 	ierr = MatLMVMSetDelta(lmP->M, delta); CHKERRQ(ierr);
140 	ierr = MatLMVMReset(lmP->M); CHKERRQ(ierr);
141 	ierr = MatLMVMUpdate(lmP->M, tao->solution, tao->gradient); CHKERRQ(ierr);
142 	ierr = MatLMVMSolve(lmP->M, tao->gradient, lmP->D); CHKERRQ(ierr);
143 
144         /* On a reset, the direction cannot be not a number; it is a
145 	   scaled gradient step.  No need to check for this condition. */
146 
147 	lmP->bfgs = 1;
148 	++lmP->sgrad;
149 	stepType = LMVM_SCALED_GRADIENT;
150 	break;
151 
152       case LMVM_SCALED_GRADIENT:
153         /* The scaled gradient step did not produce a new iterate;
154 	   attempt to use the gradient direction.
155 	   Need to make sure we are not using a different diagonal scaling */
156 	ierr = MatLMVMSetDelta(lmP->M, 1.0); CHKERRQ(ierr);
157 	ierr = MatLMVMReset(lmP->M); CHKERRQ(ierr);
158 	ierr = MatLMVMUpdate(lmP->M, tao->solution, tao->gradient); CHKERRQ(ierr);
159 	ierr = MatLMVMSolve(lmP->M, tao->gradient, lmP->D); CHKERRQ(ierr);
160 
161         lmP->bfgs = 1;
162         ++lmP->grad;
163         stepType = LMVM_GRADIENT;
164         break;
165       }
166       ierr = VecScale(lmP->D, -1.0); CHKERRQ(ierr);
167 
168       /*  Perform the linesearch */
169       ierr = TaoLineSearchApply(tao->linesearch, tao->solution, &f, tao->gradient, lmP->D, &step, &ls_status); CHKERRQ(ierr);
170       ierr = TaoAddLineSearchCounts(tao); CHKERRQ(ierr);
171 
172     }
173 
174     if (ls_status != TAOLINESEARCH_SUCCESS &&
175 	ls_status != TAOLINESEARCH_SUCCESS_USER) {
176       /*  Failed to find an improving point */
177       f = fold;
178       ierr = VecCopy(lmP->Xold, tao->solution); CHKERRQ(ierr);
179       ierr = VecCopy(lmP->Gold, tao->gradient); CHKERRQ(ierr);
180       step = 0.0;
181       reason = TAO_DIVERGED_LS_FAILURE;
182       tao->reason = TAO_DIVERGED_LS_FAILURE;
183     }
184     /*  Check for termination */
185     ierr = VecNorm(tao->gradient, NORM_2, &gnorm); CHKERRQ(ierr);
186     iter++;
187     ierr = TaoMonitor(tao,iter,f,gnorm,0.0,step,&reason); CHKERRQ(ierr);
188   }
189   PetscFunctionReturn(0);
190 }
191 #undef __FUNCT__
192 #define __FUNCT__ "TaoSetUp_LMVM"
193 static PetscErrorCode TaoSetUp_LMVM(TaoSolver tao)
194 {
195   TAO_LMVM *lmP = (TAO_LMVM *)tao->data;
196   PetscInt n,N;
197   PetscErrorCode ierr;
198 
199   PetscFunctionBegin;
200   /* Existence of tao->solution checked in TaoSetUp() */
201   if (!tao->gradient) {ierr = VecDuplicate(tao->solution,&tao->gradient); CHKERRQ(ierr);  }
202   if (!tao->stepdirection) {ierr = VecDuplicate(tao->solution,&tao->stepdirection); CHKERRQ(ierr);  }
203   if (!lmP->D) {ierr = VecDuplicate(tao->solution,&lmP->D); CHKERRQ(ierr);  }
204   if (!lmP->Xold) {ierr = VecDuplicate(tao->solution,&lmP->Xold); CHKERRQ(ierr);  }
205   if (!lmP->Gold) {ierr = VecDuplicate(tao->solution,&lmP->Gold); CHKERRQ(ierr);  }
206 
207   /*  Create matrix for the limited memory approximation */
208   ierr = VecGetLocalSize(tao->solution,&n); CHKERRQ(ierr);
209   ierr = VecGetSize(tao->solution,&N); CHKERRQ(ierr);
210   ierr = MatCreateLMVM(((PetscObject)tao)->comm,n,N,&lmP->M); CHKERRQ(ierr);
211   ierr = MatLMVMAllocateVectors(lmP->M,tao->solution); CHKERRQ(ierr);
212 
213 
214   PetscFunctionReturn(0);
215 }
216 
217 /* ---------------------------------------------------------- */
218 #undef __FUNCT__
219 #define __FUNCT__ "TaoDestroy_LMVM"
220 static PetscErrorCode TaoDestroy_LMVM(TaoSolver tao)
221 {
222 
223   TAO_LMVM *lmP = (TAO_LMVM *)tao->data;
224   PetscErrorCode ierr;
225 
226   PetscFunctionBegin;
227   if (tao->setupcalled) {
228     ierr = VecDestroy(&lmP->Xold); CHKERRQ(ierr);
229     ierr = VecDestroy(&lmP->Gold); CHKERRQ(ierr);
230     ierr = VecDestroy(&lmP->D); CHKERRQ(ierr);
231     ierr = MatDestroy(&lmP->M); CHKERRQ(ierr);
232   }
233   ierr = PetscFree(tao->data); CHKERRQ(ierr);
234   tao->data = PETSC_NULL;
235 
236   PetscFunctionReturn(0);
237 }
238 
239 /*------------------------------------------------------------*/
240 #undef __FUNCT__
241 #define __FUNCT__ "TaoSetFromOptions_LMVM"
242 static PetscErrorCode TaoSetFromOptions_LMVM(TaoSolver tao)
243 {
244 
245   PetscErrorCode ierr;
246 
247   PetscFunctionBegin;
248   ierr = PetscOptionsHead("Limited-memory variable-metric method for unconstrained optimization"); CHKERRQ(ierr);
249   ierr = TaoLineSearchSetFromOptions(tao->linesearch); CHKERRQ(ierr);
250   ierr = PetscOptionsTail(); CHKERRQ(ierr);
251   PetscFunctionReturn(0);
252 
253   PetscFunctionReturn(0);
254 }
255 
256 /*------------------------------------------------------------*/
257 #undef __FUNCT__
258 #define __FUNCT__ "TaoView_LMVM"
259 static PetscErrorCode TaoView_LMVM(TaoSolver tao, PetscViewer viewer)
260 {
261 
262     TAO_LMVM *lm = (TAO_LMVM *)tao->data;
263     PetscBool isascii;
264     PetscErrorCode ierr;
265 
266 
267     PetscFunctionBegin;
268     ierr = PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii); CHKERRQ(ierr);
269     if (isascii) {
270 
271         ierr = PetscViewerASCIIPushTab(viewer); CHKERRQ(ierr);
272 	ierr = PetscViewerASCIIPrintf(viewer, "BFGS steps: %D\n", lm->bfgs); CHKERRQ(ierr);
273 	ierr = PetscViewerASCIIPrintf(viewer, "Scaled gradient steps: %D\n", lm->sgrad); CHKERRQ(ierr);
274 	ierr = PetscViewerASCIIPrintf(viewer, "Gradient steps: %D\n", lm->grad); CHKERRQ(ierr);
275         ierr = PetscViewerASCIIPopTab(viewer); CHKERRQ(ierr);
276     } else {
277       SETERRQ1(((PetscObject)tao)->comm,PETSC_ERR_SUP,"Viewer type %s not supported for TAO LMVM",((PetscObject)viewer)->type_name);
278     }
279     PetscFunctionReturn(0);
280 }
281 
282 /* ---------------------------------------------------------- */
283 
284 EXTERN_C_BEGIN
285 #undef __FUNCT__
286 #define __FUNCT__ "TaoCreate_LMVM"
287 PetscErrorCode TaoCreate_LMVM(TaoSolver tao)
288 {
289 
290   TAO_LMVM *lmP;
291   const char *morethuente_type = TAOLINESEARCH_MT;
292   PetscErrorCode ierr;
293 
294   PetscFunctionBegin;
295   tao->ops->setup = TaoSetUp_LMVM;
296   tao->ops->solve = TaoSolve_LMVM;
297   tao->ops->view = TaoView_LMVM;
298   tao->ops->setfromoptions = TaoSetFromOptions_LMVM;
299   tao->ops->destroy = TaoDestroy_LMVM;
300 
301   ierr = PetscNewLog(tao,TAO_LMVM, &lmP); CHKERRQ(ierr);
302   lmP->D = 0;
303   lmP->M = 0;
304   lmP->Xold = 0;
305   lmP->Gold = 0;
306 
307   tao->data = (void*)lmP;
308   tao->max_it = 2000;
309   tao->max_funcs = 4000;
310   tao->fatol = 1e-4;
311   tao->frtol = 1e-4;
312 
313   ierr = TaoLineSearchCreate(((PetscObject)tao)->comm,&tao->linesearch); CHKERRQ(ierr);
314   ierr = TaoLineSearchSetType(tao->linesearch,morethuente_type); CHKERRQ(ierr);
315   ierr = TaoLineSearchUseTaoSolverRoutines(tao->linesearch,tao); CHKERRQ(ierr);
316 
317   PetscFunctionReturn(0);
318 }
319 EXTERN_C_END
320