xref: /petsc/src/tao/unconstrained/impls/lmvm/lmvm.c (revision 5e71baeff2f3138f93cd4f5927dfd596eb8325cc)
1 #include <petsctaolinesearch.h>
2 #include <../src/tao/matrix/lmvmmat.h>
3 #include <../src/tao/unconstrained/impls/lmvm/lmvm.h>
4 
5 #define LMVM_BFGS                0
6 #define LMVM_SCALED_GRADIENT     1
7 #define LMVM_GRADIENT            2
8 
9 static PetscErrorCode TaoSolve_LMVM(Tao tao)
10 {
11   TAO_LMVM                     *lmP = (TAO_LMVM *)tao->data;
12   PetscReal                    f, fold, gdx, gnorm;
13   PetscReal                    step = 1.0;
14   PetscReal                    delta;
15   PetscErrorCode               ierr;
16   PetscInt                     stepType;
17   TaoConvergedReason           reason = TAO_CONTINUE_ITERATING;
18   TaoLineSearchConvergedReason ls_status = TAOLINESEARCH_CONTINUE_ITERATING;
19 
20   PetscFunctionBegin;
21 
22   if (tao->XL || tao->XU || tao->ops->computebounds) {
23     ierr = PetscPrintf(((PetscObject)tao)->comm,"WARNING: Variable bounds have been set but will be ignored by lmvm algorithm\n");CHKERRQ(ierr);
24   }
25 
26   /*  Check convergence criteria */
27   ierr = TaoComputeObjectiveAndGradient(tao, tao->solution, &f, tao->gradient);CHKERRQ(ierr);
28   ierr = TaoGradientNorm(tao, tao->gradient,NORM_2,&gnorm);CHKERRQ(ierr);
29 
30   if (PetscIsInfOrNanReal(f) || PetscIsInfOrNanReal(gnorm)) SETERRQ(PETSC_COMM_SELF,1, "User provided compute function generated Inf or NaN");
31 
32   ierr = TaoMonitor(tao, tao->niter, f, gnorm, 0.0, step, &reason);CHKERRQ(ierr);
33   if (reason != TAO_CONTINUE_ITERATING) PetscFunctionReturn(0);
34 
35   /*  Set initial scaling for the function */
36   if (f != 0.0) {
37     delta = 2.0 * PetscAbsScalar(f) / (gnorm*gnorm);
38   } else {
39     delta = 2.0 / (gnorm*gnorm);
40   }
41   ierr = MatLMVMSetDelta(lmP->M,delta);CHKERRQ(ierr);
42 
43   /*  Set counter for gradient/reset steps */
44   lmP->bfgs = 0;
45   lmP->sgrad = 0;
46   lmP->grad = 0;
47 
48   /*  Have not converged; continue with Newton method */
49   while (reason == TAO_CONTINUE_ITERATING) {
50     /*  Compute direction */
51     ierr = MatLMVMUpdate(lmP->M,tao->solution,tao->gradient);CHKERRQ(ierr);
52     ierr = MatLMVMSolve(lmP->M, tao->gradient, lmP->D);CHKERRQ(ierr);
53     ++lmP->bfgs;
54 
55     /*  Check for success (descent direction) */
56     ierr = VecDot(lmP->D, tao->gradient, &gdx);CHKERRQ(ierr);
57     if ((gdx <= 0.0) || PetscIsInfOrNanReal(gdx)) {
58       /* Step is not descent or direction produced not a number
59          We can assert bfgsUpdates > 1 in this case because
60          the first solve produces the scaled gradient direction,
61          which is guaranteed to be descent
62 
63          Use steepest descent direction (scaled)
64       */
65 
66       ++lmP->grad;
67 
68       if (f != 0.0) {
69         delta = 2.0 * PetscAbsScalar(f) / (gnorm*gnorm);
70       } else {
71         delta = 2.0 / (gnorm*gnorm);
72       }
73       ierr = MatLMVMSetDelta(lmP->M, delta);CHKERRQ(ierr);
74       ierr = MatLMVMReset(lmP->M);CHKERRQ(ierr);
75       ierr = MatLMVMUpdate(lmP->M, tao->solution, tao->gradient);CHKERRQ(ierr);
76       ierr = MatLMVMSolve(lmP->M,tao->gradient, lmP->D);CHKERRQ(ierr);
77 
78       /* On a reset, the direction cannot be not a number; it is a
79          scaled gradient step.  No need to check for this condition. */
80 
81       lmP->bfgs = 1;
82       ++lmP->sgrad;
83       stepType = LMVM_SCALED_GRADIENT;
84     } else {
85       if (1 == lmP->bfgs) {
86         /*  The first BFGS direction is always the scaled gradient */
87         ++lmP->sgrad;
88         stepType = LMVM_SCALED_GRADIENT;
89       } else {
90         ++lmP->bfgs;
91         stepType = LMVM_BFGS;
92       }
93     }
94     ierr = VecScale(lmP->D, -1.0);CHKERRQ(ierr);
95 
96     /*  Perform the linesearch */
97     fold = f;
98     ierr = VecCopy(tao->solution, lmP->Xold);CHKERRQ(ierr);
99     ierr = VecCopy(tao->gradient, lmP->Gold);CHKERRQ(ierr);
100 
101     ierr = TaoLineSearchApply(tao->linesearch, tao->solution, &f, tao->gradient, lmP->D, &step,&ls_status);CHKERRQ(ierr);
102     ierr = TaoAddLineSearchCounts(tao);CHKERRQ(ierr);
103 
104     while (ls_status != TAOLINESEARCH_SUCCESS && ls_status != TAOLINESEARCH_SUCCESS_USER && (stepType != LMVM_GRADIENT)) {
105       /*  Linesearch failed */
106       /*  Reset factors and use scaled gradient step */
107       f = fold;
108       ierr = VecCopy(lmP->Xold, tao->solution);CHKERRQ(ierr);
109       ierr = VecCopy(lmP->Gold, tao->gradient);CHKERRQ(ierr);
110 
111       switch(stepType) {
112       case LMVM_BFGS:
113         /*  Failed to obtain acceptable iterate with BFGS step */
114         /*  Attempt to use the scaled gradient direction */
115 
116         if (f != 0.0) {
117           delta = 2.0 * PetscAbsScalar(f) / (gnorm*gnorm);
118         } else {
119           delta = 2.0 / (gnorm*gnorm);
120         }
121         ierr = MatLMVMSetDelta(lmP->M, delta);CHKERRQ(ierr);
122         ierr = MatLMVMReset(lmP->M);CHKERRQ(ierr);
123         ierr = MatLMVMUpdate(lmP->M, tao->solution, tao->gradient);CHKERRQ(ierr);
124         ierr = MatLMVMSolve(lmP->M, tao->gradient, lmP->D);CHKERRQ(ierr);
125 
126         /* On a reset, the direction cannot be not a number; it is a
127            scaled gradient step.  No need to check for this condition. */
128 
129         lmP->bfgs = 1;
130         ++lmP->sgrad;
131         stepType = LMVM_SCALED_GRADIENT;
132         break;
133 
134       case LMVM_SCALED_GRADIENT:
135         /* The scaled gradient step did not produce a new iterate;
136            attempt to use the gradient direction.
137            Need to make sure we are not using a different diagonal scaling */
138         ierr = MatLMVMSetDelta(lmP->M, 1.0);CHKERRQ(ierr);
139         ierr = MatLMVMReset(lmP->M);CHKERRQ(ierr);
140         ierr = MatLMVMUpdate(lmP->M, tao->solution, tao->gradient);CHKERRQ(ierr);
141         ierr = MatLMVMSolve(lmP->M, tao->gradient, lmP->D);CHKERRQ(ierr);
142 
143         lmP->bfgs = 1;
144         ++lmP->grad;
145         stepType = LMVM_GRADIENT;
146         break;
147       }
148       ierr = VecScale(lmP->D, -1.0);CHKERRQ(ierr);
149 
150       /*  Perform the linesearch */
151       ierr = TaoLineSearchApply(tao->linesearch, tao->solution, &f, tao->gradient, lmP->D, &step, &ls_status);CHKERRQ(ierr);
152       ierr = TaoAddLineSearchCounts(tao);CHKERRQ(ierr);
153     }
154 
155     if (ls_status != TAOLINESEARCH_SUCCESS && ls_status != TAOLINESEARCH_SUCCESS_USER) {
156       /*  Failed to find an improving point */
157       f = fold;
158       ierr = VecCopy(lmP->Xold, tao->solution);CHKERRQ(ierr);
159       ierr = VecCopy(lmP->Gold, tao->gradient);CHKERRQ(ierr);
160       step = 0.0;
161       reason = TAO_DIVERGED_LS_FAILURE;
162       tao->reason = TAO_DIVERGED_LS_FAILURE;
163     }
164 
165     /*  Check for termination */
166     ierr = TaoGradientNorm(tao, tao->gradient,NORM_2,&gnorm);CHKERRQ(ierr);
167 
168     tao->niter++;
169     ierr = TaoMonitor(tao,tao->niter,f,gnorm,0.0,step,&reason);CHKERRQ(ierr);
170   }
171   PetscFunctionReturn(0);
172 }
173 
174 static PetscErrorCode TaoSetUp_LMVM(Tao tao)
175 {
176   TAO_LMVM       *lmP = (TAO_LMVM *)tao->data;
177   PetscInt       n,N;
178   PetscErrorCode ierr;
179   KSP            H0ksp;
180 
181   PetscFunctionBegin;
182   /* Existence of tao->solution checked in TaoSetUp() */
183   if (!tao->gradient) {ierr = VecDuplicate(tao->solution,&tao->gradient);CHKERRQ(ierr);  }
184   if (!tao->stepdirection) {ierr = VecDuplicate(tao->solution,&tao->stepdirection);CHKERRQ(ierr);  }
185   if (!lmP->D) {ierr = VecDuplicate(tao->solution,&lmP->D);CHKERRQ(ierr);  }
186   if (!lmP->Xold) {ierr = VecDuplicate(tao->solution,&lmP->Xold);CHKERRQ(ierr);  }
187   if (!lmP->Gold) {ierr = VecDuplicate(tao->solution,&lmP->Gold);CHKERRQ(ierr);  }
188 
189   /*  Create matrix for the limited memory approximation */
190   ierr = VecGetLocalSize(tao->solution,&n);CHKERRQ(ierr);
191   ierr = VecGetSize(tao->solution,&N);CHKERRQ(ierr);
192   ierr = MatCreateLMVM(((PetscObject)tao)->comm,n,N,&lmP->M);CHKERRQ(ierr);
193   ierr = MatLMVMAllocateVectors(lmP->M,tao->solution);CHKERRQ(ierr);
194 
195   /* If the user has set a matrix to solve as the initial H0, set the options prefix here, and set up the KSP */
196   if (lmP->H0) {
197     const char *prefix;
198     PC H0pc;
199 
200     ierr = MatLMVMSetH0(lmP->M, lmP->H0);CHKERRQ(ierr);
201     ierr = MatLMVMGetH0KSP(lmP->M, &H0ksp);CHKERRQ(ierr);
202 
203     ierr = TaoGetOptionsPrefix(tao, &prefix);CHKERRQ(ierr);
204     ierr = KSPSetOptionsPrefix(H0ksp, prefix);CHKERRQ(ierr);
205     ierr = PetscObjectAppendOptionsPrefix((PetscObject)H0ksp, "tao_h0_");CHKERRQ(ierr);
206     ierr = KSPGetPC(H0ksp, &H0pc);CHKERRQ(ierr);
207     ierr = PetscObjectAppendOptionsPrefix((PetscObject)H0pc,  "tao_h0_");CHKERRQ(ierr);
208 
209     ierr = KSPSetFromOptions(H0ksp);CHKERRQ(ierr);
210     ierr = KSPSetUp(H0ksp);CHKERRQ(ierr);
211   }
212 
213   PetscFunctionReturn(0);
214 }
215 
216 /* ---------------------------------------------------------- */
217 static PetscErrorCode TaoDestroy_LMVM(Tao tao)
218 {
219   TAO_LMVM       *lmP = (TAO_LMVM *)tao->data;
220   PetscErrorCode ierr;
221 
222   PetscFunctionBegin;
223   if (tao->setupcalled) {
224     ierr = VecDestroy(&lmP->Xold);CHKERRQ(ierr);
225     ierr = VecDestroy(&lmP->Gold);CHKERRQ(ierr);
226     ierr = VecDestroy(&lmP->D);CHKERRQ(ierr);
227     ierr = MatDestroy(&lmP->M);CHKERRQ(ierr);
228   }
229 
230   if (lmP->H0) {
231     ierr = PetscObjectDereference((PetscObject)lmP->H0);CHKERRQ(ierr);
232   }
233 
234   ierr = PetscFree(tao->data);CHKERRQ(ierr);
235 
236   PetscFunctionReturn(0);
237 }
238 
239 /*------------------------------------------------------------*/
240 static PetscErrorCode TaoSetFromOptions_LMVM(PetscOptionItems *PetscOptionsObject,Tao tao)
241 {
242   PetscErrorCode ierr;
243 
244   PetscFunctionBegin;
245   ierr = PetscOptionsHead(PetscOptionsObject,"Limited-memory variable-metric method for unconstrained optimization");CHKERRQ(ierr);
246   ierr = TaoLineSearchSetFromOptions(tao->linesearch);CHKERRQ(ierr);
247   ierr = PetscOptionsTail();CHKERRQ(ierr);
248   PetscFunctionReturn(0);
249 }
250 
251 /*------------------------------------------------------------*/
252 static PetscErrorCode TaoView_LMVM(Tao tao, PetscViewer viewer)
253 {
254   TAO_LMVM       *lm = (TAO_LMVM *)tao->data;
255   PetscBool      isascii;
256   PetscErrorCode ierr;
257 
258   PetscFunctionBegin;
259   ierr = PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii);CHKERRQ(ierr);
260   if (isascii) {
261     ierr = PetscViewerASCIIPushTab(viewer);CHKERRQ(ierr);
262     ierr = PetscViewerASCIIPrintf(viewer, "BFGS steps: %D\n", lm->bfgs);CHKERRQ(ierr);
263     ierr = PetscViewerASCIIPrintf(viewer, "Scaled gradient steps: %D\n", lm->sgrad);CHKERRQ(ierr);
264     ierr = PetscViewerASCIIPrintf(viewer, "Gradient steps: %D\n", lm->grad);CHKERRQ(ierr);
265     ierr = PetscViewerASCIIPopTab(viewer);CHKERRQ(ierr);
266   }
267   PetscFunctionReturn(0);
268 }
269 
270 /* ---------------------------------------------------------- */
271 
272 /*MC
273      TAOLMVM - Limited Memory Variable Metric method is a quasi-Newton
274      optimization solver for unconstrained minimization. It solves
275      the Newton step
276               Hkdk = - gk
277 
278      using an approximation Bk in place of Hk, where Bk is composed using
279      the BFGS update formula. A More-Thuente line search is then used
280      to computed the steplength in the dk direction
281   Options Database Keys:
282 +     -tao_lmm_vectors - number of vectors to use for approximation
283 .     -tao_lmm_scale_type - "none","scalar","broyden"
284 .     -tao_lmm_limit_type - "none","average","relative","absolute"
285 .     -tao_lmm_rescale_type - "none","scalar","gl"
286 .     -tao_lmm_limit_mu - mu limiting factor
287 .     -tao_lmm_limit_nu - nu limiting factor
288 .     -tao_lmm_delta_min - minimum delta value
289 .     -tao_lmm_delta_max - maximum delta value
290 .     -tao_lmm_broyden_phi - phi factor for Broyden scaling
291 .     -tao_lmm_scalar_alpha - alpha factor for scalar scaling
292 .     -tao_lmm_rescale_alpha - alpha factor for rescaling diagonal
293 .     -tao_lmm_rescale_beta - beta factor for rescaling diagonal
294 .     -tao_lmm_scalar_history - amount of history for scalar scaling
295 .     -tao_lmm_rescale_history - amount of history for rescaling diagonal
296 -     -tao_lmm_eps - rejection tolerance
297 
298   Level: beginner
299 M*/
300 
301 PETSC_EXTERN PetscErrorCode TaoCreate_LMVM(Tao tao)
302 {
303   TAO_LMVM       *lmP;
304   const char     *morethuente_type = TAOLINESEARCHMT;
305   PetscErrorCode ierr;
306 
307   PetscFunctionBegin;
308   tao->ops->setup = TaoSetUp_LMVM;
309   tao->ops->solve = TaoSolve_LMVM;
310   tao->ops->view = TaoView_LMVM;
311   tao->ops->setfromoptions = TaoSetFromOptions_LMVM;
312   tao->ops->destroy = TaoDestroy_LMVM;
313 
314   ierr = PetscNewLog(tao,&lmP);CHKERRQ(ierr);
315   lmP->D = 0;
316   lmP->M = 0;
317   lmP->Xold = 0;
318   lmP->Gold = 0;
319   lmP->H0   = NULL;
320 
321   tao->data = (void*)lmP;
322   /* Override default settings (unless already changed) */
323   if (!tao->max_it_changed) tao->max_it = 2000;
324   if (!tao->max_funcs_changed) tao->max_funcs = 4000;
325 
326   ierr = TaoLineSearchCreate(((PetscObject)tao)->comm,&tao->linesearch);CHKERRQ(ierr);
327   ierr = TaoLineSearchSetType(tao->linesearch,morethuente_type);CHKERRQ(ierr);
328   ierr = TaoLineSearchUseTaoRoutines(tao->linesearch,tao);CHKERRQ(ierr);
329   ierr = TaoLineSearchSetOptionsPrefix(tao->linesearch,tao->hdr.prefix);CHKERRQ(ierr);
330   PetscFunctionReturn(0);
331 }
332