1 #include <petsctaolinesearch.h> 2 #include <../src/tao/bound/impls/bnk/bnk.h> 3 4 #include <petscksp.h> 5 6 /* Routine for BFGS preconditioner */ 7 8 PetscErrorCode MatLMVMSolveShell(PC pc, Vec b, Vec x) 9 { 10 PetscErrorCode ierr; 11 Mat M; 12 13 PetscFunctionBegin; 14 PetscValidHeaderSpecific(pc,PC_CLASSID,1); 15 PetscValidHeaderSpecific(b,VEC_CLASSID,2); 16 PetscValidHeaderSpecific(x,VEC_CLASSID,3); 17 ierr = PCShellGetContext(pc,(void**)&M);CHKERRQ(ierr); 18 ierr = MatLMVMSolveInactive(M, b, x);CHKERRQ(ierr); 19 PetscFunctionReturn(0); 20 } 21 22 PetscErrorCode TaoBNKInitialize(Tao tao) 23 { 24 PetscErrorCode ierr; 25 TAO_BNK *bnk = (TAO_BNK *)tao->data; 26 KSPType ksp_type; 27 PC pc; 28 29 PetscReal fmin, ftrial, prered, actred, kappa, sigma; 30 PetscReal tau, tau_1, tau_2, tau_max, tau_min, max_radius; 31 PetscReal delta, step = 1.0; 32 33 PetscInt n,N,needH = 1; 34 35 PetscInt i_max = 5; 36 PetscInt j_max = 1; 37 PetscInt i, j; 38 39 PetscFunctionBegin; 40 /* Number of times ksp stopped because of these reasons */ 41 bnk->ksp_atol = 0; 42 bnk->ksp_rtol = 0; 43 bnk->ksp_dtol = 0; 44 bnk->ksp_ctol = 0; 45 bnk->ksp_negc = 0; 46 bnk->ksp_iter = 0; 47 bnk->ksp_othr = 0; 48 49 /* Initialize trust-region radius when using nash, stcg, or gltr 50 Command automatically ignored for other methods 51 Will be reset during the first iteration 52 */ 53 ierr = KSPGetType(tao->ksp,&ksp_type);CHKERRQ(ierr); 54 ierr = PetscStrcmp(ksp_type,KSPCGNASH,&bnk->is_nash);CHKERRQ(ierr); 55 ierr = PetscStrcmp(ksp_type,KSPCGSTCG,&bnk->is_stcg);CHKERRQ(ierr); 56 ierr = PetscStrcmp(ksp_type,KSPCGGLTR,&bnk->is_gltr);CHKERRQ(ierr); 57 58 ierr = KSPCGSetRadius(tao->ksp,bnk->max_radius);CHKERRQ(ierr); 59 60 if (bnk->is_nash || bnk->is_stcg || bnk->is_gltr) { 61 if (tao->trust0 < 0.0) SETERRQ(PETSC_COMM_SELF,1,"Initial radius negative"); 62 tao->trust = tao->trust0; 63 tao->trust = PetscMax(tao->trust, bnk->min_radius); 64 tao->trust = PetscMin(tao->trust, bnk->max_radius); 65 } 66 67 /* Get vectors we will need */ 68 if (BNK_PC_BFGS == bnk->pc_type && !bnk->M) { 69 ierr = VecGetLocalSize(tao->solution,&n);CHKERRQ(ierr); 70 ierr = VecGetSize(tao->solution,&N);CHKERRQ(ierr); 71 ierr = MatCreateLMVM(((PetscObject)tao)->comm,n,N,&bnk->M);CHKERRQ(ierr); 72 ierr = MatLMVMAllocateVectors(bnk->M,tao->solution);CHKERRQ(ierr); 73 } 74 75 /* create vectors for the limited memory preconditioner */ 76 if ((BNK_PC_BFGS == bnk->pc_type) && (BFGS_SCALE_BFGS != bnk->bfgs_scale_type)) { 77 if (!bnk->Diag) { 78 ierr = VecDuplicate(tao->solution,&bnk->Diag);CHKERRQ(ierr); 79 } 80 } 81 82 /* Modify the preconditioner to use the bfgs approximation */ 83 ierr = KSPGetPC(tao->ksp, &pc);CHKERRQ(ierr); 84 switch(bnk->pc_type) { 85 case BNK_PC_NONE: 86 ierr = PCSetType(pc, PCNONE);CHKERRQ(ierr); 87 ierr = PCSetFromOptions(pc);CHKERRQ(ierr); 88 break; 89 90 case BNK_PC_AHESS: 91 ierr = PCSetType(pc, PCJACOBI);CHKERRQ(ierr); 92 ierr = PCSetFromOptions(pc);CHKERRQ(ierr); 93 ierr = PCJacobiSetUseAbs(pc,PETSC_TRUE);CHKERRQ(ierr); 94 break; 95 96 case BNK_PC_BFGS: 97 ierr = PCSetType(pc, PCSHELL);CHKERRQ(ierr); 98 ierr = PCSetFromOptions(pc);CHKERRQ(ierr); 99 ierr = PCShellSetName(pc, "bfgs");CHKERRQ(ierr); 100 ierr = PCShellSetContext(pc, bnk->M);CHKERRQ(ierr); 101 ierr = PCShellSetApply(pc, MatLMVMSolveShell);CHKERRQ(ierr); 102 break; 103 104 default: 105 /* Use the pc method set by pc_type */ 106 break; 107 } 108 109 /* Initialize trust-region radius. The initialization is only performed 110 when we are using Nash, Steihaug-Toint or the Generalized Lanczos method. */ 111 if (bnk->is_nash || bnk->is_stcg || bnk->is_gltr) { 112 switch(bnk->init_type) { 113 case BNK_INIT_CONSTANT: 114 /* Use the initial radius specified */ 115 break; 116 117 case BNK_INIT_INTERPOLATION: 118 /* Use the initial radius specified */ 119 max_radius = 0.0; 120 121 for (j = 0; j < j_max; ++j) { 122 fmin = bnk->f; 123 sigma = 0.0; 124 125 if (needH) { 126 ierr = TaoComputeHessian(tao, tao->solution,tao->hessian,tao->hessian_pre);CHKERRQ(ierr); 127 needH = 0; 128 } 129 130 for (i = 0; i < i_max; ++i) { 131 ierr = VecCopy(tao->solution,bnk->W);CHKERRQ(ierr); 132 ierr = VecAXPY(bnk->W,-tao->trust/bnk->gnorm,tao->gradient);CHKERRQ(ierr); 133 ierr = TaoComputeObjective(tao, bnk->W, &ftrial);CHKERRQ(ierr); 134 if (PetscIsInfOrNanReal(ftrial)) { 135 tau = bnk->gamma1_i; 136 } else { 137 if (ftrial < fmin) { 138 fmin = ftrial; 139 sigma = -tao->trust / bnk->gnorm; 140 } 141 142 ierr = MatMult(tao->hessian, tao->gradient, tao->stepdirection);CHKERRQ(ierr); 143 ierr = VecDot(tao->gradient, tao->stepdirection, &prered);CHKERRQ(ierr); 144 145 prered = tao->trust * (bnk->gnorm - 0.5 * tao->trust * prered / (bnk->gnorm * bnk->gnorm)); 146 actred = bnk->f - ftrial; 147 if ((PetscAbsScalar(actred) <= bnk->epsilon) && (PetscAbsScalar(prered) <= bnk->epsilon)) { 148 kappa = 1.0; 149 } else { 150 kappa = actred / prered; 151 } 152 153 tau_1 = bnk->theta_i * bnk->gnorm * tao->trust / (bnk->theta_i * bnk->gnorm * tao->trust + (1.0 - bnk->theta_i) * prered - actred); 154 tau_2 = bnk->theta_i * bnk->gnorm * tao->trust / (bnk->theta_i * bnk->gnorm * tao->trust - (1.0 + bnk->theta_i) * prered + actred); 155 tau_min = PetscMin(tau_1, tau_2); 156 tau_max = PetscMax(tau_1, tau_2); 157 158 if (PetscAbsScalar(kappa - 1.0) <= bnk->mu1_i) { 159 /* Great agreement */ 160 max_radius = PetscMax(max_radius, tao->trust); 161 162 if (tau_max < 1.0) { 163 tau = bnk->gamma3_i; 164 } else if (tau_max > bnk->gamma4_i) { 165 tau = bnk->gamma4_i; 166 } else if (tau_1 >= 1.0 && tau_1 <= bnk->gamma4_i && tau_2 < 1.0) { 167 tau = tau_1; 168 } else if (tau_2 >= 1.0 && tau_2 <= bnk->gamma4_i && tau_1 < 1.0) { 169 tau = tau_2; 170 } else { 171 tau = tau_max; 172 } 173 } else if (PetscAbsScalar(kappa - 1.0) <= bnk->mu2_i) { 174 /* Good agreement */ 175 max_radius = PetscMax(max_radius, tao->trust); 176 177 if (tau_max < bnk->gamma2_i) { 178 tau = bnk->gamma2_i; 179 } else if (tau_max > bnk->gamma3_i) { 180 tau = bnk->gamma3_i; 181 } else { 182 tau = tau_max; 183 } 184 } else { 185 /* Not good agreement */ 186 if (tau_min > 1.0) { 187 tau = bnk->gamma2_i; 188 } else if (tau_max < bnk->gamma1_i) { 189 tau = bnk->gamma1_i; 190 } else if ((tau_min < bnk->gamma1_i) && (tau_max >= 1.0)) { 191 tau = bnk->gamma1_i; 192 } else if ((tau_1 >= bnk->gamma1_i) && (tau_1 < 1.0) && ((tau_2 < bnk->gamma1_i) || (tau_2 >= 1.0))) { 193 tau = tau_1; 194 } else if ((tau_2 >= bnk->gamma1_i) && (tau_2 < 1.0) && ((tau_1 < bnk->gamma1_i) || (tau_2 >= 1.0))) { 195 tau = tau_2; 196 } else { 197 tau = tau_max; 198 } 199 } 200 } 201 tao->trust = tau * tao->trust; 202 } 203 204 if (fmin < bnk->f) { 205 bnk->f = fmin; 206 ierr = VecAXPY(tao->solution,sigma,tao->gradient);CHKERRQ(ierr); 207 ierr = TaoComputeGradient(tao,tao->solution,bnk->unprojected_gradient);CHKERRQ(ierr); 208 ierr = VecBoundGradientProjection(bnk->unprojected_gradient,tao->solution,tao->XL,tao->XU,tao->gradient);CHKERRQ(ierr); 209 210 ierr = TaoGradientNorm(tao, tao->gradient,NORM_2,&bnk->gnorm);CHKERRQ(ierr); 211 if (PetscIsInfOrNanReal(bnk->gnorm)) SETERRQ(PETSC_COMM_SELF,1, "User provided compute gradient generated Inf or NaN"); 212 needH = 1; 213 214 ierr = TaoLogConvergenceHistory(tao,bnk->f,bnk->gnorm,0.0,tao->ksp_its);CHKERRQ(ierr); 215 ierr = TaoMonitor(tao,tao->niter,bnk->f,bnk->gnorm,0.0,step);CHKERRQ(ierr); 216 ierr = (*tao->ops->convergencetest)(tao,tao->cnvP);CHKERRQ(ierr); 217 if (tao->reason != TAO_CONTINUE_ITERATING) PetscFunctionReturn(0); 218 } 219 } 220 tao->trust = PetscMax(tao->trust, max_radius); 221 222 /* Modify the radius if it is too large or small */ 223 tao->trust = PetscMax(tao->trust, bnk->min_radius); 224 tao->trust = PetscMin(tao->trust, bnk->max_radius); 225 break; 226 227 default: 228 /* Norm of the first direction will initialize radius */ 229 tao->trust = 0.0; 230 break; 231 } 232 } 233 234 /* Set initial scaling for the BFGS preconditioner 235 This step is done after computing the initial trust-region radius 236 since the function value may have decreased */ 237 if (BNK_PC_BFGS == bnk->pc_type) { 238 if (bnk->f != 0.0) { 239 delta = 2.0 * PetscAbsScalar(bnk->f) / (bnk->gnorm*bnk->gnorm); 240 } else { 241 delta = 2.0 / (bnk->gnorm*bnk->gnorm); 242 } 243 ierr = MatLMVMSetDelta(bnk->M,delta);CHKERRQ(ierr); 244 } 245 246 /* Set counter for gradient/reset steps*/ 247 bnk->newt = 0; 248 bnk->bfgs = 0; 249 bnk->sgrad = 0; 250 bnk->grad = 0; 251 PetscFunctionReturn(0); 252 } 253 254 PetscErrorCode TaoBNKComputeStep(Tao tao, PetscInt *stepType) 255 { 256 PetscErrorCode ierr; 257 TAO_BNK *bnk = (TAO_BNK *)tao->data; 258 KSPConvergedReason ksp_reason; 259 260 PetscReal gdx, delta, e_min; 261 262 PetscInt bfgsUpdates = 0; 263 PetscInt kspits; 264 265 PetscFunctionBegin; 266 if ((BNK_PC_BFGS == bnk->pc_type) && (BFGS_SCALE_AHESS == bnk->bfgs_scale_type)) { 267 /* Obtain diagonal for the bfgs preconditioner */ 268 ierr = MatGetDiagonal(tao->hessian, bnk->Diag);CHKERRQ(ierr); 269 ierr = VecAbs(bnk->Diag);CHKERRQ(ierr); 270 ierr = VecReciprocal(bnk->Diag);CHKERRQ(ierr); 271 ierr = MatLMVMSetScale(bnk->M,bnk->Diag);CHKERRQ(ierr); 272 } 273 274 /* Shift the Hessian matrix */ 275 bnk->pert = bnk->sval; 276 if (bnk->pert > 0) { 277 ierr = MatShift(tao->hessian, bnk->pert);CHKERRQ(ierr); 278 if (tao->hessian != tao->hessian_pre) { 279 ierr = MatShift(tao->hessian_pre, bnk->pert);CHKERRQ(ierr); 280 } 281 } 282 283 if (BNK_PC_BFGS == bnk->pc_type) { 284 if (BFGS_SCALE_PHESS == bnk->bfgs_scale_type) { 285 /* Obtain diagonal for the bfgs preconditioner */ 286 ierr = MatGetDiagonal(tao->hessian, bnk->Diag);CHKERRQ(ierr); 287 ierr = VecAbs(bnk->Diag);CHKERRQ(ierr); 288 ierr = VecReciprocal(bnk->Diag);CHKERRQ(ierr); 289 ierr = MatLMVMSetScale(bnk->M,bnk->Diag);CHKERRQ(ierr); 290 } 291 /* Update the limited memory preconditioner and get existing # of updates */ 292 ierr = MatLMVMUpdate(bnk->M, tao->solution, bnk->unprojected_gradient);CHKERRQ(ierr); 293 ierr = MatLMVMGetUpdates(bnk->M, &bfgsUpdates);CHKERRQ(ierr); 294 } 295 296 /* Determine the inactive set */ 297 ierr = ISDestroy(&bnk->inactive_idx);CHKERRQ(ierr); 298 ierr = VecWhichInactive(tao->XL,tao->solution,bnk->unprojected_gradient,tao->XU,PETSC_TRUE,&bnk->inactive_idx);CHKERRQ(ierr); 299 300 /* Prepare masked matrices for the inactive set */ 301 ierr = MatLMVMSetInactive(bnk->M, bnk->inactive_idx);CHKERRQ(ierr); 302 ierr = VecSet(tao->stepdirection, 0.0);CHKERRQ(ierr); 303 ierr = TaoMatGetSubMat(tao->hessian, bnk->inactive_idx, bnk->Xwork, TAO_SUBSET_MASK, &bnk->H_inactive);CHKERRQ(ierr); 304 if (tao->hessian == tao->hessian_pre) { 305 bnk->Hpre_inactive = bnk->H_inactive; 306 } else { 307 ierr = TaoMatGetSubMat(tao->hessian_pre, bnk->inactive_idx, bnk->Xwork, TAO_SUBSET_MASK, &bnk->Hpre_inactive);CHKERRQ(ierr); 308 } 309 310 /* Solve the Newton system of equations */ 311 ierr = KSPSetOperators(tao->ksp,bnk->H_inactive,bnk->Hpre_inactive);CHKERRQ(ierr); 312 if (bnk->is_nash || bnk->is_stcg || bnk->is_gltr) { 313 ierr = KSPCGSetRadius(tao->ksp,bnk->max_radius);CHKERRQ(ierr); 314 ierr = KSPSolve(tao->ksp, tao->gradient, tao->stepdirection);CHKERRQ(ierr); 315 ierr = KSPGetIterationNumber(tao->ksp,&kspits);CHKERRQ(ierr); 316 tao->ksp_its+=kspits; 317 tao->ksp_tot_its+=kspits; 318 ierr = KSPCGGetNormD(tao->ksp,&bnk->dnorm);CHKERRQ(ierr); 319 320 if (0.0 == tao->trust) { 321 /* Radius was uninitialized; use the norm of the direction */ 322 if (bnk->dnorm > 0.0) { 323 tao->trust = bnk->dnorm; 324 325 /* Modify the radius if it is too large or small */ 326 tao->trust = PetscMax(tao->trust, bnk->min_radius); 327 tao->trust = PetscMin(tao->trust, bnk->max_radius); 328 } else { 329 /* The direction was bad; set radius to default value and re-solve 330 the trust-region subproblem to get a direction */ 331 tao->trust = tao->trust0; 332 333 /* Modify the radius if it is too large or small */ 334 tao->trust = PetscMax(tao->trust, bnk->min_radius); 335 tao->trust = PetscMin(tao->trust, bnk->max_radius); 336 337 ierr = KSPCGSetRadius(tao->ksp,bnk->max_radius);CHKERRQ(ierr); 338 ierr = KSPSolve(tao->ksp, tao->gradient, tao->stepdirection);CHKERRQ(ierr); 339 ierr = KSPGetIterationNumber(tao->ksp,&kspits);CHKERRQ(ierr); 340 tao->ksp_its+=kspits; 341 tao->ksp_tot_its+=kspits; 342 ierr = KSPCGGetNormD(tao->ksp,&bnk->dnorm);CHKERRQ(ierr); 343 344 if (bnk->dnorm == 0.0) SETERRQ(PETSC_COMM_SELF,1, "Initial direction zero"); 345 } 346 } 347 } else { 348 ierr = KSPSolve(tao->ksp, tao->gradient, tao->stepdirection);CHKERRQ(ierr); 349 ierr = KSPGetIterationNumber(tao->ksp, &kspits);CHKERRQ(ierr); 350 tao->ksp_its += kspits; 351 tao->ksp_tot_its+=kspits; 352 } 353 /* Destroy masked matrices */ 354 if (bnk->H_inactive != bnk->Hpre_inactive) { 355 ierr = MatDestroy(&bnk->Hpre_inactive);CHKERRQ(ierr); 356 } 357 ierr = MatDestroy(&bnk->H_inactive);CHKERRQ(ierr); 358 359 ierr = VecScale(tao->stepdirection, -1.0);CHKERRQ(ierr); 360 ierr = KSPGetConvergedReason(tao->ksp, &ksp_reason);CHKERRQ(ierr); 361 if ((KSP_DIVERGED_INDEFINITE_PC == ksp_reason) && (BNK_PC_BFGS == bnk->pc_type) && (bfgsUpdates > 1)) { 362 /* Preconditioner is numerically indefinite; reset the 363 approximate if using BFGS preconditioning. */ 364 365 if (bnk->f != 0.0) { 366 delta = 2.0 * PetscAbsScalar(bnk->f) / (bnk->gnorm*bnk->gnorm); 367 } else { 368 delta = 2.0 / (bnk->gnorm*bnk->gnorm); 369 } 370 ierr = MatLMVMSetDelta(bnk->M,delta);CHKERRQ(ierr); 371 ierr = MatLMVMReset(bnk->M);CHKERRQ(ierr); 372 ierr = MatLMVMUpdate(bnk->M, tao->solution, bnk->unprojected_gradient);CHKERRQ(ierr); 373 bfgsUpdates = 1; 374 } 375 376 if (KSP_CONVERGED_ATOL == ksp_reason) { 377 ++bnk->ksp_atol; 378 } else if (KSP_CONVERGED_RTOL == ksp_reason) { 379 ++bnk->ksp_rtol; 380 } else if (KSP_CONVERGED_CG_CONSTRAINED == ksp_reason) { 381 ++bnk->ksp_ctol; 382 } else if (KSP_CONVERGED_CG_NEG_CURVE == ksp_reason) { 383 ++bnk->ksp_negc; 384 } else if (KSP_DIVERGED_DTOL == ksp_reason) { 385 ++bnk->ksp_dtol; 386 } else if (KSP_DIVERGED_ITS == ksp_reason) { 387 ++bnk->ksp_iter; 388 } else { 389 ++bnk->ksp_othr; 390 } 391 392 /* Check for success (descent direction) */ 393 ierr = VecDot(tao->stepdirection, tao->gradient, &gdx);CHKERRQ(ierr); 394 if ((gdx >= 0.0) || PetscIsInfOrNanReal(gdx)) { 395 /* Newton step is not descent or direction produced Inf or NaN 396 Update the perturbation for next time */ 397 if (bnk->pert <= 0.0) { 398 /* Initialize the perturbation */ 399 bnk->pert = PetscMin(bnk->imax, PetscMax(bnk->imin, bnk->imfac * bnk->gnorm)); 400 if (bnk->is_gltr) { 401 ierr = KSPCGGLTRGetMinEig(tao->ksp,&e_min);CHKERRQ(ierr); 402 bnk->pert = PetscMax(bnk->pert, -e_min); 403 } 404 } else { 405 /* Increase the perturbation */ 406 bnk->pert = PetscMin(bnk->pmax, PetscMax(bnk->pgfac * bnk->pert, bnk->pmgfac * bnk->gnorm)); 407 } 408 409 if (BNK_PC_BFGS != bnk->pc_type) { 410 /* We don't have the bfgs matrix around and updated 411 Must use gradient direction in this case */ 412 ierr = VecCopy(tao->gradient, tao->stepdirection);CHKERRQ(ierr); 413 ierr = VecScale(tao->stepdirection, -1.0);CHKERRQ(ierr); 414 ++bnk->grad; 415 *stepType = BNK_GRADIENT; 416 } else { 417 /* Attempt to use the BFGS direction */ 418 ierr = MatLMVMSolve(bnk->M, bnk->unprojected_gradient, tao->stepdirection);CHKERRQ(ierr); 419 ierr = VecBoundGradientProjection(tao->stepdirection,tao->solution,tao->XL,tao->XU,tao->stepdirection);CHKERRQ(ierr); 420 ierr = VecScale(tao->stepdirection, -1.0);CHKERRQ(ierr); 421 422 /* Check for success (descent direction) */ 423 ierr = VecDot(tao->gradient, tao->stepdirection, &gdx);CHKERRQ(ierr); 424 if ((gdx >= 0) || PetscIsInfOrNanReal(gdx)) { 425 /* BFGS direction is not descent or direction produced not a number 426 We can assert bfgsUpdates > 1 in this case because 427 the first solve produces the scaled gradient direction, 428 which is guaranteed to be descent */ 429 430 /* Use steepest descent direction (scaled) */ 431 432 if (bnk->f != 0.0) { 433 delta = 2.0 * PetscAbsScalar(bnk->f) / (bnk->gnorm*bnk->gnorm); 434 } else { 435 delta = 2.0 / (bnk->gnorm*bnk->gnorm); 436 } 437 ierr = MatLMVMSetDelta(bnk->M, delta);CHKERRQ(ierr); 438 ierr = MatLMVMReset(bnk->M);CHKERRQ(ierr); 439 ierr = MatLMVMUpdate(bnk->M, tao->solution, bnk->unprojected_gradient);CHKERRQ(ierr); 440 ierr = MatLMVMSolve(bnk->M, bnk->unprojected_gradient, tao->stepdirection);CHKERRQ(ierr); 441 ierr = VecBoundGradientProjection(tao->stepdirection,tao->solution,tao->XL,tao->XU,tao->stepdirection);CHKERRQ(ierr); 442 ierr = VecScale(tao->stepdirection, -1.0);CHKERRQ(ierr); 443 444 bfgsUpdates = 1; 445 ++bnk->sgrad; 446 *stepType = BNK_SCALED_GRADIENT; 447 } else { 448 if (1 == bfgsUpdates) { 449 /* The first BFGS direction is always the scaled gradient */ 450 ++bnk->sgrad; 451 *stepType = BNK_SCALED_GRADIENT; 452 } else { 453 ++bnk->bfgs; 454 *stepType = BNK_BFGS; 455 } 456 } 457 } 458 } else { 459 /* Computed Newton step is descent */ 460 switch (ksp_reason) { 461 case KSP_DIVERGED_NANORINF: 462 case KSP_DIVERGED_BREAKDOWN: 463 case KSP_DIVERGED_INDEFINITE_MAT: 464 case KSP_DIVERGED_INDEFINITE_PC: 465 case KSP_CONVERGED_CG_NEG_CURVE: 466 /* Matrix or preconditioner is indefinite; increase perturbation */ 467 if (bnk->pert <= 0.0) { 468 /* Initialize the perturbation */ 469 bnk->pert = PetscMin(bnk->imax, PetscMax(bnk->imin, bnk->imfac * bnk->gnorm)); 470 if (bnk->is_gltr) { 471 ierr = KSPCGGLTRGetMinEig(tao->ksp, &e_min);CHKERRQ(ierr); 472 bnk->pert = PetscMax(bnk->pert, -e_min); 473 } 474 } else { 475 /* Increase the perturbation */ 476 bnk->pert = PetscMin(bnk->pmax, PetscMax(bnk->pgfac * bnk->pert, bnk->pmgfac * bnk->gnorm)); 477 } 478 break; 479 480 default: 481 /* Newton step computation is good; decrease perturbation */ 482 bnk->pert = PetscMin(bnk->psfac * bnk->pert, bnk->pmsfac * bnk->gnorm); 483 if (bnk->pert < bnk->pmin) { 484 bnk->pert = 0.0; 485 } 486 break; 487 } 488 489 ++bnk->newt; 490 stepType = BNK_NEWTON; 491 } 492 PetscFunctionReturn(0); 493 } 494 495 PetscErrorCode TaoBNKUpdateTrustRadius(Tao tao, PetscReal fold, PetscReal fnew, PetscInt stepType, PetscBool *accept) 496 { 497 TAO_BNK *bnk = (TAO_BNK *)tao->data; 498 PetscErrorCode ierr; 499 500 PetscReal step, prered, actred, kappa; 501 PetscReal gdx, tau_1, tau_2, tau_min, tau_max; 502 503 PetscFunctionBegin; 504 /* Update trust region radius */ 505 *accept = PETSC_FALSE; 506 switch(bnk->update_type) { 507 case BNK_UPDATE_STEP: 508 if (stepType == BNK_NEWTON) { 509 ierr = TaoLineSearchGetStepLength(tao->linesearch, &step);CHKERRQ(ierr); 510 if (step < bnk->nu1) { 511 /* Very bad step taken; reduce radius */ 512 tao->trust = bnk->omega1 * PetscMin(bnk->dnorm, tao->trust); 513 } else if (step < bnk->nu2) { 514 /* Reasonably bad step taken; reduce radius */ 515 tao->trust = bnk->omega2 * PetscMin(bnk->dnorm, tao->trust); 516 } else if (step < bnk->nu3) { 517 /* Reasonable step was taken; leave radius alone */ 518 if (bnk->omega3 < 1.0) { 519 tao->trust = bnk->omega3 * PetscMin(bnk->dnorm, tao->trust); 520 } else if (bnk->omega3 > 1.0) { 521 tao->trust = PetscMax(bnk->omega3 * bnk->dnorm, tao->trust); 522 } 523 } else if (step < bnk->nu4) { 524 /* Full step taken; increase the radius */ 525 tao->trust = PetscMax(bnk->omega4 * bnk->dnorm, tao->trust); 526 } else { 527 /* More than full step taken; increase the radius */ 528 tao->trust = PetscMax(bnk->omega5 * bnk->dnorm, tao->trust); 529 } 530 } else { 531 /* Newton step was not good; reduce the radius */ 532 tao->trust = bnk->omega1 * PetscMin(bnk->dnorm, tao->trust); 533 } 534 break; 535 536 case BNK_UPDATE_REDUCTION: 537 if (stepType == BNK_NEWTON) { 538 /* Get predicted reduction */ 539 ierr = KSPCGGetObjFcn(tao->ksp,&prered);CHKERRQ(ierr); 540 if (prered >= 0.0) { 541 /* The predicted reduction has the wrong sign. This cannot */ 542 /* happen in infinite precision arithmetic. Step should */ 543 /* be rejected! */ 544 tao->trust = bnk->alpha1 * PetscMin(tao->trust, bnk->dnorm); 545 } else { 546 if (PetscIsInfOrNanReal(fnew)) { 547 tao->trust = bnk->alpha1 * PetscMin(tao->trust, bnk->dnorm); 548 } else { 549 /* Compute and actual reduction */ 550 actred = fold - fnew; 551 prered = -prered; 552 if ((PetscAbsScalar(actred) <= bnk->epsilon) && 553 (PetscAbsScalar(prered) <= bnk->epsilon)) { 554 kappa = 1.0; 555 } else { 556 kappa = actred / prered; 557 } 558 /* Accept of reject the step and update radius */ 559 if (kappa < bnk->eta1) { 560 /* Very bad step, rejected */ 561 tao->trust = bnk->alpha1 * PetscMin(tao->trust, bnk->dnorm); 562 } else { 563 /* Accept step here */ 564 *accept = PETSC_TRUE; 565 /* Adjust trust radius */ 566 if (kappa < bnk->eta2) { 567 /* Marginal bad step */ 568 tao->trust = bnk->alpha2 * PetscMin(tao->trust, bnk->dnorm); 569 } else if (kappa < bnk->eta3) { 570 /* Reasonable step */ 571 if (bnk->alpha3 < 1.0) { 572 tao->trust = bnk->alpha3 * PetscMin(bnk->dnorm, tao->trust); 573 } else if (bnk->alpha3 > 1.0) { 574 tao->trust = PetscMax(bnk->alpha3 * bnk->dnorm, tao->trust); 575 } 576 } else if (kappa < bnk->eta4) { 577 /* Good step */ 578 tao->trust = PetscMax(bnk->alpha4 * bnk->dnorm, tao->trust); 579 } else { 580 /* Very good step */ 581 tao->trust = PetscMax(bnk->alpha5 * bnk->dnorm, tao->trust); 582 } 583 } 584 } 585 } 586 } else { 587 /* Newton step was not good; reduce the radius */ 588 tao->trust = bnk->alpha1 * PetscMin(bnk->dnorm, tao->trust); 589 } 590 break; 591 592 default: 593 if (stepType == BNK_NEWTON) { 594 ierr = KSPCGGetObjFcn(tao->ksp,&prered);CHKERRQ(ierr); 595 if (prered >= 0.0) { 596 /* The predicted reduction has the wrong sign. This cannot */ 597 /* happen in infinite precision arithmetic. Step should */ 598 /* be rejected! */ 599 tao->trust = bnk->gamma1 * PetscMin(tao->trust, bnk->dnorm); 600 } else { 601 if (PetscIsInfOrNanReal(fnew)) { 602 tao->trust = bnk->gamma1 * PetscMin(tao->trust, bnk->dnorm); 603 } else { 604 actred = fold - fnew; 605 prered = -prered; 606 if ((PetscAbsScalar(actred) <= bnk->epsilon) && (PetscAbsScalar(prered) <= bnk->epsilon)) { 607 kappa = 1.0; 608 } else { 609 kappa = actred / prered; 610 } 611 612 ierr = VecDot(tao->gradient, tao->stepdirection, &gdx);CHKERRQ(ierr); 613 tau_1 = bnk->theta * gdx / (bnk->theta * gdx - (1.0 - bnk->theta) * prered + actred); 614 tau_2 = bnk->theta * gdx / (bnk->theta * gdx + (1.0 + bnk->theta) * prered - actred); 615 tau_min = PetscMin(tau_1, tau_2); 616 tau_max = PetscMax(tau_1, tau_2); 617 618 if (kappa >= 1.0 - bnk->mu1) { 619 /* Great agreement */ 620 *accept = PETSC_TRUE; 621 if (tau_max < 1.0) { 622 tao->trust = PetscMax(tao->trust, bnk->gamma3 * bnk->dnorm); 623 } else if (tau_max > bnk->gamma4) { 624 tao->trust = PetscMax(tao->trust, bnk->gamma4 * bnk->dnorm); 625 } else { 626 tao->trust = PetscMax(tao->trust, tau_max * bnk->dnorm); 627 } 628 } else if (kappa >= 1.0 - bnk->mu2) { 629 /* Good agreement */ 630 *accept = PETSC_TRUE; 631 if (tau_max < bnk->gamma2) { 632 tao->trust = bnk->gamma2 * PetscMin(tao->trust, bnk->dnorm); 633 } else if (tau_max > bnk->gamma3) { 634 tao->trust = PetscMax(tao->trust, bnk->gamma3 * bnk->dnorm); 635 } else if (tau_max < 1.0) { 636 tao->trust = tau_max * PetscMin(tao->trust, bnk->dnorm); 637 } else { 638 tao->trust = PetscMax(tao->trust, tau_max * bnk->dnorm); 639 } 640 } else { 641 /* Not good agreement */ 642 if (tau_min > 1.0) { 643 tao->trust = bnk->gamma2 * PetscMin(tao->trust, bnk->dnorm); 644 } else if (tau_max < bnk->gamma1) { 645 tao->trust = bnk->gamma1 * PetscMin(tao->trust, bnk->dnorm); 646 } else if ((tau_min < bnk->gamma1) && (tau_max >= 1.0)) { 647 tao->trust = bnk->gamma1 * PetscMin(tao->trust, bnk->dnorm); 648 } else if ((tau_1 >= bnk->gamma1) && (tau_1 < 1.0) && ((tau_2 < bnk->gamma1) || (tau_2 >= 1.0))) { 649 tao->trust = tau_1 * PetscMin(tao->trust, bnk->dnorm); 650 } else if ((tau_2 >= bnk->gamma1) && (tau_2 < 1.0) && ((tau_1 < bnk->gamma1) || (tau_2 >= 1.0))) { 651 tao->trust = tau_2 * PetscMin(tao->trust, bnk->dnorm); 652 } else { 653 tao->trust = tau_max * PetscMin(tao->trust, bnk->dnorm); 654 } 655 } 656 } 657 } 658 } else { 659 /* Newton step was not good; reduce the radius */ 660 tao->trust = bnk->gamma1 * PetscMin(bnk->dnorm, tao->trust); 661 } 662 /* The radius may have been increased; modify if it is too large */ 663 tao->trust = PetscMin(tao->trust, bnk->max_radius); 664 } 665 PetscFunctionReturn(0); 666 } 667 668 /* ---------------------------------------------------------- */ 669 static PetscErrorCode TaoSetUp_BNK(Tao tao) 670 { 671 TAO_BNK *bnk = (TAO_BNK *)tao->data; 672 PetscErrorCode ierr; 673 674 PetscFunctionBegin; 675 if (!tao->gradient) {ierr = VecDuplicate(tao->solution,&tao->gradient);CHKERRQ(ierr);} 676 if (!tao->stepdirection) {ierr = VecDuplicate(tao->solution,&tao->stepdirection);CHKERRQ(ierr);} 677 if (!bnk->W) {ierr = VecDuplicate(tao->solution,&bnk->W);CHKERRQ(ierr);} 678 if (!bnk->Xold) {ierr = VecDuplicate(tao->solution,&bnk->Xold);CHKERRQ(ierr);} 679 if (!bnk->Gold) {ierr = VecDuplicate(tao->solution,&bnk->Gold);CHKERRQ(ierr);} 680 if (!bnk->Xwork) {ierr = VecDuplicate(tao->solution,&bnk->Xwork);CHKERRQ(ierr);} 681 if (!bnk->Gwork) {ierr = VecDuplicate(tao->solution,&bnk->Gwork);CHKERRQ(ierr);} 682 if (!bnk->unprojected_gradient) {ierr = VecDuplicate(tao->solution,&bnk->unprojected_gradient);CHKERRQ(ierr);} 683 if (!bnk->unprojected_gradient_old) {ierr = VecDuplicate(tao->solution,&bnk->unprojected_gradient_old);CHKERRQ(ierr);} 684 bnk->Diag = 0; 685 bnk->M = 0; 686 bnk->H_inactive = 0; 687 bnk->Hpre_inactive = 0; 688 PetscFunctionReturn(0); 689 } 690 691 /*------------------------------------------------------------*/ 692 static PetscErrorCode TaoDestroy_BNK(Tao tao) 693 { 694 TAO_BNK *bnk = (TAO_BNK *)tao->data; 695 PetscErrorCode ierr; 696 697 PetscFunctionBegin; 698 if (tao->setupcalled) { 699 ierr = VecDestroy(&bnk->W);CHKERRQ(ierr); 700 ierr = VecDestroy(&bnk->Xold);CHKERRQ(ierr); 701 ierr = VecDestroy(&bnk->Gold);CHKERRQ(ierr); 702 ierr = VecDestroy(&bnk->Xwork);CHKERRQ(ierr); 703 ierr = VecDestroy(&bnk->Gwork);CHKERRQ(ierr); 704 ierr = VecDestroy(&bnk->unprojected_gradient);CHKERRQ(ierr); 705 ierr = VecDestroy(&bnk->unprojected_gradient_old);CHKERRQ(ierr); 706 } 707 ierr = VecDestroy(&bnk->Diag);CHKERRQ(ierr); 708 ierr = MatDestroy(&bnk->M);CHKERRQ(ierr); 709 ierr = PetscFree(tao->data);CHKERRQ(ierr); 710 PetscFunctionReturn(0); 711 } 712 713 /*------------------------------------------------------------*/ 714 static PetscErrorCode TaoSetFromOptions_BNK(PetscOptionItems *PetscOptionsObject,Tao tao) 715 { 716 TAO_BNK *bnk = (TAO_BNK *)tao->data; 717 PetscErrorCode ierr; 718 719 PetscFunctionBegin; 720 ierr = PetscOptionsHead(PetscOptionsObject,"Newton line search method for unconstrained optimization");CHKERRQ(ierr); 721 ierr = PetscOptionsEList("-tao_BNK_pc_type", "pc type", "", BNK_PC, BNK_PC_TYPES, BNK_PC[bnk->pc_type], &bnk->pc_type, 0);CHKERRQ(ierr); 722 ierr = PetscOptionsEList("-tao_BNK_bfgs_scale_type", "bfgs scale type", "", BFGS_SCALE, BFGS_SCALE_TYPES, BFGS_SCALE[bnk->bfgs_scale_type], &bnk->bfgs_scale_type, 0);CHKERRQ(ierr); 723 ierr = PetscOptionsEList("-tao_BNK_init_type", "radius initialization type", "", BNK_INIT, BNK_INIT_TYPES, BNK_INIT[bnk->init_type], &bnk->init_type, 0);CHKERRQ(ierr); 724 ierr = PetscOptionsEList("-tao_BNK_update_type", "radius update type", "", BNK_UPDATE, BNK_UPDATE_TYPES, BNK_UPDATE[bnk->update_type], &bnk->update_type, 0);CHKERRQ(ierr); 725 ierr = PetscOptionsReal("-tao_BNK_sval", "perturbation starting value", "", bnk->sval, &bnk->sval,NULL);CHKERRQ(ierr); 726 ierr = PetscOptionsReal("-tao_BNK_imin", "minimum initial perturbation", "", bnk->imin, &bnk->imin,NULL);CHKERRQ(ierr); 727 ierr = PetscOptionsReal("-tao_BNK_imax", "maximum initial perturbation", "", bnk->imax, &bnk->imax,NULL);CHKERRQ(ierr); 728 ierr = PetscOptionsReal("-tao_BNK_imfac", "initial merit factor", "", bnk->imfac, &bnk->imfac,NULL);CHKERRQ(ierr); 729 ierr = PetscOptionsReal("-tao_BNK_pmin", "minimum perturbation", "", bnk->pmin, &bnk->pmin,NULL);CHKERRQ(ierr); 730 ierr = PetscOptionsReal("-tao_BNK_pmax", "maximum perturbation", "", bnk->pmax, &bnk->pmax,NULL);CHKERRQ(ierr); 731 ierr = PetscOptionsReal("-tao_BNK_pgfac", "growth factor", "", bnk->pgfac, &bnk->pgfac,NULL);CHKERRQ(ierr); 732 ierr = PetscOptionsReal("-tao_BNK_psfac", "shrink factor", "", bnk->psfac, &bnk->psfac,NULL);CHKERRQ(ierr); 733 ierr = PetscOptionsReal("-tao_BNK_pmgfac", "merit growth factor", "", bnk->pmgfac, &bnk->pmgfac,NULL);CHKERRQ(ierr); 734 ierr = PetscOptionsReal("-tao_BNK_pmsfac", "merit shrink factor", "", bnk->pmsfac, &bnk->pmsfac,NULL);CHKERRQ(ierr); 735 ierr = PetscOptionsReal("-tao_BNK_eta1", "poor steplength; reduce radius", "", bnk->eta1, &bnk->eta1,NULL);CHKERRQ(ierr); 736 ierr = PetscOptionsReal("-tao_BNK_eta2", "reasonable steplength; leave radius alone", "", bnk->eta2, &bnk->eta2,NULL);CHKERRQ(ierr); 737 ierr = PetscOptionsReal("-tao_BNK_eta3", "good steplength; increase radius", "", bnk->eta3, &bnk->eta3,NULL);CHKERRQ(ierr); 738 ierr = PetscOptionsReal("-tao_BNK_eta4", "excellent steplength; greatly increase radius", "", bnk->eta4, &bnk->eta4,NULL);CHKERRQ(ierr); 739 ierr = PetscOptionsReal("-tao_BNK_alpha1", "", "", bnk->alpha1, &bnk->alpha1,NULL);CHKERRQ(ierr); 740 ierr = PetscOptionsReal("-tao_BNK_alpha2", "", "", bnk->alpha2, &bnk->alpha2,NULL);CHKERRQ(ierr); 741 ierr = PetscOptionsReal("-tao_BNK_alpha3", "", "", bnk->alpha3, &bnk->alpha3,NULL);CHKERRQ(ierr); 742 ierr = PetscOptionsReal("-tao_BNK_alpha4", "", "", bnk->alpha4, &bnk->alpha4,NULL);CHKERRQ(ierr); 743 ierr = PetscOptionsReal("-tao_BNK_alpha5", "", "", bnk->alpha5, &bnk->alpha5,NULL);CHKERRQ(ierr); 744 ierr = PetscOptionsReal("-tao_BNK_nu1", "poor steplength; reduce radius", "", bnk->nu1, &bnk->nu1,NULL);CHKERRQ(ierr); 745 ierr = PetscOptionsReal("-tao_BNK_nu2", "reasonable steplength; leave radius alone", "", bnk->nu2, &bnk->nu2,NULL);CHKERRQ(ierr); 746 ierr = PetscOptionsReal("-tao_BNK_nu3", "good steplength; increase radius", "", bnk->nu3, &bnk->nu3,NULL);CHKERRQ(ierr); 747 ierr = PetscOptionsReal("-tao_BNK_nu4", "excellent steplength; greatly increase radius", "", bnk->nu4, &bnk->nu4,NULL);CHKERRQ(ierr); 748 ierr = PetscOptionsReal("-tao_BNK_omega1", "", "", bnk->omega1, &bnk->omega1,NULL);CHKERRQ(ierr); 749 ierr = PetscOptionsReal("-tao_BNK_omega2", "", "", bnk->omega2, &bnk->omega2,NULL);CHKERRQ(ierr); 750 ierr = PetscOptionsReal("-tao_BNK_omega3", "", "", bnk->omega3, &bnk->omega3,NULL);CHKERRQ(ierr); 751 ierr = PetscOptionsReal("-tao_BNK_omega4", "", "", bnk->omega4, &bnk->omega4,NULL);CHKERRQ(ierr); 752 ierr = PetscOptionsReal("-tao_BNK_omega5", "", "", bnk->omega5, &bnk->omega5,NULL);CHKERRQ(ierr); 753 ierr = PetscOptionsReal("-tao_BNK_mu1_i", "", "", bnk->mu1_i, &bnk->mu1_i,NULL);CHKERRQ(ierr); 754 ierr = PetscOptionsReal("-tao_BNK_mu2_i", "", "", bnk->mu2_i, &bnk->mu2_i,NULL);CHKERRQ(ierr); 755 ierr = PetscOptionsReal("-tao_BNK_gamma1_i", "", "", bnk->gamma1_i, &bnk->gamma1_i,NULL);CHKERRQ(ierr); 756 ierr = PetscOptionsReal("-tao_BNK_gamma2_i", "", "", bnk->gamma2_i, &bnk->gamma2_i,NULL);CHKERRQ(ierr); 757 ierr = PetscOptionsReal("-tao_BNK_gamma3_i", "", "", bnk->gamma3_i, &bnk->gamma3_i,NULL);CHKERRQ(ierr); 758 ierr = PetscOptionsReal("-tao_BNK_gamma4_i", "", "", bnk->gamma4_i, &bnk->gamma4_i,NULL);CHKERRQ(ierr); 759 ierr = PetscOptionsReal("-tao_BNK_theta_i", "", "", bnk->theta_i, &bnk->theta_i,NULL);CHKERRQ(ierr); 760 ierr = PetscOptionsReal("-tao_BNK_mu1", "", "", bnk->mu1, &bnk->mu1,NULL);CHKERRQ(ierr); 761 ierr = PetscOptionsReal("-tao_BNK_mu2", "", "", bnk->mu2, &bnk->mu2,NULL);CHKERRQ(ierr); 762 ierr = PetscOptionsReal("-tao_BNK_gamma1", "", "", bnk->gamma1, &bnk->gamma1,NULL);CHKERRQ(ierr); 763 ierr = PetscOptionsReal("-tao_BNK_gamma2", "", "", bnk->gamma2, &bnk->gamma2,NULL);CHKERRQ(ierr); 764 ierr = PetscOptionsReal("-tao_BNK_gamma3", "", "", bnk->gamma3, &bnk->gamma3,NULL);CHKERRQ(ierr); 765 ierr = PetscOptionsReal("-tao_BNK_gamma4", "", "", bnk->gamma4, &bnk->gamma4,NULL);CHKERRQ(ierr); 766 ierr = PetscOptionsReal("-tao_BNK_theta", "", "", bnk->theta, &bnk->theta,NULL);CHKERRQ(ierr); 767 ierr = PetscOptionsReal("-tao_BNK_min_radius", "lower bound on initial radius", "", bnk->min_radius, &bnk->min_radius,NULL);CHKERRQ(ierr); 768 ierr = PetscOptionsReal("-tao_BNK_max_radius", "upper bound on radius", "", bnk->max_radius, &bnk->max_radius,NULL);CHKERRQ(ierr); 769 ierr = PetscOptionsReal("-tao_BNK_epsilon", "tolerance used when computing actual and predicted reduction", "", bnk->epsilon, &bnk->epsilon,NULL);CHKERRQ(ierr); 770 ierr = PetscOptionsTail();CHKERRQ(ierr); 771 ierr = TaoLineSearchSetFromOptions(tao->linesearch);CHKERRQ(ierr); 772 ierr = KSPSetFromOptions(tao->ksp);CHKERRQ(ierr); 773 PetscFunctionReturn(0); 774 } 775 776 /*------------------------------------------------------------*/ 777 static PetscErrorCode TaoView_BNK(Tao tao, PetscViewer viewer) 778 { 779 TAO_BNK *bnk = (TAO_BNK *)tao->data; 780 PetscInt nrejects; 781 PetscBool isascii; 782 PetscErrorCode ierr; 783 784 PetscFunctionBegin; 785 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&isascii);CHKERRQ(ierr); 786 if (isascii) { 787 ierr = PetscViewerASCIIPushTab(viewer);CHKERRQ(ierr); 788 if (BNK_PC_BFGS == bnk->pc_type && bnk->M) { 789 ierr = MatLMVMGetRejects(bnk->M,&nrejects);CHKERRQ(ierr); 790 ierr = PetscViewerASCIIPrintf(viewer, "Rejected matrix updates: %D\n",nrejects);CHKERRQ(ierr); 791 } 792 ierr = PetscViewerASCIIPrintf(viewer, "Newton steps: %D\n", bnk->newt);CHKERRQ(ierr); 793 ierr = PetscViewerASCIIPrintf(viewer, "BFGS steps: %D\n", bnk->bfgs);CHKERRQ(ierr); 794 ierr = PetscViewerASCIIPrintf(viewer, "Scaled gradient steps: %D\n", bnk->sgrad);CHKERRQ(ierr); 795 ierr = PetscViewerASCIIPrintf(viewer, "Gradient steps: %D\n", bnk->grad);CHKERRQ(ierr); 796 ierr = PetscViewerASCIIPrintf(viewer, "KSP termination reasons:\n");CHKERRQ(ierr); 797 ierr = PetscViewerASCIIPrintf(viewer, " atol: %D\n", bnk->ksp_atol);CHKERRQ(ierr); 798 ierr = PetscViewerASCIIPrintf(viewer, " rtol: %D\n", bnk->ksp_rtol);CHKERRQ(ierr); 799 ierr = PetscViewerASCIIPrintf(viewer, " ctol: %D\n", bnk->ksp_ctol);CHKERRQ(ierr); 800 ierr = PetscViewerASCIIPrintf(viewer, " negc: %D\n", bnk->ksp_negc);CHKERRQ(ierr); 801 ierr = PetscViewerASCIIPrintf(viewer, " dtol: %D\n", bnk->ksp_dtol);CHKERRQ(ierr); 802 ierr = PetscViewerASCIIPrintf(viewer, " iter: %D\n", bnk->ksp_iter);CHKERRQ(ierr); 803 ierr = PetscViewerASCIIPrintf(viewer, " othr: %D\n", bnk->ksp_othr);CHKERRQ(ierr); 804 ierr = PetscViewerASCIIPopTab(viewer);CHKERRQ(ierr); 805 } 806 PetscFunctionReturn(0); 807 } 808 809 /* ---------------------------------------------------------- */ 810 /*MC 811 TAOBNK - Shared base-type for Bounded Newton-Krylov type algorithms. 812 At each iteration, the BNK method solves the symmetric 813 system of equations to obtain the step diretion dk: 814 Hk dk = -gk 815 for free variables only. The step can be globalized either through 816 trust-region methods, or a line search, or a heuristic mixture of both. 817 818 Options Database Keys: 819 + -tao_BNK_pc_type - "none","ahess","bfgs","petsc" 820 . -tao_BNK_bfgs_scale_type - "ahess","phess","bfgs" 821 . -tao_BNK_init_type - "constant","direction","interpolation" 822 . -tao_BNK_update_type - "step","direction","interpolation" 823 . -tao_BNK_sval - perturbation starting value 824 . -tao_BNK_imin - minimum initial perturbation 825 . -tao_BNK_imax - maximum initial perturbation 826 . -tao_BNK_pmin - minimum perturbation 827 . -tao_BNK_pmax - maximum perturbation 828 . -tao_BNK_pgfac - growth factor 829 . -tao_BNK_psfac - shrink factor 830 . -tao_BNK_imfac - initial merit factor 831 . -tao_BNK_pmgfac - merit growth factor 832 . -tao_BNK_pmsfac - merit shrink factor 833 . -tao_BNK_eta1 - poor steplength; reduce radius 834 . -tao_BNK_eta2 - reasonable steplength; leave radius 835 . -tao_BNK_eta3 - good steplength; increase readius 836 . -tao_BNK_eta4 - excellent steplength; greatly increase radius 837 . -tao_BNK_alpha1 - alpha1 reduction 838 . -tao_BNK_alpha2 - alpha2 reduction 839 . -tao_BNK_alpha3 - alpha3 reduction 840 . -tao_BNK_alpha4 - alpha4 reduction 841 . -tao_BNK_alpha - alpha5 reduction 842 . -tao_BNK_mu1 - mu1 interpolation update 843 . -tao_BNK_mu2 - mu2 interpolation update 844 . -tao_BNK_gamma1 - gamma1 interpolation update 845 . -tao_BNK_gamma2 - gamma2 interpolation update 846 . -tao_BNK_gamma3 - gamma3 interpolation update 847 . -tao_BNK_gamma4 - gamma4 interpolation update 848 . -tao_BNK_theta - theta interpolation update 849 . -tao_BNK_omega1 - omega1 step update 850 . -tao_BNK_omega2 - omega2 step update 851 . -tao_BNK_omega3 - omega3 step update 852 . -tao_BNK_omega4 - omega4 step update 853 . -tao_BNK_omega5 - omega5 step update 854 . -tao_BNK_mu1_i - mu1 interpolation init factor 855 . -tao_BNK_mu2_i - mu2 interpolation init factor 856 . -tao_BNK_gamma1_i - gamma1 interpolation init factor 857 . -tao_BNK_gamma2_i - gamma2 interpolation init factor 858 . -tao_BNK_gamma3_i - gamma3 interpolation init factor 859 . -tao_BNK_gamma4_i - gamma4 interpolation init factor 860 - -tao_BNK_theta_i - theta interpolation init factor 861 862 Level: beginner 863 M*/ 864 865 PetscErrorCode TaoCreate_BNK(Tao tao) 866 { 867 TAO_BNK *bnk; 868 const char *morethuente_type = TAOLINESEARCHMT; 869 PetscErrorCode ierr; 870 871 PetscFunctionBegin; 872 ierr = PetscNewLog(tao,&bnk);CHKERRQ(ierr); 873 874 tao->ops->setup = TaoSetUp_BNK; 875 tao->ops->view = TaoView_BNK; 876 tao->ops->setfromoptions = TaoSetFromOptions_BNK; 877 tao->ops->destroy = TaoDestroy_BNK; 878 879 /* Override default settings (unless already changed) */ 880 if (!tao->max_it_changed) tao->max_it = 50; 881 if (!tao->trust0_changed) tao->trust0 = 100.0; 882 883 tao->data = (void*)bnk; 884 885 bnk->sval = 0.0; 886 bnk->imin = 1.0e-4; 887 bnk->imax = 1.0e+2; 888 bnk->imfac = 1.0e-1; 889 890 bnk->pmin = 1.0e-12; 891 bnk->pmax = 1.0e+2; 892 bnk->pgfac = 1.0e+1; 893 bnk->psfac = 4.0e-1; 894 bnk->pmgfac = 1.0e-1; 895 bnk->pmsfac = 1.0e-1; 896 897 /* Default values for trust-region radius update based on steplength */ 898 bnk->nu1 = 0.25; 899 bnk->nu2 = 0.50; 900 bnk->nu3 = 1.00; 901 bnk->nu4 = 1.25; 902 903 bnk->omega1 = 0.25; 904 bnk->omega2 = 0.50; 905 bnk->omega3 = 1.00; 906 bnk->omega4 = 2.00; 907 bnk->omega5 = 4.00; 908 909 /* Default values for trust-region radius update based on reduction */ 910 bnk->eta1 = 1.0e-4; 911 bnk->eta2 = 0.25; 912 bnk->eta3 = 0.50; 913 bnk->eta4 = 0.90; 914 915 bnk->alpha1 = 0.25; 916 bnk->alpha2 = 0.50; 917 bnk->alpha3 = 1.00; 918 bnk->alpha4 = 2.00; 919 bnk->alpha5 = 4.00; 920 921 /* Default values for trust-region radius update based on interpolation */ 922 bnk->mu1 = 0.10; 923 bnk->mu2 = 0.50; 924 925 bnk->gamma1 = 0.25; 926 bnk->gamma2 = 0.50; 927 bnk->gamma3 = 2.00; 928 bnk->gamma4 = 4.00; 929 930 bnk->theta = 0.05; 931 932 /* Default values for trust region initialization based on interpolation */ 933 bnk->mu1_i = 0.35; 934 bnk->mu2_i = 0.50; 935 936 bnk->gamma1_i = 0.0625; 937 bnk->gamma2_i = 0.5; 938 bnk->gamma3_i = 2.0; 939 bnk->gamma4_i = 5.0; 940 941 bnk->theta_i = 0.25; 942 943 /* Remaining parameters */ 944 bnk->min_radius = 1.0e-10; 945 bnk->max_radius = 1.0e10; 946 bnk->epsilon = 1.0e-6; 947 948 bnk->pc_type = BNK_PC_BFGS; 949 bnk->bfgs_scale_type = BFGS_SCALE_PHESS; 950 bnk->init_type = BNK_INIT_INTERPOLATION; 951 bnk->update_type = BNK_UPDATE_STEP; 952 953 ierr = TaoLineSearchCreate(((PetscObject)tao)->comm,&tao->linesearch);CHKERRQ(ierr); 954 ierr = PetscObjectIncrementTabLevel((PetscObject)tao->linesearch, (PetscObject)tao, 1);CHKERRQ(ierr); 955 ierr = TaoLineSearchSetType(tao->linesearch,morethuente_type);CHKERRQ(ierr); 956 ierr = TaoLineSearchUseTaoRoutines(tao->linesearch,tao);CHKERRQ(ierr); 957 ierr = TaoLineSearchSetOptionsPrefix(tao->linesearch,tao->hdr.prefix);CHKERRQ(ierr); 958 959 /* Set linear solver to default for symmetric matrices */ 960 ierr = KSPCreate(((PetscObject)tao)->comm,&tao->ksp);CHKERRQ(ierr); 961 ierr = PetscObjectIncrementTabLevel((PetscObject)tao->ksp, (PetscObject)tao, 1);CHKERRQ(ierr); 962 ierr = KSPSetOptionsPrefix(tao->ksp,tao->hdr.prefix);CHKERRQ(ierr); 963 ierr = KSPSetType(tao->ksp,KSPCGSTCG);CHKERRQ(ierr); 964 PetscFunctionReturn(0); 965 } 966