1 #include <../src/tao/unconstrained/impls/ntl/ntlimpl.h> 2 3 #include <petscksp.h> 4 5 #define NTL_INIT_CONSTANT 0 6 #define NTL_INIT_DIRECTION 1 7 #define NTL_INIT_INTERPOLATION 2 8 #define NTL_INIT_TYPES 3 9 10 #define NTL_UPDATE_REDUCTION 0 11 #define NTL_UPDATE_INTERPOLATION 1 12 #define NTL_UPDATE_TYPES 2 13 14 static const char *NTL_INIT[64] = {"constant","direction","interpolation"}; 15 16 static const char *NTL_UPDATE[64] = {"reduction","interpolation"}; 17 18 /* Implements Newton's Method with a trust-region, line-search approach for 19 solving unconstrained minimization problems. A More'-Thuente line search 20 is used to guarantee that the bfgs preconditioner remains positive 21 definite. */ 22 23 #define NTL_NEWTON 0 24 #define NTL_BFGS 1 25 #define NTL_SCALED_GRADIENT 2 26 #define NTL_GRADIENT 3 27 28 static PetscErrorCode TaoSolve_NTL(Tao tao) 29 { 30 TAO_NTL *tl = (TAO_NTL *)tao->data; 31 KSPType ksp_type; 32 PetscBool is_nash,is_stcg,is_gltr,is_bfgs,is_jacobi,is_symmetric,sym_set; 33 KSPConvergedReason ksp_reason; 34 PC pc; 35 TaoLineSearchConvergedReason ls_reason; 36 37 PetscReal fmin, ftrial, prered, actred, kappa, sigma; 38 PetscReal tau, tau_1, tau_2, tau_max, tau_min, max_radius; 39 PetscReal f, fold, gdx, gnorm; 40 PetscReal step = 1.0; 41 42 PetscReal norm_d = 0.0; 43 PetscErrorCode ierr; 44 PetscInt stepType; 45 PetscInt its; 46 47 PetscInt bfgsUpdates = 0; 48 PetscInt needH; 49 50 PetscInt i_max = 5; 51 PetscInt j_max = 1; 52 PetscInt i, j, n, N; 53 54 PetscInt tr_reject; 55 56 PetscFunctionBegin; 57 if (tao->XL || tao->XU || tao->ops->computebounds) { 58 ierr = PetscInfo(tao,"WARNING: Variable bounds have been set but will be ignored by ntl algorithm\n");CHKERRQ(ierr); 59 } 60 61 ierr = KSPGetType(tao->ksp,&ksp_type);CHKERRQ(ierr); 62 ierr = PetscStrcmp(ksp_type,KSPNASH,&is_nash);CHKERRQ(ierr); 63 ierr = PetscStrcmp(ksp_type,KSPSTCG,&is_stcg);CHKERRQ(ierr); 64 ierr = PetscStrcmp(ksp_type,KSPGLTR,&is_gltr);CHKERRQ(ierr); 65 if (!is_nash && !is_stcg && !is_gltr) SETERRQ(PetscObjectComm((PetscObject)tao),PETSC_ERR_SUP,"TAO_NTR requires nash, stcg, or gltr for the KSP"); 66 67 /* Initialize the radius and modify if it is too large or small */ 68 tao->trust = tao->trust0; 69 tao->trust = PetscMax(tao->trust, tl->min_radius); 70 tao->trust = PetscMin(tao->trust, tl->max_radius); 71 72 /* Allocate the vectors needed for the BFGS approximation */ 73 ierr = KSPGetPC(tao->ksp, &pc);CHKERRQ(ierr); 74 ierr = PetscObjectTypeCompare((PetscObject)pc, PCLMVM, &is_bfgs);CHKERRQ(ierr); 75 ierr = PetscObjectTypeCompare((PetscObject)pc, PCJACOBI, &is_jacobi);CHKERRQ(ierr); 76 if (is_bfgs) { 77 tl->bfgs_pre = pc; 78 ierr = PCLMVMGetMatLMVM(tl->bfgs_pre, &tl->M);CHKERRQ(ierr); 79 ierr = VecGetLocalSize(tao->solution, &n);CHKERRQ(ierr); 80 ierr = VecGetSize(tao->solution, &N);CHKERRQ(ierr); 81 ierr = MatSetSizes(tl->M, n, n, N, N);CHKERRQ(ierr); 82 ierr = MatLMVMAllocate(tl->M, tao->solution, tao->gradient);CHKERRQ(ierr); 83 ierr = MatIsSymmetricKnown(tl->M, &sym_set, &is_symmetric);CHKERRQ(ierr); 84 if (!sym_set || !is_symmetric) SETERRQ(PetscObjectComm((PetscObject)tao), PETSC_ERR_ARG_INCOMP, "LMVM matrix in the LMVM preconditioner must be symmetric."); 85 } else if (is_jacobi) { 86 ierr = PCJacobiSetUseAbs(pc,PETSC_TRUE);CHKERRQ(ierr); 87 } 88 89 /* Check convergence criteria */ 90 ierr = TaoComputeObjectiveAndGradient(tao, tao->solution, &f, tao->gradient);CHKERRQ(ierr); 91 ierr = VecNorm(tao->gradient, NORM_2, &gnorm);CHKERRQ(ierr); 92 if (PetscIsInfOrNanReal(f) || PetscIsInfOrNanReal(gnorm)) SETERRQ(PetscObjectComm((PetscObject)tao),PETSC_ERR_USER, "User provided compute function generated Inf or NaN"); 93 needH = 1; 94 95 tao->reason = TAO_CONTINUE_ITERATING; 96 ierr = TaoLogConvergenceHistory(tao,f,gnorm,0.0,tao->ksp_its);CHKERRQ(ierr); 97 ierr = TaoMonitor(tao,tao->niter,f,gnorm,0.0,step);CHKERRQ(ierr); 98 ierr = (*tao->ops->convergencetest)(tao,tao->cnvP);CHKERRQ(ierr); 99 if (tao->reason != TAO_CONTINUE_ITERATING) PetscFunctionReturn(0); 100 101 /* Initialize trust-region radius */ 102 switch(tl->init_type) { 103 case NTL_INIT_CONSTANT: 104 /* Use the initial radius specified */ 105 break; 106 107 case NTL_INIT_INTERPOLATION: 108 /* Use the initial radius specified */ 109 max_radius = 0.0; 110 111 for (j = 0; j < j_max; ++j) { 112 fmin = f; 113 sigma = 0.0; 114 115 if (needH) { 116 ierr = TaoComputeHessian(tao,tao->solution,tao->hessian,tao->hessian_pre);CHKERRQ(ierr); 117 needH = 0; 118 } 119 120 for (i = 0; i < i_max; ++i) { 121 ierr = VecCopy(tao->solution, tl->W);CHKERRQ(ierr); 122 ierr = VecAXPY(tl->W, -tao->trust/gnorm, tao->gradient);CHKERRQ(ierr); 123 124 ierr = TaoComputeObjective(tao, tl->W, &ftrial);CHKERRQ(ierr); 125 if (PetscIsInfOrNanReal(ftrial)) { 126 tau = tl->gamma1_i; 127 } else { 128 if (ftrial < fmin) { 129 fmin = ftrial; 130 sigma = -tao->trust / gnorm; 131 } 132 133 ierr = MatMult(tao->hessian, tao->gradient, tao->stepdirection);CHKERRQ(ierr); 134 ierr = VecDot(tao->gradient, tao->stepdirection, &prered);CHKERRQ(ierr); 135 136 prered = tao->trust * (gnorm - 0.5 * tao->trust * prered / (gnorm * gnorm)); 137 actred = f - ftrial; 138 if ((PetscAbsScalar(actred) <= tl->epsilon) && (PetscAbsScalar(prered) <= tl->epsilon)) { 139 kappa = 1.0; 140 } else { 141 kappa = actred / prered; 142 } 143 144 tau_1 = tl->theta_i * gnorm * tao->trust / (tl->theta_i * gnorm * tao->trust + (1.0 - tl->theta_i) * prered - actred); 145 tau_2 = tl->theta_i * gnorm * tao->trust / (tl->theta_i * gnorm * tao->trust - (1.0 + tl->theta_i) * prered + actred); 146 tau_min = PetscMin(tau_1, tau_2); 147 tau_max = PetscMax(tau_1, tau_2); 148 149 if (PetscAbsScalar(kappa - (PetscReal)1.0) <= tl->mu1_i) { 150 /* Great agreement */ 151 max_radius = PetscMax(max_radius, tao->trust); 152 153 if (tau_max < 1.0) { 154 tau = tl->gamma3_i; 155 } else if (tau_max > tl->gamma4_i) { 156 tau = tl->gamma4_i; 157 } else if (tau_1 >= 1.0 && tau_1 <= tl->gamma4_i && tau_2 < 1.0) { 158 tau = tau_1; 159 } else if (tau_2 >= 1.0 && tau_2 <= tl->gamma4_i && tau_1 < 1.0) { 160 tau = tau_2; 161 } else { 162 tau = tau_max; 163 } 164 } else if (PetscAbsScalar(kappa - (PetscReal)1.0) <= tl->mu2_i) { 165 /* Good agreement */ 166 max_radius = PetscMax(max_radius, tao->trust); 167 168 if (tau_max < tl->gamma2_i) { 169 tau = tl->gamma2_i; 170 } else if (tau_max > tl->gamma3_i) { 171 tau = tl->gamma3_i; 172 } else { 173 tau = tau_max; 174 } 175 } else { 176 /* Not good agreement */ 177 if (tau_min > 1.0) { 178 tau = tl->gamma2_i; 179 } else if (tau_max < tl->gamma1_i) { 180 tau = tl->gamma1_i; 181 } else if ((tau_min < tl->gamma1_i) && (tau_max >= 1.0)) { 182 tau = tl->gamma1_i; 183 } else if ((tau_1 >= tl->gamma1_i) && (tau_1 < 1.0) && ((tau_2 < tl->gamma1_i) || (tau_2 >= 1.0))) { 184 tau = tau_1; 185 } else if ((tau_2 >= tl->gamma1_i) && (tau_2 < 1.0) && ((tau_1 < tl->gamma1_i) || (tau_2 >= 1.0))) { 186 tau = tau_2; 187 } else { 188 tau = tau_max; 189 } 190 } 191 } 192 tao->trust = tau * tao->trust; 193 } 194 195 if (fmin < f) { 196 f = fmin; 197 ierr = VecAXPY(tao->solution, sigma, tao->gradient);CHKERRQ(ierr); 198 ierr = TaoComputeGradient(tao, tao->solution, tao->gradient);CHKERRQ(ierr); 199 200 ierr = VecNorm(tao->gradient, NORM_2, &gnorm);CHKERRQ(ierr); 201 if (PetscIsInfOrNanReal(f) || PetscIsInfOrNanReal(gnorm)) SETERRQ(PetscObjectComm((PetscObject)tao),PETSC_ERR_USER, "User provided compute function generated Inf or NaN"); 202 needH = 1; 203 204 ierr = TaoLogConvergenceHistory(tao,f,gnorm,0.0,tao->ksp_its);CHKERRQ(ierr); 205 ierr = TaoMonitor(tao,tao->niter,f,gnorm,0.0,step);CHKERRQ(ierr); 206 ierr = (*tao->ops->convergencetest)(tao,tao->cnvP);CHKERRQ(ierr); 207 if (tao->reason != TAO_CONTINUE_ITERATING) PetscFunctionReturn(0); 208 } 209 } 210 tao->trust = PetscMax(tao->trust, max_radius); 211 212 /* Modify the radius if it is too large or small */ 213 tao->trust = PetscMax(tao->trust, tl->min_radius); 214 tao->trust = PetscMin(tao->trust, tl->max_radius); 215 break; 216 217 default: 218 /* Norm of the first direction will initialize radius */ 219 tao->trust = 0.0; 220 break; 221 } 222 223 /* Set counter for gradient/reset steps */ 224 tl->ntrust = 0; 225 tl->newt = 0; 226 tl->bfgs = 0; 227 tl->grad = 0; 228 229 /* Have not converged; continue with Newton method */ 230 while (tao->reason == TAO_CONTINUE_ITERATING) { 231 /* Call general purpose update function */ 232 if (tao->ops->update) { 233 ierr = (*tao->ops->update)(tao, tao->niter, tao->user_update);CHKERRQ(ierr); 234 } 235 ++tao->niter; 236 tao->ksp_its=0; 237 /* Compute the Hessian */ 238 if (needH) { 239 ierr = TaoComputeHessian(tao,tao->solution,tao->hessian,tao->hessian_pre);CHKERRQ(ierr); 240 } 241 242 if (tl->bfgs_pre) { 243 /* Update the limited memory preconditioner */ 244 ierr = MatLMVMUpdate(tl->M,tao->solution, tao->gradient);CHKERRQ(ierr); 245 ++bfgsUpdates; 246 } 247 ierr = KSPSetOperators(tao->ksp, tao->hessian, tao->hessian_pre);CHKERRQ(ierr); 248 /* Solve the Newton system of equations */ 249 ierr = KSPCGSetRadius(tao->ksp,tl->max_radius);CHKERRQ(ierr); 250 ierr = KSPSolve(tao->ksp, tao->gradient, tao->stepdirection);CHKERRQ(ierr); 251 ierr = KSPGetIterationNumber(tao->ksp,&its);CHKERRQ(ierr); 252 tao->ksp_its+=its; 253 tao->ksp_tot_its+=its; 254 ierr = KSPCGGetNormD(tao->ksp, &norm_d);CHKERRQ(ierr); 255 256 if (0.0 == tao->trust) { 257 /* Radius was uninitialized; use the norm of the direction */ 258 if (norm_d > 0.0) { 259 tao->trust = norm_d; 260 261 /* Modify the radius if it is too large or small */ 262 tao->trust = PetscMax(tao->trust, tl->min_radius); 263 tao->trust = PetscMin(tao->trust, tl->max_radius); 264 } else { 265 /* The direction was bad; set radius to default value and re-solve 266 the trust-region subproblem to get a direction */ 267 tao->trust = tao->trust0; 268 269 /* Modify the radius if it is too large or small */ 270 tao->trust = PetscMax(tao->trust, tl->min_radius); 271 tao->trust = PetscMin(tao->trust, tl->max_radius); 272 273 ierr = KSPCGSetRadius(tao->ksp,tl->max_radius);CHKERRQ(ierr); 274 ierr = KSPSolve(tao->ksp, tao->gradient, tao->stepdirection);CHKERRQ(ierr); 275 ierr = KSPGetIterationNumber(tao->ksp,&its);CHKERRQ(ierr); 276 tao->ksp_its+=its; 277 tao->ksp_tot_its+=its; 278 ierr = KSPCGGetNormD(tao->ksp, &norm_d);CHKERRQ(ierr); 279 280 if (norm_d == 0.0) SETERRQ(PetscObjectComm((PetscObject)tao),PETSC_ERR_PLIB, "Initial direction zero"); 281 } 282 } 283 284 ierr = VecScale(tao->stepdirection, -1.0);CHKERRQ(ierr); 285 ierr = KSPGetConvergedReason(tao->ksp, &ksp_reason);CHKERRQ(ierr); 286 if ((KSP_DIVERGED_INDEFINITE_PC == ksp_reason) && (tl->bfgs_pre)) { 287 /* Preconditioner is numerically indefinite; reset the 288 approximate if using BFGS preconditioning. */ 289 ierr = MatLMVMReset(tl->M, PETSC_FALSE);CHKERRQ(ierr); 290 ierr = MatLMVMUpdate(tl->M, tao->solution, tao->gradient);CHKERRQ(ierr); 291 bfgsUpdates = 1; 292 } 293 294 /* Check trust-region reduction conditions */ 295 tr_reject = 0; 296 if (NTL_UPDATE_REDUCTION == tl->update_type) { 297 /* Get predicted reduction */ 298 ierr = KSPCGGetObjFcn(tao->ksp,&prered);CHKERRQ(ierr); 299 if (prered >= 0.0) { 300 /* The predicted reduction has the wrong sign. This cannot 301 happen in infinite precision arithmetic. Step should 302 be rejected! */ 303 tao->trust = tl->alpha1 * PetscMin(tao->trust, norm_d); 304 tr_reject = 1; 305 } else { 306 /* Compute trial step and function value */ 307 ierr = VecCopy(tao->solution, tl->W);CHKERRQ(ierr); 308 ierr = VecAXPY(tl->W, 1.0, tao->stepdirection);CHKERRQ(ierr); 309 ierr = TaoComputeObjective(tao, tl->W, &ftrial);CHKERRQ(ierr); 310 311 if (PetscIsInfOrNanReal(ftrial)) { 312 tao->trust = tl->alpha1 * PetscMin(tao->trust, norm_d); 313 tr_reject = 1; 314 } else { 315 /* Compute and actual reduction */ 316 actred = f - ftrial; 317 prered = -prered; 318 if ((PetscAbsScalar(actred) <= tl->epsilon) && 319 (PetscAbsScalar(prered) <= tl->epsilon)) { 320 kappa = 1.0; 321 } else { 322 kappa = actred / prered; 323 } 324 325 /* Accept of reject the step and update radius */ 326 if (kappa < tl->eta1) { 327 /* Reject the step */ 328 tao->trust = tl->alpha1 * PetscMin(tao->trust, norm_d); 329 tr_reject = 1; 330 } else { 331 /* Accept the step */ 332 if (kappa < tl->eta2) { 333 /* Marginal bad step */ 334 tao->trust = tl->alpha2 * PetscMin(tao->trust, norm_d); 335 } else if (kappa < tl->eta3) { 336 /* Reasonable step */ 337 tao->trust = tl->alpha3 * tao->trust; 338 } else if (kappa < tl->eta4) { 339 /* Good step */ 340 tao->trust = PetscMax(tl->alpha4 * norm_d, tao->trust); 341 } else { 342 /* Very good step */ 343 tao->trust = PetscMax(tl->alpha5 * norm_d, tao->trust); 344 } 345 } 346 } 347 } 348 } else { 349 /* Get predicted reduction */ 350 ierr = KSPCGGetObjFcn(tao->ksp,&prered);CHKERRQ(ierr); 351 if (prered >= 0.0) { 352 /* The predicted reduction has the wrong sign. This cannot 353 happen in infinite precision arithmetic. Step should 354 be rejected! */ 355 tao->trust = tl->gamma1 * PetscMin(tao->trust, norm_d); 356 tr_reject = 1; 357 } else { 358 ierr = VecCopy(tao->solution, tl->W);CHKERRQ(ierr); 359 ierr = VecAXPY(tl->W, 1.0, tao->stepdirection);CHKERRQ(ierr); 360 ierr = TaoComputeObjective(tao, tl->W, &ftrial);CHKERRQ(ierr); 361 if (PetscIsInfOrNanReal(ftrial)) { 362 tao->trust = tl->gamma1 * PetscMin(tao->trust, norm_d); 363 tr_reject = 1; 364 } else { 365 ierr = VecDot(tao->gradient, tao->stepdirection, &gdx);CHKERRQ(ierr); 366 367 actred = f - ftrial; 368 prered = -prered; 369 if ((PetscAbsScalar(actred) <= tl->epsilon) && 370 (PetscAbsScalar(prered) <= tl->epsilon)) { 371 kappa = 1.0; 372 } else { 373 kappa = actred / prered; 374 } 375 376 tau_1 = tl->theta * gdx / (tl->theta * gdx - (1.0 - tl->theta) * prered + actred); 377 tau_2 = tl->theta * gdx / (tl->theta * gdx + (1.0 + tl->theta) * prered - actred); 378 tau_min = PetscMin(tau_1, tau_2); 379 tau_max = PetscMax(tau_1, tau_2); 380 381 if (kappa >= 1.0 - tl->mu1) { 382 /* Great agreement; accept step and update radius */ 383 if (tau_max < 1.0) { 384 tao->trust = PetscMax(tao->trust, tl->gamma3 * norm_d); 385 } else if (tau_max > tl->gamma4) { 386 tao->trust = PetscMax(tao->trust, tl->gamma4 * norm_d); 387 } else { 388 tao->trust = PetscMax(tao->trust, tau_max * norm_d); 389 } 390 } else if (kappa >= 1.0 - tl->mu2) { 391 /* Good agreement */ 392 393 if (tau_max < tl->gamma2) { 394 tao->trust = tl->gamma2 * PetscMin(tao->trust, norm_d); 395 } else if (tau_max > tl->gamma3) { 396 tao->trust = PetscMax(tao->trust, tl->gamma3 * norm_d); 397 } else if (tau_max < 1.0) { 398 tao->trust = tau_max * PetscMin(tao->trust, norm_d); 399 } else { 400 tao->trust = PetscMax(tao->trust, tau_max * norm_d); 401 } 402 } else { 403 /* Not good agreement */ 404 if (tau_min > 1.0) { 405 tao->trust = tl->gamma2 * PetscMin(tao->trust, norm_d); 406 } else if (tau_max < tl->gamma1) { 407 tao->trust = tl->gamma1 * PetscMin(tao->trust, norm_d); 408 } else if ((tau_min < tl->gamma1) && (tau_max >= 1.0)) { 409 tao->trust = tl->gamma1 * PetscMin(tao->trust, norm_d); 410 } else if ((tau_1 >= tl->gamma1) && (tau_1 < 1.0) && ((tau_2 < tl->gamma1) || (tau_2 >= 1.0))) { 411 tao->trust = tau_1 * PetscMin(tao->trust, norm_d); 412 } else if ((tau_2 >= tl->gamma1) && (tau_2 < 1.0) && ((tau_1 < tl->gamma1) || (tau_2 >= 1.0))) { 413 tao->trust = tau_2 * PetscMin(tao->trust, norm_d); 414 } else { 415 tao->trust = tau_max * PetscMin(tao->trust, norm_d); 416 } 417 tr_reject = 1; 418 } 419 } 420 } 421 } 422 423 if (tr_reject) { 424 /* The trust-region constraints rejected the step. Apply a linesearch. 425 Check for descent direction. */ 426 ierr = VecDot(tao->stepdirection, tao->gradient, &gdx);CHKERRQ(ierr); 427 if ((gdx >= 0.0) || PetscIsInfOrNanReal(gdx)) { 428 /* Newton step is not descent or direction produced Inf or NaN */ 429 430 if (!tl->bfgs_pre) { 431 /* We don't have the bfgs matrix around and updated 432 Must use gradient direction in this case */ 433 ierr = VecCopy(tao->gradient, tao->stepdirection);CHKERRQ(ierr); 434 ierr = VecScale(tao->stepdirection, -1.0);CHKERRQ(ierr); 435 ++tl->grad; 436 stepType = NTL_GRADIENT; 437 } else { 438 /* Attempt to use the BFGS direction */ 439 ierr = MatSolve(tl->M, tao->gradient, tao->stepdirection);CHKERRQ(ierr); 440 ierr = VecScale(tao->stepdirection, -1.0);CHKERRQ(ierr); 441 442 /* Check for success (descent direction) */ 443 ierr = VecDot(tao->stepdirection, tao->gradient, &gdx);CHKERRQ(ierr); 444 if ((gdx >= 0) || PetscIsInfOrNanReal(gdx)) { 445 /* BFGS direction is not descent or direction produced not a number 446 We can assert bfgsUpdates > 1 in this case because 447 the first solve produces the scaled gradient direction, 448 which is guaranteed to be descent */ 449 450 /* Use steepest descent direction (scaled) */ 451 ierr = MatLMVMReset(tl->M, PETSC_FALSE);CHKERRQ(ierr); 452 ierr = MatLMVMUpdate(tl->M, tao->solution, tao->gradient);CHKERRQ(ierr); 453 ierr = MatSolve(tl->M, tao->gradient, tao->stepdirection);CHKERRQ(ierr); 454 ierr = VecScale(tao->stepdirection, -1.0);CHKERRQ(ierr); 455 456 bfgsUpdates = 1; 457 ++tl->grad; 458 stepType = NTL_GRADIENT; 459 } else { 460 ierr = MatLMVMGetUpdateCount(tl->M, &bfgsUpdates);CHKERRQ(ierr); 461 if (1 == bfgsUpdates) { 462 /* The first BFGS direction is always the scaled gradient */ 463 ++tl->grad; 464 stepType = NTL_GRADIENT; 465 } else { 466 ++tl->bfgs; 467 stepType = NTL_BFGS; 468 } 469 } 470 } 471 } else { 472 /* Computed Newton step is descent */ 473 ++tl->newt; 474 stepType = NTL_NEWTON; 475 } 476 477 /* Perform the linesearch */ 478 fold = f; 479 ierr = VecCopy(tao->solution, tl->Xold);CHKERRQ(ierr); 480 ierr = VecCopy(tao->gradient, tl->Gold);CHKERRQ(ierr); 481 482 step = 1.0; 483 ierr = TaoLineSearchApply(tao->linesearch, tao->solution, &f, tao->gradient, tao->stepdirection, &step, &ls_reason);CHKERRQ(ierr); 484 ierr = TaoAddLineSearchCounts(tao);CHKERRQ(ierr); 485 486 while (ls_reason != TAOLINESEARCH_SUCCESS && ls_reason != TAOLINESEARCH_SUCCESS_USER && stepType != NTL_GRADIENT) { /* Linesearch failed */ 487 /* Linesearch failed */ 488 f = fold; 489 ierr = VecCopy(tl->Xold, tao->solution);CHKERRQ(ierr); 490 ierr = VecCopy(tl->Gold, tao->gradient);CHKERRQ(ierr); 491 492 switch(stepType) { 493 case NTL_NEWTON: 494 /* Failed to obtain acceptable iterate with Newton step */ 495 496 if (tl->bfgs_pre) { 497 /* We don't have the bfgs matrix around and being updated 498 Must use gradient direction in this case */ 499 ierr = VecCopy(tao->gradient, tao->stepdirection);CHKERRQ(ierr); 500 ++tl->grad; 501 stepType = NTL_GRADIENT; 502 } else { 503 /* Attempt to use the BFGS direction */ 504 ierr = MatSolve(tl->M, tao->gradient, tao->stepdirection);CHKERRQ(ierr); 505 506 /* Check for success (descent direction) */ 507 ierr = VecDot(tao->stepdirection, tao->gradient, &gdx);CHKERRQ(ierr); 508 if ((gdx <= 0) || PetscIsInfOrNanReal(gdx)) { 509 /* BFGS direction is not descent or direction produced 510 not a number. We can assert bfgsUpdates > 1 in this case 511 Use steepest descent direction (scaled) */ 512 ierr = MatLMVMReset(tl->M, PETSC_FALSE);CHKERRQ(ierr); 513 ierr = MatLMVMUpdate(tl->M, tao->solution, tao->gradient);CHKERRQ(ierr); 514 ierr = MatSolve(tl->M, tao->gradient, tao->stepdirection);CHKERRQ(ierr); 515 516 bfgsUpdates = 1; 517 ++tl->grad; 518 stepType = NTL_GRADIENT; 519 } else { 520 ierr = MatLMVMGetUpdateCount(tl->M, &bfgsUpdates);CHKERRQ(ierr); 521 if (1 == bfgsUpdates) { 522 /* The first BFGS direction is always the scaled gradient */ 523 ++tl->grad; 524 stepType = NTL_GRADIENT; 525 } else { 526 ++tl->bfgs; 527 stepType = NTL_BFGS; 528 } 529 } 530 } 531 break; 532 533 case NTL_BFGS: 534 /* Can only enter if pc_type == NTL_PC_BFGS 535 Failed to obtain acceptable iterate with BFGS step 536 Attempt to use the scaled gradient direction */ 537 ierr = MatLMVMReset(tl->M, PETSC_FALSE);CHKERRQ(ierr); 538 ierr = MatLMVMUpdate(tl->M, tao->solution, tao->gradient);CHKERRQ(ierr); 539 ierr = MatSolve(tl->M, tao->gradient, tao->stepdirection);CHKERRQ(ierr); 540 541 bfgsUpdates = 1; 542 ++tl->grad; 543 stepType = NTL_GRADIENT; 544 break; 545 } 546 ierr = VecScale(tao->stepdirection, -1.0);CHKERRQ(ierr); 547 548 /* This may be incorrect; linesearch has values for stepmax and stepmin 549 that should be reset. */ 550 step = 1.0; 551 ierr = TaoLineSearchApply(tao->linesearch, tao->solution, &f, tao->gradient, tao->stepdirection, &step, &ls_reason);CHKERRQ(ierr); 552 ierr = TaoAddLineSearchCounts(tao);CHKERRQ(ierr); 553 } 554 555 if (ls_reason != TAOLINESEARCH_SUCCESS && ls_reason != TAOLINESEARCH_SUCCESS_USER) { 556 /* Failed to find an improving point */ 557 f = fold; 558 ierr = VecCopy(tl->Xold, tao->solution);CHKERRQ(ierr); 559 ierr = VecCopy(tl->Gold, tao->gradient);CHKERRQ(ierr); 560 tao->trust = 0.0; 561 step = 0.0; 562 tao->reason = TAO_DIVERGED_LS_FAILURE; 563 break; 564 } else if (stepType == NTL_NEWTON) { 565 if (step < tl->nu1) { 566 /* Very bad step taken; reduce radius */ 567 tao->trust = tl->omega1 * PetscMin(norm_d, tao->trust); 568 } else if (step < tl->nu2) { 569 /* Reasonably bad step taken; reduce radius */ 570 tao->trust = tl->omega2 * PetscMin(norm_d, tao->trust); 571 } else if (step < tl->nu3) { 572 /* Reasonable step was taken; leave radius alone */ 573 if (tl->omega3 < 1.0) { 574 tao->trust = tl->omega3 * PetscMin(norm_d, tao->trust); 575 } else if (tl->omega3 > 1.0) { 576 tao->trust = PetscMax(tl->omega3 * norm_d, tao->trust); 577 } 578 } else if (step < tl->nu4) { 579 /* Full step taken; increase the radius */ 580 tao->trust = PetscMax(tl->omega4 * norm_d, tao->trust); 581 } else { 582 /* More than full step taken; increase the radius */ 583 tao->trust = PetscMax(tl->omega5 * norm_d, tao->trust); 584 } 585 } else { 586 /* Newton step was not good; reduce the radius */ 587 tao->trust = tl->omega1 * PetscMin(norm_d, tao->trust); 588 } 589 } else { 590 /* Trust-region step is accepted */ 591 ierr = VecCopy(tl->W, tao->solution);CHKERRQ(ierr); 592 f = ftrial; 593 ierr = TaoComputeGradient(tao, tao->solution, tao->gradient);CHKERRQ(ierr); 594 ++tl->ntrust; 595 } 596 597 /* The radius may have been increased; modify if it is too large */ 598 tao->trust = PetscMin(tao->trust, tl->max_radius); 599 600 /* Check for converged */ 601 ierr = VecNorm(tao->gradient, NORM_2, &gnorm);CHKERRQ(ierr); 602 if (PetscIsInfOrNanReal(f) || PetscIsInfOrNanReal(gnorm)) SETERRQ(PetscObjectComm((PetscObject)tao),PETSC_ERR_USER,"User provided compute function generated Not-a-Number"); 603 needH = 1; 604 605 ierr = TaoLogConvergenceHistory(tao,f,gnorm,0.0,tao->ksp_its);CHKERRQ(ierr); 606 ierr = TaoMonitor(tao,tao->niter,f,gnorm,0.0,step);CHKERRQ(ierr); 607 ierr = (*tao->ops->convergencetest)(tao,tao->cnvP);CHKERRQ(ierr); 608 } 609 PetscFunctionReturn(0); 610 } 611 612 /* ---------------------------------------------------------- */ 613 static PetscErrorCode TaoSetUp_NTL(Tao tao) 614 { 615 TAO_NTL *tl = (TAO_NTL *)tao->data; 616 PetscErrorCode ierr; 617 618 PetscFunctionBegin; 619 if (!tao->gradient) {ierr = VecDuplicate(tao->solution, &tao->gradient);CHKERRQ(ierr); } 620 if (!tao->stepdirection) {ierr = VecDuplicate(tao->solution, &tao->stepdirection);CHKERRQ(ierr);} 621 if (!tl->W) { ierr = VecDuplicate(tao->solution, &tl->W);CHKERRQ(ierr);} 622 if (!tl->Xold) { ierr = VecDuplicate(tao->solution, &tl->Xold);CHKERRQ(ierr);} 623 if (!tl->Gold) { ierr = VecDuplicate(tao->solution, &tl->Gold);CHKERRQ(ierr);} 624 tl->bfgs_pre = NULL; 625 tl->M = NULL; 626 PetscFunctionReturn(0); 627 } 628 629 /*------------------------------------------------------------*/ 630 static PetscErrorCode TaoDestroy_NTL(Tao tao) 631 { 632 TAO_NTL *tl = (TAO_NTL *)tao->data; 633 PetscErrorCode ierr; 634 635 PetscFunctionBegin; 636 if (tao->setupcalled) { 637 ierr = VecDestroy(&tl->W);CHKERRQ(ierr); 638 ierr = VecDestroy(&tl->Xold);CHKERRQ(ierr); 639 ierr = VecDestroy(&tl->Gold);CHKERRQ(ierr); 640 } 641 ierr = PetscFree(tao->data);CHKERRQ(ierr); 642 PetscFunctionReturn(0); 643 } 644 645 /*------------------------------------------------------------*/ 646 static PetscErrorCode TaoSetFromOptions_NTL(PetscOptionItems *PetscOptionsObject,Tao tao) 647 { 648 TAO_NTL *tl = (TAO_NTL *)tao->data; 649 PetscErrorCode ierr; 650 651 PetscFunctionBegin; 652 ierr = PetscOptionsHead(PetscOptionsObject,"Newton trust region with line search method for unconstrained optimization");CHKERRQ(ierr); 653 ierr = PetscOptionsEList("-tao_ntl_init_type", "radius initialization type", "", NTL_INIT, NTL_INIT_TYPES, NTL_INIT[tl->init_type], &tl->init_type,NULL);CHKERRQ(ierr); 654 ierr = PetscOptionsEList("-tao_ntl_update_type", "radius update type", "", NTL_UPDATE, NTL_UPDATE_TYPES, NTL_UPDATE[tl->update_type], &tl->update_type,NULL);CHKERRQ(ierr); 655 ierr = PetscOptionsReal("-tao_ntl_eta1", "poor steplength; reduce radius", "", tl->eta1, &tl->eta1,NULL);CHKERRQ(ierr); 656 ierr = PetscOptionsReal("-tao_ntl_eta2", "reasonable steplength; leave radius alone", "", tl->eta2, &tl->eta2,NULL);CHKERRQ(ierr); 657 ierr = PetscOptionsReal("-tao_ntl_eta3", "good steplength; increase radius", "", tl->eta3, &tl->eta3,NULL);CHKERRQ(ierr); 658 ierr = PetscOptionsReal("-tao_ntl_eta4", "excellent steplength; greatly increase radius", "", tl->eta4, &tl->eta4,NULL);CHKERRQ(ierr); 659 ierr = PetscOptionsReal("-tao_ntl_alpha1", "", "", tl->alpha1, &tl->alpha1,NULL);CHKERRQ(ierr); 660 ierr = PetscOptionsReal("-tao_ntl_alpha2", "", "", tl->alpha2, &tl->alpha2,NULL);CHKERRQ(ierr); 661 ierr = PetscOptionsReal("-tao_ntl_alpha3", "", "", tl->alpha3, &tl->alpha3,NULL);CHKERRQ(ierr); 662 ierr = PetscOptionsReal("-tao_ntl_alpha4", "", "", tl->alpha4, &tl->alpha4,NULL);CHKERRQ(ierr); 663 ierr = PetscOptionsReal("-tao_ntl_alpha5", "", "", tl->alpha5, &tl->alpha5,NULL);CHKERRQ(ierr); 664 ierr = PetscOptionsReal("-tao_ntl_nu1", "poor steplength; reduce radius", "", tl->nu1, &tl->nu1,NULL);CHKERRQ(ierr); 665 ierr = PetscOptionsReal("-tao_ntl_nu2", "reasonable steplength; leave radius alone", "", tl->nu2, &tl->nu2,NULL);CHKERRQ(ierr); 666 ierr = PetscOptionsReal("-tao_ntl_nu3", "good steplength; increase radius", "", tl->nu3, &tl->nu3,NULL);CHKERRQ(ierr); 667 ierr = PetscOptionsReal("-tao_ntl_nu4", "excellent steplength; greatly increase radius", "", tl->nu4, &tl->nu4,NULL);CHKERRQ(ierr); 668 ierr = PetscOptionsReal("-tao_ntl_omega1", "", "", tl->omega1, &tl->omega1,NULL);CHKERRQ(ierr); 669 ierr = PetscOptionsReal("-tao_ntl_omega2", "", "", tl->omega2, &tl->omega2,NULL);CHKERRQ(ierr); 670 ierr = PetscOptionsReal("-tao_ntl_omega3", "", "", tl->omega3, &tl->omega3,NULL);CHKERRQ(ierr); 671 ierr = PetscOptionsReal("-tao_ntl_omega4", "", "", tl->omega4, &tl->omega4,NULL);CHKERRQ(ierr); 672 ierr = PetscOptionsReal("-tao_ntl_omega5", "", "", tl->omega5, &tl->omega5,NULL);CHKERRQ(ierr); 673 ierr = PetscOptionsReal("-tao_ntl_mu1_i", "", "", tl->mu1_i, &tl->mu1_i,NULL);CHKERRQ(ierr); 674 ierr = PetscOptionsReal("-tao_ntl_mu2_i", "", "", tl->mu2_i, &tl->mu2_i,NULL);CHKERRQ(ierr); 675 ierr = PetscOptionsReal("-tao_ntl_gamma1_i", "", "", tl->gamma1_i, &tl->gamma1_i,NULL);CHKERRQ(ierr); 676 ierr = PetscOptionsReal("-tao_ntl_gamma2_i", "", "", tl->gamma2_i, &tl->gamma2_i,NULL);CHKERRQ(ierr); 677 ierr = PetscOptionsReal("-tao_ntl_gamma3_i", "", "", tl->gamma3_i, &tl->gamma3_i,NULL);CHKERRQ(ierr); 678 ierr = PetscOptionsReal("-tao_ntl_gamma4_i", "", "", tl->gamma4_i, &tl->gamma4_i,NULL);CHKERRQ(ierr); 679 ierr = PetscOptionsReal("-tao_ntl_theta_i", "", "", tl->theta_i, &tl->theta_i,NULL);CHKERRQ(ierr); 680 ierr = PetscOptionsReal("-tao_ntl_mu1", "", "", tl->mu1, &tl->mu1,NULL);CHKERRQ(ierr); 681 ierr = PetscOptionsReal("-tao_ntl_mu2", "", "", tl->mu2, &tl->mu2,NULL);CHKERRQ(ierr); 682 ierr = PetscOptionsReal("-tao_ntl_gamma1", "", "", tl->gamma1, &tl->gamma1,NULL);CHKERRQ(ierr); 683 ierr = PetscOptionsReal("-tao_ntl_gamma2", "", "", tl->gamma2, &tl->gamma2,NULL);CHKERRQ(ierr); 684 ierr = PetscOptionsReal("-tao_ntl_gamma3", "", "", tl->gamma3, &tl->gamma3,NULL);CHKERRQ(ierr); 685 ierr = PetscOptionsReal("-tao_ntl_gamma4", "", "", tl->gamma4, &tl->gamma4,NULL);CHKERRQ(ierr); 686 ierr = PetscOptionsReal("-tao_ntl_theta", "", "", tl->theta, &tl->theta,NULL);CHKERRQ(ierr); 687 ierr = PetscOptionsReal("-tao_ntl_min_radius", "lower bound on initial radius", "", tl->min_radius, &tl->min_radius,NULL);CHKERRQ(ierr); 688 ierr = PetscOptionsReal("-tao_ntl_max_radius", "upper bound on radius", "", tl->max_radius, &tl->max_radius,NULL);CHKERRQ(ierr); 689 ierr = PetscOptionsReal("-tao_ntl_epsilon", "tolerance used when computing actual and predicted reduction", "", tl->epsilon, &tl->epsilon,NULL);CHKERRQ(ierr); 690 ierr = PetscOptionsTail();CHKERRQ(ierr); 691 ierr = TaoLineSearchSetFromOptions(tao->linesearch);CHKERRQ(ierr); 692 ierr = KSPSetFromOptions(tao->ksp);CHKERRQ(ierr); 693 PetscFunctionReturn(0); 694 } 695 696 /*------------------------------------------------------------*/ 697 static PetscErrorCode TaoView_NTL(Tao tao, PetscViewer viewer) 698 { 699 TAO_NTL *tl = (TAO_NTL *)tao->data; 700 PetscBool isascii; 701 PetscErrorCode ierr; 702 703 PetscFunctionBegin; 704 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&isascii);CHKERRQ(ierr); 705 if (isascii) { 706 ierr = PetscViewerASCIIPushTab(viewer);CHKERRQ(ierr); 707 ierr = PetscViewerASCIIPrintf(viewer, "Trust-region steps: %D\n", tl->ntrust);CHKERRQ(ierr); 708 ierr = PetscViewerASCIIPrintf(viewer, "Newton search steps: %D\n", tl->newt);CHKERRQ(ierr); 709 ierr = PetscViewerASCIIPrintf(viewer, "BFGS search steps: %D\n", tl->bfgs);CHKERRQ(ierr); 710 ierr = PetscViewerASCIIPrintf(viewer, "Gradient search steps: %D\n", tl->grad);CHKERRQ(ierr); 711 ierr = PetscViewerASCIIPopTab(viewer);CHKERRQ(ierr); 712 } 713 PetscFunctionReturn(0); 714 } 715 716 /* ---------------------------------------------------------- */ 717 /*MC 718 TAONTL - Newton's method with trust region globalization and line search fallback. 719 At each iteration, the Newton trust region method solves the system for d 720 and performs a line search in the d direction: 721 722 min_d .5 dT Hk d + gkT d, s.t. ||d|| < Delta_k 723 724 Options Database Keys: 725 + -tao_ntl_init_type - "constant","direction","interpolation" 726 . -tao_ntl_update_type - "reduction","interpolation" 727 . -tao_ntl_min_radius - lower bound on trust region radius 728 . -tao_ntl_max_radius - upper bound on trust region radius 729 . -tao_ntl_epsilon - tolerance for accepting actual / predicted reduction 730 . -tao_ntl_mu1_i - mu1 interpolation init factor 731 . -tao_ntl_mu2_i - mu2 interpolation init factor 732 . -tao_ntl_gamma1_i - gamma1 interpolation init factor 733 . -tao_ntl_gamma2_i - gamma2 interpolation init factor 734 . -tao_ntl_gamma3_i - gamma3 interpolation init factor 735 . -tao_ntl_gamma4_i - gamma4 interpolation init factor 736 . -tao_ntl_theta_i - theta1 interpolation init factor 737 . -tao_ntl_eta1 - eta1 reduction update factor 738 . -tao_ntl_eta2 - eta2 reduction update factor 739 . -tao_ntl_eta3 - eta3 reduction update factor 740 . -tao_ntl_eta4 - eta4 reduction update factor 741 . -tao_ntl_alpha1 - alpha1 reduction update factor 742 . -tao_ntl_alpha2 - alpha2 reduction update factor 743 . -tao_ntl_alpha3 - alpha3 reduction update factor 744 . -tao_ntl_alpha4 - alpha4 reduction update factor 745 . -tao_ntl_alpha4 - alpha4 reduction update factor 746 . -tao_ntl_mu1 - mu1 interpolation update 747 . -tao_ntl_mu2 - mu2 interpolation update 748 . -tao_ntl_gamma1 - gamma1 interpolcation update 749 . -tao_ntl_gamma2 - gamma2 interpolcation update 750 . -tao_ntl_gamma3 - gamma3 interpolcation update 751 . -tao_ntl_gamma4 - gamma4 interpolation update 752 - -tao_ntl_theta - theta1 interpolation update 753 754 Level: beginner 755 M*/ 756 PETSC_EXTERN PetscErrorCode TaoCreate_NTL(Tao tao) 757 { 758 TAO_NTL *tl; 759 PetscErrorCode ierr; 760 const char *morethuente_type = TAOLINESEARCHMT; 761 762 PetscFunctionBegin; 763 ierr = PetscNewLog(tao,&tl);CHKERRQ(ierr); 764 tao->ops->setup = TaoSetUp_NTL; 765 tao->ops->solve = TaoSolve_NTL; 766 tao->ops->view = TaoView_NTL; 767 tao->ops->setfromoptions = TaoSetFromOptions_NTL; 768 tao->ops->destroy = TaoDestroy_NTL; 769 770 /* Override default settings (unless already changed) */ 771 if (!tao->max_it_changed) tao->max_it = 50; 772 if (!tao->trust0_changed) tao->trust0 = 100.0; 773 774 tao->data = (void*)tl; 775 776 /* Default values for trust-region radius update based on steplength */ 777 tl->nu1 = 0.25; 778 tl->nu2 = 0.50; 779 tl->nu3 = 1.00; 780 tl->nu4 = 1.25; 781 782 tl->omega1 = 0.25; 783 tl->omega2 = 0.50; 784 tl->omega3 = 1.00; 785 tl->omega4 = 2.00; 786 tl->omega5 = 4.00; 787 788 /* Default values for trust-region radius update based on reduction */ 789 tl->eta1 = 1.0e-4; 790 tl->eta2 = 0.25; 791 tl->eta3 = 0.50; 792 tl->eta4 = 0.90; 793 794 tl->alpha1 = 0.25; 795 tl->alpha2 = 0.50; 796 tl->alpha3 = 1.00; 797 tl->alpha4 = 2.00; 798 tl->alpha5 = 4.00; 799 800 /* Default values for trust-region radius update based on interpolation */ 801 tl->mu1 = 0.10; 802 tl->mu2 = 0.50; 803 804 tl->gamma1 = 0.25; 805 tl->gamma2 = 0.50; 806 tl->gamma3 = 2.00; 807 tl->gamma4 = 4.00; 808 809 tl->theta = 0.05; 810 811 /* Default values for trust region initialization based on interpolation */ 812 tl->mu1_i = 0.35; 813 tl->mu2_i = 0.50; 814 815 tl->gamma1_i = 0.0625; 816 tl->gamma2_i = 0.5; 817 tl->gamma3_i = 2.0; 818 tl->gamma4_i = 5.0; 819 820 tl->theta_i = 0.25; 821 822 /* Remaining parameters */ 823 tl->min_radius = 1.0e-10; 824 tl->max_radius = 1.0e10; 825 tl->epsilon = 1.0e-6; 826 827 tl->init_type = NTL_INIT_INTERPOLATION; 828 tl->update_type = NTL_UPDATE_REDUCTION; 829 830 ierr = TaoLineSearchCreate(((PetscObject)tao)->comm, &tao->linesearch);CHKERRQ(ierr); 831 ierr = PetscObjectIncrementTabLevel((PetscObject)tao->linesearch,(PetscObject)tao,1);CHKERRQ(ierr); 832 ierr = TaoLineSearchSetType(tao->linesearch, morethuente_type);CHKERRQ(ierr); 833 ierr = TaoLineSearchUseTaoRoutines(tao->linesearch, tao);CHKERRQ(ierr); 834 ierr = TaoLineSearchSetOptionsPrefix(tao->linesearch,tao->hdr.prefix);CHKERRQ(ierr); 835 ierr = KSPCreate(((PetscObject)tao)->comm,&tao->ksp);CHKERRQ(ierr); 836 ierr = PetscObjectIncrementTabLevel((PetscObject)tao->ksp,(PetscObject)tao,1);CHKERRQ(ierr); 837 ierr = KSPSetOptionsPrefix(tao->ksp,tao->hdr.prefix);CHKERRQ(ierr); 838 ierr = KSPAppendOptionsPrefix(tao->ksp,"tao_ntl_");CHKERRQ(ierr); 839 ierr = KSPSetType(tao->ksp,KSPSTCG);CHKERRQ(ierr); 840 PetscFunctionReturn(0); 841 } 842