| 1 |
#include <math.h> |
| 2 |
#include "OOPSEMinimizer.hpp" |
| 3 |
#include "ShakeMin.hpp" |
| 4 |
#include "Integrator.cpp" |
| 5 |
|
| 6 |
OOPSEMinimizer::OOPSEMinimizer( SimInfo *theInfo, ForceFields* the_ff , |
| 7 |
MinimizerParameterSet * param) : |
| 8 |
RealIntegrator(theInfo, the_ff), bShake(true), bVerbose(false) { |
| 9 |
dumpOut = NULL; |
| 10 |
statOut = NULL; |
| 11 |
|
| 12 |
tStats = new Thermo(info); |
| 13 |
|
| 14 |
|
| 15 |
paramSet = param; |
| 16 |
|
| 17 |
calcDim(); |
| 18 |
|
| 19 |
curX = getCoor(); |
| 20 |
curG.resize(ndim); |
| 21 |
|
| 22 |
shakeAlgo = new ShakeMinFramework(theInfo); |
| 23 |
shakeAlgo->doPreConstraint(); |
| 24 |
} |
| 25 |
|
| 26 |
OOPSEMinimizer::~OOPSEMinimizer(){ |
| 27 |
delete tStats; |
| 28 |
if(dumpOut) |
| 29 |
delete dumpOut; |
| 30 |
if(statOut) |
| 31 |
delete statOut; |
| 32 |
delete paramSet; |
| 33 |
} |
| 34 |
|
| 35 |
void OOPSEMinimizer::calcEnergyGradient(vector<double>& x, vector<double>& grad, |
| 36 |
double& energy, int& status){ |
| 37 |
|
| 38 |
DirectionalAtom* dAtom; |
| 39 |
int index; |
| 40 |
double force[3]; |
| 41 |
double dAtomGrad[6]; |
| 42 |
int shakeStatus; |
| 43 |
|
| 44 |
status = 1; |
| 45 |
|
| 46 |
setCoor(x); |
| 47 |
|
| 48 |
if (bShake){ |
| 49 |
shakeStatus = shakeAlgo->doShakeR(); |
| 50 |
if(shakeStatus < 0) |
| 51 |
status = -1; |
| 52 |
} |
| 53 |
|
| 54 |
calcForce(1, 1); |
| 55 |
|
| 56 |
if (bShake){ |
| 57 |
shakeStatus = shakeAlgo->doShakeF(); |
| 58 |
if(shakeStatus < 0) |
| 59 |
status = -1; |
| 60 |
} |
| 61 |
|
| 62 |
x = getCoor(); |
| 63 |
|
| 64 |
|
| 65 |
index = 0; |
| 66 |
|
| 67 |
for(int i = 0; i < integrableObjects.size(); i++){ |
| 68 |
|
| 69 |
if (integrableObjects[i]->isDirectional()) { |
| 70 |
|
| 71 |
integrableObjects[i]->getGrad(dAtomGrad); |
| 72 |
|
| 73 |
//gradient is equal to -f |
| 74 |
grad[index++] = -dAtomGrad[0]; |
| 75 |
grad[index++] = -dAtomGrad[1]; |
| 76 |
grad[index++] = -dAtomGrad[2]; |
| 77 |
grad[index++] = -dAtomGrad[3]; |
| 78 |
grad[index++] = -dAtomGrad[4]; |
| 79 |
grad[index++] = -dAtomGrad[5]; |
| 80 |
|
| 81 |
} |
| 82 |
else{ |
| 83 |
integrableObjects[i]->getFrc(force); |
| 84 |
|
| 85 |
grad[index++] = -force[0]; |
| 86 |
grad[index++] = -force[1]; |
| 87 |
grad[index++] = -force[2]; |
| 88 |
|
| 89 |
} |
| 90 |
|
| 91 |
} |
| 92 |
|
| 93 |
energy = tStats->getPotential(); |
| 94 |
|
| 95 |
} |
| 96 |
|
| 97 |
/** |
| 98 |
* |
| 99 |
*/ |
| 100 |
|
| 101 |
void OOPSEMinimizer::setCoor(vector<double>& x){ |
| 102 |
|
| 103 |
DirectionalAtom* dAtom; |
| 104 |
int index; |
| 105 |
double position[3]; |
| 106 |
double eulerAngle[3]; |
| 107 |
|
| 108 |
|
| 109 |
index = 0; |
| 110 |
|
| 111 |
for(int i = 0; i < integrableObjects.size(); i++){ |
| 112 |
|
| 113 |
position[0] = x[index++]; |
| 114 |
position[1] = x[index++]; |
| 115 |
position[2] = x[index++]; |
| 116 |
|
| 117 |
integrableObjects[i]->setPos(position); |
| 118 |
|
| 119 |
if (integrableObjects[i]->isDirectional()){ |
| 120 |
|
| 121 |
eulerAngle[0] = x[index++]; |
| 122 |
eulerAngle[1] = x[index++]; |
| 123 |
eulerAngle[2] = x[index++]; |
| 124 |
|
| 125 |
integrableObjects[i]->setEuler(eulerAngle[0], |
| 126 |
eulerAngle[1], |
| 127 |
eulerAngle[2]); |
| 128 |
|
| 129 |
} |
| 130 |
|
| 131 |
} |
| 132 |
|
| 133 |
} |
| 134 |
|
| 135 |
/** |
| 136 |
* |
| 137 |
*/ |
| 138 |
vector<double> OOPSEMinimizer::getCoor(){ |
| 139 |
|
| 140 |
DirectionalAtom* dAtom; |
| 141 |
int index; |
| 142 |
double position[3]; |
| 143 |
double eulerAngle[3]; |
| 144 |
vector<double> x; |
| 145 |
|
| 146 |
x.resize(getDim()); |
| 147 |
|
| 148 |
index = 0; |
| 149 |
|
| 150 |
for(int i = 0; i < integrableObjects.size(); i++){ |
| 151 |
integrableObjects[i]->getPos(position); |
| 152 |
|
| 153 |
x[index++] = position[0]; |
| 154 |
x[index++] = position[1]; |
| 155 |
x[index++] = position[2]; |
| 156 |
|
| 157 |
if (integrableObjects[i]->isDirectional()){ |
| 158 |
|
| 159 |
integrableObjects[i]->getEulerAngles(eulerAngle); |
| 160 |
|
| 161 |
x[index++] = eulerAngle[0]; |
| 162 |
x[index++] = eulerAngle[1]; |
| 163 |
x[index++] = eulerAngle[2]; |
| 164 |
|
| 165 |
} |
| 166 |
|
| 167 |
} |
| 168 |
|
| 169 |
return x; |
| 170 |
|
| 171 |
} |
| 172 |
|
| 173 |
/* |
| 174 |
int OOPSEMinimizer::shakeR(){ |
| 175 |
int i, j; |
| 176 |
int done; |
| 177 |
double posA[3], posB[3]; |
| 178 |
double velA[3], velB[3]; |
| 179 |
double pab[3]; |
| 180 |
double rab[3]; |
| 181 |
int a, b, ax, ay, az, bx, by, bz; |
| 182 |
double rma, rmb; |
| 183 |
double dx, dy, dz; |
| 184 |
double rpab; |
| 185 |
double rabsq, pabsq, rpabsq; |
| 186 |
double diffsq; |
| 187 |
double gab; |
| 188 |
int iteration; |
| 189 |
|
| 190 |
for (i = 0; i < nAtoms; i++){ |
| 191 |
moving[i] = 0; |
| 192 |
moved[i] = 1; |
| 193 |
} |
| 194 |
|
| 195 |
iteration = 0; |
| 196 |
done = 0; |
| 197 |
while (!done && (iteration < maxIteration)){ |
| 198 |
done = 1; |
| 199 |
for (i = 0; i < nConstrained; i++){ |
| 200 |
a = constrainedA[i]; |
| 201 |
b = constrainedB[i]; |
| 202 |
|
| 203 |
ax = (a * 3) + 0; |
| 204 |
ay = (a * 3) + 1; |
| 205 |
az = (a * 3) + 2; |
| 206 |
|
| 207 |
bx = (b * 3) + 0; |
| 208 |
by = (b * 3) + 1; |
| 209 |
bz = (b * 3) + 2; |
| 210 |
|
| 211 |
if (moved[a] || moved[b]){ |
| 212 |
atoms[a]->getPos(posA); |
| 213 |
atoms[b]->getPos(posB); |
| 214 |
|
| 215 |
for (j = 0; j < 3; j++) |
| 216 |
pab[j] = posA[j] - posB[j]; |
| 217 |
|
| 218 |
//periodic boundary condition |
| 219 |
|
| 220 |
info->wrapVector(pab); |
| 221 |
|
| 222 |
pabsq = pab[0] * pab[0] + pab[1] * pab[1] + pab[2] * pab[2]; |
| 223 |
|
| 224 |
rabsq = constrainedDsqr[i]; |
| 225 |
diffsq = rabsq - pabsq; |
| 226 |
|
| 227 |
// the original rattle code from alan tidesley |
| 228 |
if (fabs(diffsq) > (tol * rabsq * 2)){ |
| 229 |
rab[0] = oldPos[ax] - oldPos[bx]; |
| 230 |
rab[1] = oldPos[ay] - oldPos[by]; |
| 231 |
rab[2] = oldPos[az] - oldPos[bz]; |
| 232 |
|
| 233 |
info->wrapVector(rab); |
| 234 |
|
| 235 |
rpab = rab[0] * pab[0] + rab[1] * pab[1] + rab[2] * pab[2]; |
| 236 |
|
| 237 |
rpabsq = rpab * rpab; |
| 238 |
|
| 239 |
|
| 240 |
if (rpabsq < (rabsq * -diffsq)){ |
| 241 |
#ifdef IS_MPI |
| 242 |
a = atoms[a]->getGlobalIndex(); |
| 243 |
b = atoms[b]->getGlobalIndex(); |
| 244 |
#endif //is_mpi |
| 245 |
//cerr << "Waring: constraint failure" << endl; |
| 246 |
gab = sqrt(rabsq/pabsq); |
| 247 |
rab[0] = (posA[0] - posB[0])*gab; |
| 248 |
rab[1]= (posA[1] - posB[1])*gab; |
| 249 |
rab[2] = (posA[2] - posB[2])*gab; |
| 250 |
|
| 251 |
info->wrapVector(rab); |
| 252 |
|
| 253 |
rpab = rab[0] * pab[0] + rab[1] * pab[1] + rab[2] * pab[2]; |
| 254 |
|
| 255 |
} |
| 256 |
|
| 257 |
//rma = 1.0 / atoms[a]->getMass(); |
| 258 |
//rmb = 1.0 / atoms[b]->getMass(); |
| 259 |
rma = 1.0; |
| 260 |
rmb =1.0; |
| 261 |
|
| 262 |
gab = diffsq / (2.0 * (rma + rmb) * rpab); |
| 263 |
|
| 264 |
dx = rab[0] * gab; |
| 265 |
dy = rab[1] * gab; |
| 266 |
dz = rab[2] * gab; |
| 267 |
|
| 268 |
posA[0] += rma * dx; |
| 269 |
posA[1] += rma * dy; |
| 270 |
posA[2] += rma * dz; |
| 271 |
|
| 272 |
atoms[a]->setPos(posA); |
| 273 |
|
| 274 |
posB[0] -= rmb * dx; |
| 275 |
posB[1] -= rmb * dy; |
| 276 |
posB[2] -= rmb * dz; |
| 277 |
|
| 278 |
atoms[b]->setPos(posB); |
| 279 |
|
| 280 |
moving[a] = 1; |
| 281 |
moving[b] = 1; |
| 282 |
done = 0; |
| 283 |
} |
| 284 |
} |
| 285 |
} |
| 286 |
|
| 287 |
for (i = 0; i < nAtoms; i++){ |
| 288 |
moved[i] = moving[i]; |
| 289 |
moving[i] = 0; |
| 290 |
} |
| 291 |
|
| 292 |
iteration++; |
| 293 |
} |
| 294 |
|
| 295 |
if (!done){ |
| 296 |
cerr << "Waring: can not constraint within maxIteration" << endl; |
| 297 |
return -1; |
| 298 |
} |
| 299 |
else |
| 300 |
return 1; |
| 301 |
} |
| 302 |
|
| 303 |
|
| 304 |
//remove constraint force along the bond direction |
| 305 |
int OOPSEMinimizer::shakeF(){ |
| 306 |
int i, j; |
| 307 |
int done; |
| 308 |
double posA[3], posB[3]; |
| 309 |
double frcA[3], frcB[3]; |
| 310 |
double rab[3], fpab[3]; |
| 311 |
int a, b, ax, ay, az, bx, by, bz; |
| 312 |
double rma, rmb; |
| 313 |
double rvab; |
| 314 |
double gab; |
| 315 |
double rabsq; |
| 316 |
double rfab; |
| 317 |
int iteration; |
| 318 |
|
| 319 |
for (i = 0; i < nAtoms; i++){ |
| 320 |
moving[i] = 0; |
| 321 |
moved[i] = 1; |
| 322 |
} |
| 323 |
|
| 324 |
done = 0; |
| 325 |
iteration = 0; |
| 326 |
while (!done && (iteration < maxIteration)){ |
| 327 |
done = 1; |
| 328 |
|
| 329 |
for (i = 0; i < nConstrained; i++){ |
| 330 |
a = constrainedA[i]; |
| 331 |
b = constrainedB[i]; |
| 332 |
|
| 333 |
ax = (a * 3) + 0; |
| 334 |
ay = (a * 3) + 1; |
| 335 |
az = (a * 3) + 2; |
| 336 |
|
| 337 |
bx = (b * 3) + 0; |
| 338 |
by = (b * 3) + 1; |
| 339 |
bz = (b * 3) + 2; |
| 340 |
|
| 341 |
if (moved[a] || moved[b]){ |
| 342 |
|
| 343 |
atoms[a]->getPos(posA); |
| 344 |
atoms[b]->getPos(posB); |
| 345 |
|
| 346 |
for (j = 0; j < 3; j++) |
| 347 |
rab[j] = posA[j] - posB[j]; |
| 348 |
|
| 349 |
info->wrapVector(rab); |
| 350 |
|
| 351 |
atoms[a]->getFrc(frcA); |
| 352 |
atoms[b]->getFrc(frcB); |
| 353 |
|
| 354 |
//rma = 1.0 / atoms[a]->getMass(); |
| 355 |
//rmb = 1.0 / atoms[b]->getMass(); |
| 356 |
rma = 1.0; |
| 357 |
rmb = 1.0; |
| 358 |
|
| 359 |
|
| 360 |
fpab[0] = frcA[0] * rma - frcB[0] * rmb; |
| 361 |
fpab[1] = frcA[1] * rma - frcB[1] * rmb; |
| 362 |
fpab[2] = frcA[2] * rma - frcB[2] * rmb; |
| 363 |
|
| 364 |
|
| 365 |
gab=fpab[0] * fpab[0] + fpab[1] * fpab[1] + fpab[2] * fpab[2]; |
| 366 |
|
| 367 |
if (gab < 1.0) |
| 368 |
gab = 1.0; |
| 369 |
|
| 370 |
rabsq = rab[0] * rab[0] + rab[1] * rab[1] + rab[2] * rab[2]; |
| 371 |
rfab = rab[0] * fpab[0] + rab[1] * fpab[1] + rab[2] * fpab[2]; |
| 372 |
|
| 373 |
if (fabs(rfab) > sqrt(rabsq*gab) * 0.00001){ |
| 374 |
|
| 375 |
gab = -rfab /(rabsq*(rma + rmb)); |
| 376 |
|
| 377 |
frcA[0] = rab[0] * gab; |
| 378 |
frcA[1] = rab[1] * gab; |
| 379 |
frcA[2] = rab[2] * gab; |
| 380 |
|
| 381 |
atoms[a]->addFrc(frcA); |
| 382 |
|
| 383 |
|
| 384 |
frcB[0] = -rab[0] * gab; |
| 385 |
frcB[1] = -rab[1] * gab; |
| 386 |
frcB[2] = -rab[2] * gab; |
| 387 |
|
| 388 |
atoms[b]->addFrc(frcB); |
| 389 |
|
| 390 |
moving[a] = 1; |
| 391 |
moving[b] = 1; |
| 392 |
done = 0; |
| 393 |
} |
| 394 |
} |
| 395 |
} |
| 396 |
|
| 397 |
for (i = 0; i < nAtoms; i++){ |
| 398 |
moved[i] = moving[i]; |
| 399 |
moving[i] = 0; |
| 400 |
} |
| 401 |
|
| 402 |
iteration++; |
| 403 |
} |
| 404 |
|
| 405 |
if (!done){ |
| 406 |
cerr << "Waring: can not constraint within maxIteration" << endl; |
| 407 |
return -1; |
| 408 |
} |
| 409 |
else |
| 410 |
return 1; |
| 411 |
} |
| 412 |
|
| 413 |
*/ |
| 414 |
|
| 415 |
//calculate the value of object function |
| 416 |
void OOPSEMinimizer::calcF(){ |
| 417 |
calcEnergyGradient(curX, curG, curF, egEvalStatus); |
| 418 |
} |
| 419 |
|
| 420 |
void OOPSEMinimizer::calcF(vector<double>& x, double&f, int& status){ |
| 421 |
vector<double> tempG; |
| 422 |
tempG.resize(x.size()); |
| 423 |
|
| 424 |
calcEnergyGradient(x, tempG, f, status); |
| 425 |
} |
| 426 |
|
| 427 |
//calculate the gradient |
| 428 |
void OOPSEMinimizer::calcG(){ |
| 429 |
calcEnergyGradient(curX, curG, curF, egEvalStatus); |
| 430 |
} |
| 431 |
|
| 432 |
void OOPSEMinimizer::calcG(vector<double>& x, vector<double>& g, double& f, int& status){ |
| 433 |
calcEnergyGradient(x, g, f, status); |
| 434 |
} |
| 435 |
|
| 436 |
void OOPSEMinimizer::calcDim(){ |
| 437 |
DirectionalAtom* dAtom; |
| 438 |
|
| 439 |
ndim = 0; |
| 440 |
|
| 441 |
for(int i = 0; i < integrableObjects.size(); i++){ |
| 442 |
ndim += 3; |
| 443 |
if (integrableObjects[i]->isDirectional()) |
| 444 |
ndim += 3; |
| 445 |
} |
| 446 |
} |
| 447 |
|
| 448 |
void OOPSEMinimizer::setX(vector < double > & x){ |
| 449 |
|
| 450 |
if (x.size() != ndim && bVerbose){ |
| 451 |
//sprintf(painCave.errMsg, |
| 452 |
// "OOPSEMinimizer Error: dimesion of x and curX does not match\n"); |
| 453 |
// painCave.isFatal = 1; |
| 454 |
// simError(); |
| 455 |
} |
| 456 |
|
| 457 |
curX = x; |
| 458 |
} |
| 459 |
|
| 460 |
void OOPSEMinimizer::setG(vector < double > & g){ |
| 461 |
|
| 462 |
if (g.size() != ndim && bVerbose){ |
| 463 |
//sprintf(painCave.errMsg, |
| 464 |
// "OOPSEMinimizer Error: dimesion of g and curG does not match\n"); |
| 465 |
// painCave.isFatal = 1; |
| 466 |
//simError(); |
| 467 |
} |
| 468 |
|
| 469 |
curG = g; |
| 470 |
} |
| 471 |
|
| 472 |
void OOPSEMinimizer::writeOut(vector<double>& x, double iter){ |
| 473 |
|
| 474 |
setX(x); |
| 475 |
|
| 476 |
calcG(); |
| 477 |
|
| 478 |
dumpOut->writeDump(iter); |
| 479 |
statOut->writeStat(iter); |
| 480 |
} |
| 481 |
|
| 482 |
|
| 483 |
void OOPSEMinimizer::printMinimizerInfo(){ |
| 484 |
cout << "--------------------------------------------------------------------" << endl; |
| 485 |
cout << minimizerName << endl; |
| 486 |
cout << "minimization parameter set" << endl; |
| 487 |
cout << "function tolerance = " << paramSet->getFTol() << endl; |
| 488 |
cout << "gradient tolerance = " << paramSet->getGTol() << endl; |
| 489 |
cout << "step tolerance = "<< paramSet->getFTol() << endl; |
| 490 |
cout << "absolute gradient tolerance = " << endl; |
| 491 |
cout << "max iteration = " << paramSet->getMaxIteration() << endl; |
| 492 |
cout << "max line search iteration = " << paramSet->getLineSearchMaxIteration() <<endl; |
| 493 |
cout << "shake algorithm = " << bShake << endl; |
| 494 |
cout << "--------------------------------------------------------------------" << endl; |
| 495 |
|
| 496 |
} |
| 497 |
|
| 498 |
/** |
| 499 |
* In thoery, we need to find the minimum along the search direction |
| 500 |
* However, function evaluation is too expensive. |
| 501 |
* At the very begining of the problem, we check the search direction and make sure |
| 502 |
* it is a descent direction |
| 503 |
* we will compare the energy of two end points, |
| 504 |
* if the right end point has lower energy, we just take it |
| 505 |
* |
| 506 |
* |
| 507 |
* |
| 508 |
*/ |
| 509 |
|
| 510 |
int OOPSEMinimizer::doLineSearch(vector<double>& direction, double stepSize){ |
| 511 |
vector<double> xa; |
| 512 |
vector<double> xb; |
| 513 |
vector<double> xc; |
| 514 |
vector<double> ga; |
| 515 |
vector<double> gb; |
| 516 |
vector<double> gc; |
| 517 |
double fa; |
| 518 |
double fb; |
| 519 |
double fc; |
| 520 |
double a; |
| 521 |
double b; |
| 522 |
double c; |
| 523 |
int status; |
| 524 |
double initSlope; |
| 525 |
double slopeA; |
| 526 |
double slopeB; |
| 527 |
double slopeC; |
| 528 |
bool foundLower; |
| 529 |
int iter; |
| 530 |
int maxLSIter; |
| 531 |
double mu; |
| 532 |
double eta; |
| 533 |
double ftol; |
| 534 |
double lsTol; |
| 535 |
|
| 536 |
xa.resize(ndim); |
| 537 |
xb.resize(ndim); |
| 538 |
xc.resize(ndim); |
| 539 |
|
| 540 |
ga.resize(ndim); |
| 541 |
gb.resize(ndim); |
| 542 |
gc.resize(ndim); |
| 543 |
|
| 544 |
a = 0.0; |
| 545 |
fa = curF; |
| 546 |
xa = curX; |
| 547 |
ga = curG; |
| 548 |
c = a + stepSize; |
| 549 |
ftol = paramSet->getFTol(); |
| 550 |
lsTol = paramSet->getLineSearchTol(); |
| 551 |
|
| 552 |
//calculate the derivative at a = 0 |
| 553 |
slopeA = 0; |
| 554 |
for (size_t i = 0; i < ndim; i++) |
| 555 |
slopeA += curG[i]*direction[i]; |
| 556 |
|
| 557 |
initSlope = slopeA; |
| 558 |
|
| 559 |
// if going uphill, use negative gradient as searching direction |
| 560 |
if (slopeA > 0) { |
| 561 |
|
| 562 |
if (bVerbose){ |
| 563 |
cout << "LineSearch Warning: initial searching direction is not a descent searching direction, " |
| 564 |
<< " use negative gradient instead. Therefore, finding a smaller vaule of function " |
| 565 |
<< " is guaranteed" |
| 566 |
<< endl; |
| 567 |
} |
| 568 |
|
| 569 |
for (size_t i = 0; i < ndim; i++) |
| 570 |
direction[i] = -curG[i]; |
| 571 |
|
| 572 |
for (size_t i = 0; i < ndim; i++) |
| 573 |
slopeA += curG[i]*direction[i]; |
| 574 |
|
| 575 |
initSlope = slopeA; |
| 576 |
} |
| 577 |
|
| 578 |
// Take a trial step |
| 579 |
for(size_t i = 0; i < ndim; i++) |
| 580 |
xc[i] = curX[i] + direction[i] * c; |
| 581 |
|
| 582 |
calcG(xc, gc, fc, status); |
| 583 |
|
| 584 |
if (status < 0){ |
| 585 |
if (bVerbose) |
| 586 |
cerr << "Function Evaluation Error" << endl; |
| 587 |
} |
| 588 |
|
| 589 |
//calculate the derivative at c |
| 590 |
slopeC = 0; |
| 591 |
for (size_t i = 0; i < ndim; i++) |
| 592 |
slopeC += gc[i]*direction[i]; |
| 593 |
|
| 594 |
// found a lower point |
| 595 |
if (fc < fa) { |
| 596 |
curX = xc; |
| 597 |
curG = gc; |
| 598 |
curF = fc; |
| 599 |
return LS_SUCCEED; |
| 600 |
} |
| 601 |
else { |
| 602 |
|
| 603 |
if (slopeC > 0) |
| 604 |
stepSize *= 0.618034; |
| 605 |
} |
| 606 |
|
| 607 |
maxLSIter = paramSet->getLineSearchMaxIteration(); |
| 608 |
|
| 609 |
iter = 0; |
| 610 |
|
| 611 |
do { |
| 612 |
// Select a new trial point. |
| 613 |
// If the derivatives at points a & c have different sign we use cubic interpolate |
| 614 |
//if (slopeC > 0){ |
| 615 |
eta = 3 *(fa -fc) /(c - a) + slopeA + slopeC; |
| 616 |
mu = sqrt(eta * eta - slopeA * slopeC); |
| 617 |
b = a + (c - a) * (1 - (slopeC + mu - eta) /(slopeC - slopeA + 2 * mu)); |
| 618 |
|
| 619 |
if (b < lsTol){ |
| 620 |
if (bVerbose) |
| 621 |
cout << "stepSize is less than line search tolerance" << endl; |
| 622 |
break; |
| 623 |
} |
| 624 |
//} |
| 625 |
|
| 626 |
// Take a trial step to this new point - new coords in xb |
| 627 |
for(size_t i = 0; i < ndim; i++) |
| 628 |
xb[i] = curX[i] + direction[i] * b; |
| 629 |
|
| 630 |
//function evaluation |
| 631 |
calcG(xb, gb, fb, status); |
| 632 |
|
| 633 |
if (status < 0){ |
| 634 |
if (bVerbose) |
| 635 |
cerr << "Function Evaluation Error" << endl; |
| 636 |
} |
| 637 |
|
| 638 |
//calculate the derivative at c |
| 639 |
slopeB = 0; |
| 640 |
for (size_t i = 0; i < ndim; i++) |
| 641 |
slopeB += gb[i]*direction[i]; |
| 642 |
|
| 643 |
//Amijo Rule to stop the line search |
| 644 |
if (fb <= curF + initSlope * ftol * b) { |
| 645 |
curF = fb; |
| 646 |
curX = xb; |
| 647 |
curG = gb; |
| 648 |
return LS_SUCCEED; |
| 649 |
} |
| 650 |
|
| 651 |
if (slopeB <0 && fb < fa) { |
| 652 |
//replace a by b |
| 653 |
fa = fb; |
| 654 |
a = b; |
| 655 |
slopeA = slopeB; |
| 656 |
|
| 657 |
// swap coord a/b |
| 658 |
std::swap(xa, xb); |
| 659 |
std::swap(ga, gb); |
| 660 |
} |
| 661 |
else { |
| 662 |
//replace c by b |
| 663 |
fc = fb; |
| 664 |
c = b; |
| 665 |
slopeC = slopeB; |
| 666 |
|
| 667 |
// swap coord b/c |
| 668 |
std::swap(gb, gc); |
| 669 |
std::swap(xb, xc); |
| 670 |
} |
| 671 |
|
| 672 |
|
| 673 |
iter++; |
| 674 |
} while((fb > fa || fb > fc) && (iter < maxLSIter)); |
| 675 |
|
| 676 |
if(fb < curF || iter >= maxLSIter) { |
| 677 |
//could not find a lower value, we might just go uphill. |
| 678 |
return LS_ERROR; |
| 679 |
} |
| 680 |
|
| 681 |
//select the end point |
| 682 |
|
| 683 |
if (fa <= fc) { |
| 684 |
curX = xa; |
| 685 |
curG = ga; |
| 686 |
curF = fa; |
| 687 |
} |
| 688 |
else { |
| 689 |
curX = xc; |
| 690 |
curG = gc; |
| 691 |
curF = fc; |
| 692 |
} |
| 693 |
|
| 694 |
return LS_SUCCEED; |
| 695 |
|
| 696 |
} |
| 697 |
|
| 698 |
void OOPSEMinimizer::minimize(){ |
| 699 |
|
| 700 |
int convgStatus; |
| 701 |
int stepStatus; |
| 702 |
int maxIter; |
| 703 |
//int resetFrq; |
| 704 |
//int nextResetIter; |
| 705 |
int writeFrq; |
| 706 |
int nextWriteIter; |
| 707 |
|
| 708 |
if (bVerbose) |
| 709 |
printMinimizerInfo(); |
| 710 |
|
| 711 |
dumpOut = new DumpWriter(info); |
| 712 |
statOut = new StatWriter(info); |
| 713 |
|
| 714 |
init(); |
| 715 |
|
| 716 |
//resetFrq = paramSet->getResetFrq(); |
| 717 |
//nextResetIter = resetFrq; |
| 718 |
|
| 719 |
writeFrq = paramSet->getWriteFrq(); |
| 720 |
nextWriteIter = writeFrq; |
| 721 |
|
| 722 |
maxIter = paramSet->getMaxIteration(); |
| 723 |
|
| 724 |
for (curIter = 1; curIter <= maxIter; curIter++){ |
| 725 |
|
| 726 |
stepStatus = step(); |
| 727 |
|
| 728 |
if (bShake) |
| 729 |
shakeAlgo->doPreConstraint(); |
| 730 |
|
| 731 |
if (stepStatus < 0){ |
| 732 |
saveResult(); |
| 733 |
minStatus = MIN_LSERROR; |
| 734 |
cerr << "OOPSEMinimizer Error: line search error, please try a small stepsize" << endl; |
| 735 |
return; |
| 736 |
} |
| 737 |
|
| 738 |
if (curIter == nextWriteIter){ |
| 739 |
nextWriteIter += writeFrq; |
| 740 |
writeOut(curX, curIter); |
| 741 |
} |
| 742 |
|
| 743 |
convgStatus = checkConvg(); |
| 744 |
|
| 745 |
if (convgStatus > 0){ |
| 746 |
saveResult(); |
| 747 |
minStatus = MIN_CONVERGE; |
| 748 |
return; |
| 749 |
} |
| 750 |
|
| 751 |
prepareStep(); |
| 752 |
|
| 753 |
} |
| 754 |
|
| 755 |
|
| 756 |
if (bVerbose) { |
| 757 |
cout << "OOPSEMinimizer Warning: " |
| 758 |
<< minimizerName << " algorithm did not converge within " |
| 759 |
<< maxIter << " iteration" << endl; |
| 760 |
} |
| 761 |
minStatus = MIN_MAXITER; |
| 762 |
saveResult(); |
| 763 |
|
| 764 |
} |
