1 |
tim |
1064 |
#include <math.h> |
2 |
|
|
#include "OOPSEMinimizer.hpp" |
3 |
|
|
|
4 |
|
|
OOPSEMinimizer::OOPSEMinimizer( SimInfo *theInfo, ForceFields* the_ff , |
5 |
|
|
MinimizerParameterSet * param) |
6 |
|
|
:RealIntegrator(theInfo, the_ff), bVerbose(false), bShake(true){ |
7 |
tim |
1144 |
dumpOut = NULL; |
8 |
|
|
statOut = NULL; |
9 |
tim |
1064 |
|
10 |
|
|
tStats = new Thermo(info); |
11 |
tim |
1144 |
|
12 |
tim |
1064 |
|
13 |
|
|
paramSet = param; |
14 |
|
|
|
15 |
|
|
calcDim(); |
16 |
|
|
|
17 |
|
|
curX = getCoor(); |
18 |
|
|
curG.resize(ndim); |
19 |
|
|
|
20 |
tim |
1074 |
preMove(); |
21 |
tim |
1064 |
} |
22 |
|
|
|
23 |
|
|
OOPSEMinimizer::~OOPSEMinimizer(){ |
24 |
|
|
delete tStats; |
25 |
tim |
1144 |
if(dumpOut) |
26 |
|
|
delete dumpOut; |
27 |
|
|
if(statOut) |
28 |
|
|
delete statOut; |
29 |
tim |
1064 |
delete paramSet; |
30 |
|
|
} |
31 |
|
|
|
32 |
|
|
void OOPSEMinimizer::calcEnergyGradient(vector<double>& x, vector<double>& grad, |
33 |
|
|
double& energy, int& status){ |
34 |
|
|
|
35 |
|
|
DirectionalAtom* dAtom; |
36 |
|
|
int index; |
37 |
|
|
double force[3]; |
38 |
|
|
double dAtomGrad[6]; |
39 |
|
|
int shakeStatus; |
40 |
|
|
|
41 |
|
|
setCoor(x); |
42 |
|
|
|
43 |
|
|
if (nConstrained && bShake){ |
44 |
|
|
shakeStatus = shakeR(); |
45 |
|
|
} |
46 |
|
|
|
47 |
|
|
calcForce(1, 1); |
48 |
|
|
|
49 |
|
|
if (nConstrained && bShake){ |
50 |
|
|
shakeStatus |= shakeF(); |
51 |
|
|
} |
52 |
|
|
|
53 |
|
|
x = getCoor(); |
54 |
|
|
|
55 |
|
|
|
56 |
|
|
index = 0; |
57 |
|
|
|
58 |
gezelter |
1097 |
for(int i = 0; i < integrableObjects.size(); i++){ |
59 |
tim |
1064 |
|
60 |
gezelter |
1097 |
if (integrableObjects[i]->isDirectional()) { |
61 |
tim |
1064 |
|
62 |
gezelter |
1097 |
integrableObjects[i]->getGrad(dAtomGrad); |
63 |
|
|
|
64 |
tim |
1064 |
//gradient is equal to -f |
65 |
|
|
grad[index++] = -dAtomGrad[0]; |
66 |
|
|
grad[index++] = -dAtomGrad[1]; |
67 |
|
|
grad[index++] = -dAtomGrad[2]; |
68 |
|
|
grad[index++] = -dAtomGrad[3]; |
69 |
|
|
grad[index++] = -dAtomGrad[4]; |
70 |
|
|
grad[index++] = -dAtomGrad[5]; |
71 |
|
|
|
72 |
|
|
} |
73 |
|
|
else{ |
74 |
gezelter |
1097 |
integrableObjects[i]->getFrc(force); |
75 |
tim |
1064 |
|
76 |
|
|
grad[index++] = -force[0]; |
77 |
|
|
grad[index++] = -force[1]; |
78 |
|
|
grad[index++] = -force[2]; |
79 |
|
|
|
80 |
|
|
} |
81 |
|
|
|
82 |
|
|
} |
83 |
|
|
|
84 |
|
|
energy = tStats->getPotential(); |
85 |
|
|
|
86 |
|
|
status = shakeStatus; |
87 |
|
|
} |
88 |
|
|
|
89 |
|
|
/** |
90 |
|
|
* |
91 |
|
|
*/ |
92 |
|
|
|
93 |
|
|
void OOPSEMinimizer::setCoor(vector<double>& x){ |
94 |
|
|
|
95 |
|
|
DirectionalAtom* dAtom; |
96 |
|
|
int index; |
97 |
|
|
double position[3]; |
98 |
|
|
double eulerAngle[3]; |
99 |
|
|
|
100 |
|
|
|
101 |
|
|
index = 0; |
102 |
|
|
|
103 |
gezelter |
1097 |
for(int i = 0; i < integrableObjects.size(); i++){ |
104 |
tim |
1064 |
|
105 |
|
|
position[0] = x[index++]; |
106 |
|
|
position[1] = x[index++]; |
107 |
|
|
position[2] = x[index++]; |
108 |
|
|
|
109 |
gezelter |
1097 |
integrableObjects[i]->setPos(position); |
110 |
tim |
1064 |
|
111 |
gezelter |
1097 |
if (integrableObjects[i]->isDirectional()){ |
112 |
tim |
1064 |
|
113 |
|
|
eulerAngle[0] = x[index++]; |
114 |
|
|
eulerAngle[1] = x[index++]; |
115 |
|
|
eulerAngle[2] = x[index++]; |
116 |
|
|
|
117 |
gezelter |
1097 |
integrableObjects[i]->setEuler(eulerAngle[0], |
118 |
|
|
eulerAngle[1], |
119 |
|
|
eulerAngle[2]); |
120 |
tim |
1064 |
|
121 |
|
|
} |
122 |
|
|
|
123 |
|
|
} |
124 |
|
|
|
125 |
|
|
} |
126 |
|
|
|
127 |
|
|
/** |
128 |
|
|
* |
129 |
|
|
*/ |
130 |
|
|
vector<double> OOPSEMinimizer::getCoor(){ |
131 |
|
|
|
132 |
|
|
DirectionalAtom* dAtom; |
133 |
|
|
int index; |
134 |
|
|
double position[3]; |
135 |
|
|
double eulerAngle[3]; |
136 |
|
|
vector<double> x; |
137 |
|
|
|
138 |
|
|
x.resize(getDim()); |
139 |
|
|
|
140 |
|
|
index = 0; |
141 |
|
|
|
142 |
gezelter |
1097 |
for(int i = 0; i < integrableObjects.size(); i++){ |
143 |
|
|
integrableObjects[i]->getPos(position); |
144 |
tim |
1064 |
|
145 |
|
|
x[index++] = position[0]; |
146 |
|
|
x[index++] = position[1]; |
147 |
|
|
x[index++] = position[2]; |
148 |
|
|
|
149 |
gezelter |
1097 |
if (integrableObjects[i]->isDirectional()){ |
150 |
tim |
1064 |
|
151 |
gezelter |
1097 |
integrableObjects[i]->getEulerAngles(eulerAngle); |
152 |
|
|
|
153 |
tim |
1064 |
x[index++] = eulerAngle[0]; |
154 |
|
|
x[index++] = eulerAngle[1]; |
155 |
|
|
x[index++] = eulerAngle[2]; |
156 |
|
|
|
157 |
|
|
} |
158 |
|
|
|
159 |
|
|
} |
160 |
|
|
|
161 |
|
|
return x; |
162 |
|
|
|
163 |
|
|
} |
164 |
|
|
|
165 |
|
|
int OOPSEMinimizer::shakeR(){ |
166 |
|
|
int i, j; |
167 |
|
|
int done; |
168 |
|
|
double posA[3], posB[3]; |
169 |
|
|
double velA[3], velB[3]; |
170 |
|
|
double pab[3]; |
171 |
|
|
double rab[3]; |
172 |
|
|
int a, b, ax, ay, az, bx, by, bz; |
173 |
|
|
double rma, rmb; |
174 |
|
|
double dx, dy, dz; |
175 |
|
|
double rpab; |
176 |
|
|
double rabsq, pabsq, rpabsq; |
177 |
|
|
double diffsq; |
178 |
|
|
double gab; |
179 |
|
|
int iteration; |
180 |
|
|
|
181 |
|
|
for (i = 0; i < nAtoms; i++){ |
182 |
|
|
moving[i] = 0; |
183 |
|
|
moved[i] = 1; |
184 |
|
|
} |
185 |
|
|
|
186 |
|
|
iteration = 0; |
187 |
|
|
done = 0; |
188 |
|
|
while (!done && (iteration < maxIteration)){ |
189 |
|
|
done = 1; |
190 |
|
|
for (i = 0; i < nConstrained; i++){ |
191 |
|
|
a = constrainedA[i]; |
192 |
|
|
b = constrainedB[i]; |
193 |
|
|
|
194 |
|
|
ax = (a * 3) + 0; |
195 |
|
|
ay = (a * 3) + 1; |
196 |
|
|
az = (a * 3) + 2; |
197 |
|
|
|
198 |
|
|
bx = (b * 3) + 0; |
199 |
|
|
by = (b * 3) + 1; |
200 |
|
|
bz = (b * 3) + 2; |
201 |
|
|
|
202 |
|
|
if (moved[a] || moved[b]){ |
203 |
|
|
atoms[a]->getPos(posA); |
204 |
|
|
atoms[b]->getPos(posB); |
205 |
|
|
|
206 |
|
|
for (j = 0; j < 3; j++) |
207 |
|
|
pab[j] = posA[j] - posB[j]; |
208 |
|
|
|
209 |
|
|
//periodic boundary condition |
210 |
|
|
|
211 |
|
|
info->wrapVector(pab); |
212 |
|
|
|
213 |
|
|
pabsq = pab[0] * pab[0] + pab[1] * pab[1] + pab[2] * pab[2]; |
214 |
|
|
|
215 |
|
|
rabsq = constrainedDsqr[i]; |
216 |
|
|
diffsq = rabsq - pabsq; |
217 |
|
|
|
218 |
|
|
// the original rattle code from alan tidesley |
219 |
|
|
if (fabs(diffsq) > (tol * rabsq * 2)){ |
220 |
|
|
rab[0] = oldPos[ax] - oldPos[bx]; |
221 |
|
|
rab[1] = oldPos[ay] - oldPos[by]; |
222 |
|
|
rab[2] = oldPos[az] - oldPos[bz]; |
223 |
|
|
|
224 |
|
|
info->wrapVector(rab); |
225 |
|
|
|
226 |
|
|
rpab = rab[0] * pab[0] + rab[1] * pab[1] + rab[2] * pab[2]; |
227 |
|
|
|
228 |
|
|
rpabsq = rpab * rpab; |
229 |
|
|
|
230 |
|
|
|
231 |
|
|
if (rpabsq < (rabsq * -diffsq)){ |
232 |
|
|
#ifdef IS_MPI |
233 |
|
|
a = atoms[a]->getGlobalIndex(); |
234 |
|
|
b = atoms[b]->getGlobalIndex(); |
235 |
|
|
#endif //is_mpi |
236 |
|
|
//cerr << "Waring: constraint failure" << endl; |
237 |
|
|
gab = sqrt(rabsq/pabsq); |
238 |
|
|
rab[0] = (posA[0] - posB[0])*gab; |
239 |
|
|
rab[1]= (posA[1] - posB[1])*gab; |
240 |
|
|
rab[2] = (posA[2] - posB[2])*gab; |
241 |
|
|
|
242 |
|
|
info->wrapVector(rab); |
243 |
|
|
|
244 |
|
|
rpab = rab[0] * pab[0] + rab[1] * pab[1] + rab[2] * pab[2]; |
245 |
|
|
|
246 |
|
|
} |
247 |
|
|
|
248 |
|
|
//rma = 1.0 / atoms[a]->getMass(); |
249 |
|
|
//rmb = 1.0 / atoms[b]->getMass(); |
250 |
|
|
rma = 1.0; |
251 |
|
|
rmb =1.0; |
252 |
|
|
|
253 |
|
|
gab = diffsq / (2.0 * (rma + rmb) * rpab); |
254 |
|
|
|
255 |
|
|
dx = rab[0] * gab; |
256 |
|
|
dy = rab[1] * gab; |
257 |
|
|
dz = rab[2] * gab; |
258 |
|
|
|
259 |
|
|
posA[0] += rma * dx; |
260 |
|
|
posA[1] += rma * dy; |
261 |
|
|
posA[2] += rma * dz; |
262 |
|
|
|
263 |
|
|
atoms[a]->setPos(posA); |
264 |
|
|
|
265 |
|
|
posB[0] -= rmb * dx; |
266 |
|
|
posB[1] -= rmb * dy; |
267 |
|
|
posB[2] -= rmb * dz; |
268 |
|
|
|
269 |
|
|
atoms[b]->setPos(posB); |
270 |
|
|
|
271 |
|
|
moving[a] = 1; |
272 |
|
|
moving[b] = 1; |
273 |
|
|
done = 0; |
274 |
|
|
} |
275 |
|
|
} |
276 |
|
|
} |
277 |
|
|
|
278 |
|
|
for (i = 0; i < nAtoms; i++){ |
279 |
|
|
moved[i] = moving[i]; |
280 |
|
|
moving[i] = 0; |
281 |
|
|
} |
282 |
|
|
|
283 |
|
|
iteration++; |
284 |
|
|
} |
285 |
|
|
|
286 |
|
|
if (!done){ |
287 |
|
|
cerr << "Waring: can not constraint within maxIteration" << endl; |
288 |
|
|
return -1; |
289 |
|
|
} |
290 |
|
|
else |
291 |
|
|
return 1; |
292 |
|
|
} |
293 |
|
|
|
294 |
|
|
|
295 |
|
|
//remove constraint force along the bond direction |
296 |
|
|
int OOPSEMinimizer::shakeF(){ |
297 |
|
|
int i, j; |
298 |
|
|
int done; |
299 |
|
|
double posA[3], posB[3]; |
300 |
|
|
double frcA[3], frcB[3]; |
301 |
|
|
double rab[3], fpab[3]; |
302 |
|
|
int a, b, ax, ay, az, bx, by, bz; |
303 |
|
|
double rma, rmb; |
304 |
|
|
double rvab; |
305 |
|
|
double gab; |
306 |
|
|
double rabsq; |
307 |
|
|
double rfab; |
308 |
|
|
int iteration; |
309 |
|
|
|
310 |
|
|
for (i = 0; i < nAtoms; i++){ |
311 |
|
|
moving[i] = 0; |
312 |
|
|
moved[i] = 1; |
313 |
|
|
} |
314 |
|
|
|
315 |
|
|
done = 0; |
316 |
|
|
iteration = 0; |
317 |
|
|
while (!done && (iteration < maxIteration)){ |
318 |
|
|
done = 1; |
319 |
|
|
|
320 |
|
|
for (i = 0; i < nConstrained; i++){ |
321 |
|
|
a = constrainedA[i]; |
322 |
|
|
b = constrainedB[i]; |
323 |
|
|
|
324 |
|
|
ax = (a * 3) + 0; |
325 |
|
|
ay = (a * 3) + 1; |
326 |
|
|
az = (a * 3) + 2; |
327 |
|
|
|
328 |
|
|
bx = (b * 3) + 0; |
329 |
|
|
by = (b * 3) + 1; |
330 |
|
|
bz = (b * 3) + 2; |
331 |
|
|
|
332 |
|
|
if (moved[a] || moved[b]){ |
333 |
|
|
|
334 |
|
|
atoms[a]->getPos(posA); |
335 |
|
|
atoms[b]->getPos(posB); |
336 |
|
|
|
337 |
|
|
for (j = 0; j < 3; j++) |
338 |
|
|
rab[j] = posA[j] - posB[j]; |
339 |
|
|
|
340 |
|
|
info->wrapVector(rab); |
341 |
|
|
|
342 |
|
|
atoms[a]->getFrc(frcA); |
343 |
|
|
atoms[b]->getFrc(frcB); |
344 |
|
|
|
345 |
|
|
//rma = 1.0 / atoms[a]->getMass(); |
346 |
|
|
//rmb = 1.0 / atoms[b]->getMass(); |
347 |
|
|
rma = 1.0; |
348 |
|
|
rmb = 1.0; |
349 |
|
|
|
350 |
|
|
|
351 |
|
|
fpab[0] = frcA[0] * rma - frcB[0] * rmb; |
352 |
|
|
fpab[1] = frcA[1] * rma - frcB[1] * rmb; |
353 |
|
|
fpab[2] = frcA[2] * rma - frcB[2] * rmb; |
354 |
|
|
|
355 |
|
|
|
356 |
|
|
gab=fpab[0] * fpab[0] + fpab[1] * fpab[1] + fpab[2] * fpab[2]; |
357 |
|
|
|
358 |
|
|
if (gab < 1.0) |
359 |
|
|
gab = 1.0; |
360 |
|
|
|
361 |
|
|
rabsq = rab[0] * rab[0] + rab[1] * rab[1] + rab[2] * rab[2]; |
362 |
|
|
rfab = rab[0] * fpab[0] + rab[1] * fpab[1] + rab[2] * fpab[2]; |
363 |
|
|
|
364 |
|
|
if (fabs(rfab) > sqrt(rabsq*gab) * 0.00001){ |
365 |
|
|
|
366 |
|
|
gab = -rfab /(rabsq*(rma + rmb)); |
367 |
|
|
|
368 |
|
|
frcA[0] = rab[0] * gab; |
369 |
|
|
frcA[1] = rab[1] * gab; |
370 |
|
|
frcA[2] = rab[2] * gab; |
371 |
|
|
|
372 |
|
|
atoms[a]->addFrc(frcA); |
373 |
|
|
|
374 |
|
|
|
375 |
|
|
frcB[0] = -rab[0] * gab; |
376 |
|
|
frcB[1] = -rab[1] * gab; |
377 |
|
|
frcB[2] = -rab[2] * gab; |
378 |
|
|
|
379 |
|
|
atoms[b]->addFrc(frcB); |
380 |
|
|
|
381 |
|
|
moving[a] = 1; |
382 |
|
|
moving[b] = 1; |
383 |
|
|
done = 0; |
384 |
|
|
} |
385 |
|
|
} |
386 |
|
|
} |
387 |
|
|
|
388 |
|
|
for (i = 0; i < nAtoms; i++){ |
389 |
|
|
moved[i] = moving[i]; |
390 |
|
|
moving[i] = 0; |
391 |
|
|
} |
392 |
|
|
|
393 |
|
|
iteration++; |
394 |
|
|
} |
395 |
|
|
|
396 |
|
|
if (!done){ |
397 |
|
|
cerr << "Waring: can not constraint within maxIteration" << endl; |
398 |
|
|
return -1; |
399 |
|
|
} |
400 |
|
|
else |
401 |
|
|
return 1; |
402 |
|
|
} |
403 |
|
|
|
404 |
|
|
//calculate the value of object function |
405 |
|
|
void OOPSEMinimizer::calcF(){ |
406 |
|
|
calcEnergyGradient(curX, curG, curF, egEvalStatus); |
407 |
|
|
} |
408 |
|
|
|
409 |
|
|
void OOPSEMinimizer::calcF(vector<double>& x, double&f, int& status){ |
410 |
|
|
vector<double> tempG; |
411 |
|
|
tempG.resize(x.size()); |
412 |
|
|
|
413 |
|
|
calcEnergyGradient(x, tempG, f, status); |
414 |
|
|
} |
415 |
|
|
|
416 |
|
|
//calculate the gradient |
417 |
|
|
void OOPSEMinimizer::calcG(){ |
418 |
|
|
calcEnergyGradient(curX, curG, curF, egEvalStatus); |
419 |
|
|
} |
420 |
|
|
|
421 |
|
|
void OOPSEMinimizer::calcG(vector<double>& x, vector<double>& g, double& f, int& status){ |
422 |
|
|
calcEnergyGradient(x, g, f, status); |
423 |
|
|
} |
424 |
|
|
|
425 |
|
|
void OOPSEMinimizer::calcDim(){ |
426 |
|
|
DirectionalAtom* dAtom; |
427 |
|
|
|
428 |
|
|
ndim = 0; |
429 |
|
|
|
430 |
gezelter |
1097 |
for(int i = 0; i < integrableObjects.size(); i++){ |
431 |
tim |
1064 |
ndim += 3; |
432 |
gezelter |
1097 |
if (integrableObjects[i]->isDirectional()) |
433 |
tim |
1064 |
ndim += 3; |
434 |
|
|
} |
435 |
|
|
} |
436 |
|
|
|
437 |
|
|
void OOPSEMinimizer::setX(vector < double > & x){ |
438 |
|
|
|
439 |
|
|
if (x.size() != ndim && bVerbose){ |
440 |
|
|
//sprintf(painCave.errMsg, |
441 |
|
|
// "OOPSEMinimizer Error: dimesion of x and curX does not match\n"); |
442 |
|
|
// painCave.isFatal = 1; |
443 |
|
|
// simError(); |
444 |
|
|
} |
445 |
|
|
|
446 |
|
|
curX = x; |
447 |
|
|
} |
448 |
|
|
|
449 |
|
|
void OOPSEMinimizer::setG(vector < double > & g){ |
450 |
|
|
|
451 |
|
|
if (g.size() != ndim && bVerbose){ |
452 |
|
|
//sprintf(painCave.errMsg, |
453 |
|
|
// "OOPSEMinimizer Error: dimesion of g and curG does not match\n"); |
454 |
|
|
// painCave.isFatal = 1; |
455 |
|
|
//simError(); |
456 |
|
|
} |
457 |
|
|
|
458 |
|
|
curG = g; |
459 |
|
|
} |
460 |
|
|
|
461 |
|
|
void OOPSEMinimizer::writeOut(vector<double>& x, double iter){ |
462 |
|
|
|
463 |
|
|
setX(x); |
464 |
|
|
|
465 |
|
|
calcG(); |
466 |
|
|
|
467 |
|
|
dumpOut->writeDump(iter); |
468 |
|
|
statOut->writeStat(iter); |
469 |
|
|
} |
470 |
|
|
|
471 |
|
|
|
472 |
|
|
void OOPSEMinimizer::printMinimizerInfo(){ |
473 |
|
|
cout << "--------------------------------------------------------------------" << endl; |
474 |
|
|
cout << minimizerName << endl; |
475 |
|
|
cout << "minimization parameter set" << endl; |
476 |
|
|
cout << "function tolerance = " << paramSet->getFTol() << endl; |
477 |
|
|
cout << "gradient tolerance = " << paramSet->getGTol() << endl; |
478 |
|
|
cout << "step tolerance = "<< paramSet->getFTol() << endl; |
479 |
|
|
cout << "absolute gradient tolerance = " << endl; |
480 |
|
|
cout << "max iteration = " << paramSet->getMaxIteration() << endl; |
481 |
|
|
cout << "max line search iteration = " << paramSet->getLineSearchMaxIteration() <<endl; |
482 |
|
|
cout << "shake algorithm = " << bShake << endl; |
483 |
|
|
cout << "--------------------------------------------------------------------" << endl; |
484 |
|
|
|
485 |
|
|
} |
486 |
|
|
|
487 |
|
|
/** |
488 |
|
|
* In thoery, we need to find the minimum along the search direction |
489 |
tim |
1108 |
* However, function evaluation is too expensive. |
490 |
tim |
1064 |
* At the very begining of the problem, we check the search direction and make sure |
491 |
|
|
* it is a descent direction |
492 |
|
|
* we will compare the energy of two end points, |
493 |
|
|
* if the right end point has lower energy, we just take it |
494 |
|
|
* |
495 |
|
|
* |
496 |
|
|
* |
497 |
|
|
*/ |
498 |
|
|
|
499 |
|
|
int OOPSEMinimizer::doLineSearch(vector<double>& direction, double stepSize){ |
500 |
|
|
vector<double> xa; |
501 |
|
|
vector<double> xb; |
502 |
|
|
vector<double> xc; |
503 |
|
|
vector<double> ga; |
504 |
|
|
vector<double> gb; |
505 |
|
|
vector<double> gc; |
506 |
|
|
double fa; |
507 |
|
|
double fb; |
508 |
|
|
double fc; |
509 |
|
|
double a; |
510 |
|
|
double b; |
511 |
|
|
double c; |
512 |
|
|
int status; |
513 |
|
|
double initSlope; |
514 |
|
|
double slopeA; |
515 |
|
|
double slopeB; |
516 |
|
|
double slopeC; |
517 |
|
|
bool foundLower; |
518 |
|
|
int iter; |
519 |
|
|
int maxLSIter; |
520 |
|
|
double mu; |
521 |
|
|
double eta; |
522 |
|
|
double ftol; |
523 |
tim |
1074 |
double lsTol; |
524 |
tim |
1064 |
|
525 |
|
|
xa.resize(ndim); |
526 |
|
|
xb.resize(ndim); |
527 |
|
|
xc.resize(ndim); |
528 |
|
|
|
529 |
|
|
ga.resize(ndim); |
530 |
|
|
gb.resize(ndim); |
531 |
|
|
gc.resize(ndim); |
532 |
|
|
|
533 |
|
|
a = 0.0; |
534 |
|
|
fa = curF; |
535 |
|
|
xa = curX; |
536 |
|
|
ga = curG; |
537 |
|
|
c = a + stepSize; |
538 |
|
|
ftol = paramSet->getFTol(); |
539 |
tim |
1074 |
lsTol = paramSet->getLineSearchTol(); |
540 |
tim |
1064 |
|
541 |
|
|
//calculate the derivative at a = 0 |
542 |
|
|
for (size_t i = 0; i < ndim; i++) |
543 |
|
|
slopeA += curG[i]*direction[i]; |
544 |
|
|
|
545 |
|
|
initSlope = slopeA; |
546 |
|
|
|
547 |
|
|
// if going uphill, use negative gradient as searching direction |
548 |
|
|
if (slopeA > 0) { |
549 |
|
|
|
550 |
|
|
if (bVerbose){ |
551 |
|
|
cout << "LineSearch Warning: initial searching direction is not a descent searching direction, " |
552 |
|
|
<< " use negative gradient instead. Therefore, finding a smaller vaule of function " |
553 |
|
|
<< " is guaranteed" |
554 |
|
|
<< endl; |
555 |
|
|
} |
556 |
|
|
|
557 |
|
|
for (size_t i = 0; i < ndim; i++) |
558 |
|
|
direction[i] = -curG[i]; |
559 |
|
|
|
560 |
|
|
for (size_t i = 0; i < ndim; i++) |
561 |
|
|
slopeA += curG[i]*direction[i]; |
562 |
|
|
|
563 |
|
|
initSlope = slopeA; |
564 |
|
|
} |
565 |
|
|
|
566 |
|
|
// Take a trial step |
567 |
|
|
for(size_t i = 0; i < ndim; i++) |
568 |
|
|
xc[i] = curX[i] + direction[i] * c; |
569 |
|
|
|
570 |
|
|
calcG(xc, gc, fc, status); |
571 |
|
|
|
572 |
|
|
if (status < 0){ |
573 |
|
|
if (bVerbose) |
574 |
|
|
cerr << "Function Evaluation Error" << endl; |
575 |
|
|
} |
576 |
|
|
|
577 |
|
|
//calculate the derivative at c |
578 |
|
|
slopeC = 0; |
579 |
|
|
for (size_t i = 0; i < ndim; i++) |
580 |
|
|
slopeC += gc[i]*direction[i]; |
581 |
|
|
|
582 |
|
|
// found a lower point |
583 |
|
|
if (fc < fa) { |
584 |
|
|
curX = xc; |
585 |
|
|
curG = gc; |
586 |
|
|
curF = fc; |
587 |
|
|
return LS_SUCCEED; |
588 |
|
|
} |
589 |
|
|
else { |
590 |
|
|
|
591 |
|
|
if (slopeC > 0) |
592 |
|
|
stepSize *= 0.618034; |
593 |
|
|
} |
594 |
|
|
|
595 |
|
|
maxLSIter = paramSet->getLineSearchMaxIteration(); |
596 |
|
|
|
597 |
|
|
iter = 0; |
598 |
|
|
|
599 |
|
|
do { |
600 |
|
|
// Select a new trial point. |
601 |
|
|
// If the derivatives at points a & c have different sign we use cubic interpolate |
602 |
|
|
//if (slopeC > 0){ |
603 |
|
|
eta = 3 *(fa -fc) /(c - a) + slopeA + slopeC; |
604 |
|
|
mu = sqrt(eta * eta - slopeA * slopeC); |
605 |
|
|
b = a + (c - a) * (1 - (slopeC + mu - eta) /(slopeC - slopeA + 2 * mu)); |
606 |
tim |
1074 |
|
607 |
|
|
if (b < lsTol){ |
608 |
|
|
if (bVerbose) |
609 |
|
|
cout << "stepSize is less than line search tolerance" << endl; |
610 |
|
|
break; |
611 |
|
|
} |
612 |
tim |
1064 |
//} |
613 |
|
|
|
614 |
|
|
// Take a trial step to this new point - new coords in xb |
615 |
|
|
for(size_t i = 0; i < ndim; i++) |
616 |
|
|
xb[i] = curX[i] + direction[i] * b; |
617 |
|
|
|
618 |
|
|
//function evaluation |
619 |
|
|
calcG(xb, gb, fb, status); |
620 |
|
|
|
621 |
|
|
if (status < 0){ |
622 |
|
|
if (bVerbose) |
623 |
|
|
cerr << "Function Evaluation Error" << endl; |
624 |
|
|
} |
625 |
|
|
|
626 |
|
|
//calculate the derivative at c |
627 |
|
|
slopeB = 0; |
628 |
|
|
for (size_t i = 0; i < ndim; i++) |
629 |
|
|
slopeB += gb[i]*direction[i]; |
630 |
|
|
|
631 |
|
|
//Amijo Rule to stop the line search |
632 |
|
|
if (fb <= curF + initSlope * ftol * b) { |
633 |
|
|
curF = fb; |
634 |
|
|
curX = xb; |
635 |
|
|
curG = gb; |
636 |
|
|
return LS_SUCCEED; |
637 |
|
|
} |
638 |
|
|
|
639 |
|
|
if (slopeB <0 && fb < fa) { |
640 |
|
|
//replace a by b |
641 |
|
|
fa = fb; |
642 |
|
|
a = b; |
643 |
|
|
slopeA = slopeB; |
644 |
|
|
|
645 |
|
|
// swap coord a/b |
646 |
|
|
std::swap(xa, xb); |
647 |
|
|
std::swap(ga, gb); |
648 |
|
|
} |
649 |
|
|
else { |
650 |
|
|
//replace c by b |
651 |
|
|
fc = fb; |
652 |
|
|
c = b; |
653 |
|
|
slopeC = slopeB; |
654 |
|
|
|
655 |
|
|
// swap coord b/c |
656 |
|
|
std::swap(gb, gc); |
657 |
|
|
std::swap(xb, xc); |
658 |
|
|
} |
659 |
|
|
|
660 |
|
|
|
661 |
|
|
iter++; |
662 |
|
|
} while((fb > fa || fb > fc) && (iter < maxLSIter)); |
663 |
|
|
|
664 |
|
|
if(fb < curF || iter >= maxLSIter) { |
665 |
|
|
//could not find a lower value, we might just go uphill. |
666 |
|
|
return LS_ERROR; |
667 |
|
|
} |
668 |
|
|
|
669 |
|
|
//select the end point |
670 |
|
|
|
671 |
|
|
if (fa <= fc) { |
672 |
|
|
curX = xa; |
673 |
|
|
curG = ga; |
674 |
|
|
curF = fa; |
675 |
|
|
} |
676 |
|
|
else { |
677 |
|
|
curX = xc; |
678 |
|
|
curG = gc; |
679 |
|
|
curF = fc; |
680 |
|
|
} |
681 |
|
|
|
682 |
|
|
return LS_SUCCEED; |
683 |
|
|
|
684 |
|
|
} |
685 |
|
|
|
686 |
|
|
void OOPSEMinimizer::minimize(){ |
687 |
|
|
|
688 |
|
|
int convgStatus; |
689 |
|
|
int stepStatus; |
690 |
|
|
int maxIter; |
691 |
|
|
//int resetFrq; |
692 |
|
|
//int nextResetIter; |
693 |
|
|
int writeFrq; |
694 |
|
|
int nextWriteIter; |
695 |
|
|
|
696 |
|
|
if (bVerbose) |
697 |
|
|
printMinimizerInfo(); |
698 |
tim |
1144 |
|
699 |
|
|
dumpOut = new DumpWriter(info); |
700 |
|
|
statOut = new StatWriter(info); |
701 |
tim |
1064 |
|
702 |
|
|
init(); |
703 |
|
|
|
704 |
|
|
//resetFrq = paramSet->getResetFrq(); |
705 |
|
|
//nextResetIter = resetFrq; |
706 |
|
|
|
707 |
|
|
writeFrq = paramSet->getWriteFrq(); |
708 |
|
|
nextWriteIter = writeFrq; |
709 |
|
|
|
710 |
|
|
maxIter = paramSet->getMaxIteration(); |
711 |
|
|
|
712 |
|
|
for (curIter = 1; curIter <= maxIter; curIter++){ |
713 |
|
|
|
714 |
|
|
stepStatus = step(); |
715 |
|
|
|
716 |
|
|
if (stepStatus < 0){ |
717 |
|
|
saveResult(); |
718 |
|
|
minStatus = MIN_LSERROR; |
719 |
|
|
cerr << "OOPSEMinimizer Error: line search error, please try a small stepsize" << endl; |
720 |
|
|
return; |
721 |
|
|
} |
722 |
|
|
|
723 |
|
|
if (curIter == nextWriteIter){ |
724 |
|
|
nextWriteIter += writeFrq; |
725 |
|
|
writeOut(curX, curIter); |
726 |
|
|
} |
727 |
|
|
|
728 |
|
|
//if (curIter == nextResetIter){ |
729 |
|
|
// reset(); |
730 |
|
|
// nextResetIter += resetFrq; |
731 |
|
|
//} |
732 |
|
|
|
733 |
|
|
convgStatus = checkConvg(); |
734 |
|
|
|
735 |
|
|
if (convgStatus > 0){ |
736 |
|
|
saveResult(); |
737 |
|
|
minStatus = MIN_CONVERGE; |
738 |
|
|
return; |
739 |
|
|
} |
740 |
|
|
|
741 |
|
|
prepareStep(); |
742 |
|
|
|
743 |
|
|
} |
744 |
|
|
|
745 |
|
|
|
746 |
|
|
if (bVerbose) { |
747 |
|
|
cout << "OOPSEMinimizer Warning: " |
748 |
|
|
<< minimizerName << " algorithm did not converge within " |
749 |
|
|
<< maxIter << " iteration" << endl; |
750 |
|
|
} |
751 |
|
|
minStatus = MIN_MAXITER; |
752 |
|
|
saveResult(); |
753 |
|
|
|
754 |
|
|
} |