| 1 |
#include "Minimizer1D.hpp"
|
| 2 |
#include "math.h"
|
| 3 |
|
| 4 |
//----------------------------------------------------------------------------//
|
| 5 |
void Minimizer1D::Minimize(vector<double>& direction, double left, double right){
|
| 6 |
setDirection(direction);
|
| 7 |
setRange(left,right);
|
| 8 |
minimize();
|
| 9 |
}
|
| 10 |
|
| 11 |
//----------------------------------------------------------------------------//
|
| 12 |
GoldenSectionMinimizer::GoldenSectionMinimizer(NLModel* nlp)
|
| 13 |
:Minimizer1D(nlp){
|
| 14 |
setName("GoldenSection");
|
| 15 |
}
|
| 16 |
|
| 17 |
int GoldenSectionMinimizer::checkConvergence(){
|
| 18 |
|
| 19 |
if ((rightVar - leftVar) < stepTol)
|
| 20 |
return 1;
|
| 21 |
else
|
| 22 |
return -1;
|
| 23 |
}
|
| 24 |
void GoldenSectionMinimizer::minimize(){
|
| 25 |
vector<double> tempX;
|
| 26 |
vector <double> currentX;
|
| 27 |
|
| 28 |
const double goldenRatio = 0.618034;
|
| 29 |
|
| 30 |
currentX = model->getX();
|
| 31 |
|
| 32 |
alpha = leftVar + (1 - goldenRatio) * (rightVar - leftVar);
|
| 33 |
beta = leftVar + goldenRatio * (rightVar - leftVar);
|
| 34 |
|
| 35 |
tempX = currentX + direction * alpha;
|
| 36 |
fAlpha = model->calcF(tempX);
|
| 37 |
|
| 38 |
tempX = currentX + direction * beta;
|
| 39 |
fBeta = model->calcF(tempX);
|
| 40 |
|
| 41 |
for(currentIter = 0; currentIter < maxIteration; currentIter++){
|
| 42 |
|
| 43 |
if (checkConvergence() > 0){
|
| 44 |
minStatus = MINSTATUS_CONVERGE;
|
| 45 |
return;
|
| 46 |
}
|
| 47 |
|
| 48 |
if (fAlpha > fBeta){
|
| 49 |
leftVar = alpha;
|
| 50 |
alpha = beta;
|
| 51 |
beta = leftVar + goldenRatio * (rightVar - leftVar);
|
| 52 |
|
| 53 |
tempX = currentX + beta * direction;
|
| 54 |
|
| 55 |
prevMinVar = minVar;
|
| 56 |
fPrevMinVar = fMinVar;
|
| 57 |
|
| 58 |
minVar = beta;
|
| 59 |
fMinVar = model->calcF(tempX);
|
| 60 |
|
| 61 |
}
|
| 62 |
else{
|
| 63 |
rightVar = beta;
|
| 64 |
beta = alpha;
|
| 65 |
alpha = leftVar + (1 - goldenRatio) * (rightVar - leftVar);
|
| 66 |
|
| 67 |
tempX = currentX + alpha * direction;
|
| 68 |
|
| 69 |
prevMinVar = minVar;
|
| 70 |
fPrevMinVar = fMinVar;
|
| 71 |
|
| 72 |
minVar = alpha;
|
| 73 |
fMinVar = model->calcF(tempX);
|
| 74 |
}
|
| 75 |
|
| 76 |
}
|
| 77 |
|
| 78 |
minStatus = MINSTATUS_MAXITER;
|
| 79 |
|
| 80 |
}
|
| 81 |
|
| 82 |
/**
|
| 83 |
* Brent's method is a root-finding algorithm which combines root bracketing, interval bisection,
|
| 84 |
* and inverse quadratic interpolation.
|
| 85 |
*/
|
| 86 |
BrentMinimizer::BrentMinimizer(NLModel* nlp)
|
| 87 |
:Minimizer1D(nlp){
|
| 88 |
setName("Brent");
|
| 89 |
}
|
| 90 |
|
| 91 |
void BrentMinimizer::minimize(){
|
| 92 |
|
| 93 |
double fu, fv, fw;
|
| 94 |
double p, q, r;
|
| 95 |
double u, v, w;
|
| 96 |
double d;
|
| 97 |
double e;
|
| 98 |
double etemp;
|
| 99 |
double stepTol2;
|
| 100 |
double fLeftVar, fRightVar;
|
| 101 |
const double goldenRatio = 0.3819660;
|
| 102 |
vector<double> tempX, currentX;
|
| 103 |
|
| 104 |
stepTol2 = 2 * stepTol;
|
| 105 |
e = 0;
|
| 106 |
d = 0;
|
| 107 |
|
| 108 |
currentX = tempX = model->getX();
|
| 109 |
|
| 110 |
tempX = currentX + leftVar * direction;
|
| 111 |
fLeftVar = model->calcF(tempX);
|
| 112 |
|
| 113 |
tempX = currentX + rightVar * direction;
|
| 114 |
fRightVar = model->calcF(tempX);
|
| 115 |
|
| 116 |
if(fRightVar < fLeftVar) {
|
| 117 |
prevMinVar = rightVar;
|
| 118 |
fPrevMinVar = fRightVar;
|
| 119 |
v = leftVar;
|
| 120 |
fv = fLeftVar;
|
| 121 |
}
|
| 122 |
else {
|
| 123 |
prevMinVar = leftVar;
|
| 124 |
fPrevMinVar = fLeftVar;
|
| 125 |
v = rightVar;
|
| 126 |
fv = fRightVar;
|
| 127 |
}
|
| 128 |
|
| 129 |
midVar = leftVar + rightVar;
|
| 130 |
|
| 131 |
for(currentIter = 0; currentIter < maxIteration; currentIter){
|
| 132 |
|
| 133 |
// a trial parabolic fit
|
| 134 |
if (fabs(e) > stepTol){
|
| 135 |
|
| 136 |
r = (minVar - prevMinVar) * (fMinVar - fv);
|
| 137 |
q = (minVar - v) * (fMinVar - fPrevMinVar);
|
| 138 |
p = (minVar - v) *q -(minVar - prevMinVar)*r;
|
| 139 |
q = 2.0 *(q-r);
|
| 140 |
|
| 141 |
if (q > 0.0)
|
| 142 |
p = -p;
|
| 143 |
|
| 144 |
q = fabs(q);
|
| 145 |
|
| 146 |
etemp = e;
|
| 147 |
e = d;
|
| 148 |
|
| 149 |
if(fabs(p) >= fabs(0.5*q*etemp) || p <= q*(leftVar - minVar) || p >= q*(rightVar - minVar)){
|
| 150 |
e = minVar >= midVar ? leftVar - minVar : rightVar - minVar;
|
| 151 |
d = goldenRatio * e;
|
| 152 |
}
|
| 153 |
else{
|
| 154 |
d = p/q;
|
| 155 |
u = minVar + d;
|
| 156 |
if ( u - leftVar < stepTol2 || rightVar - u < stepTol2)
|
| 157 |
d = midVar > minVar ? stepTol : - stepTol;
|
| 158 |
}
|
| 159 |
}
|
| 160 |
//golden section
|
| 161 |
else{
|
| 162 |
e = minVar >=midVar? leftVar - minVar : rightVar - minVar;
|
| 163 |
d =goldenRatio * e;
|
| 164 |
}
|
| 165 |
|
| 166 |
u = fabs(d) >= stepTol ? minVar + d : minVar + copysign(d, stepTol);
|
| 167 |
|
| 168 |
tempX = currentX + u * direction;
|
| 169 |
fu = model->calcF();
|
| 170 |
|
| 171 |
if(fu <= fMinVar){
|
| 172 |
|
| 173 |
if(u >= minVar)
|
| 174 |
leftVar = minVar;
|
| 175 |
else
|
| 176 |
rightVar = minVar;
|
| 177 |
|
| 178 |
v = prevMinVar;
|
| 179 |
fv = fPrevMinVar;
|
| 180 |
prevMinVar = minVar;
|
| 181 |
fPrevMinVar = fMinVar;
|
| 182 |
minVar = u;
|
| 183 |
fMinVar = fu;
|
| 184 |
|
| 185 |
}
|
| 186 |
else{
|
| 187 |
if (u < minVar) leftVar = u;
|
| 188 |
else rightVar= u;
|
| 189 |
if(fu <= fPrevMinVar || prevMinVar == minVar) {
|
| 190 |
v = prevMinVar;
|
| 191 |
fv = fPrevMinVar;
|
| 192 |
prevMinVar = u;
|
| 193 |
fPrevMinVar = fu;
|
| 194 |
}
|
| 195 |
else if ( fu <= fv || v == minVar || v == prevMinVar ) {
|
| 196 |
v = u;
|
| 197 |
fv = fu;
|
| 198 |
}
|
| 199 |
}
|
| 200 |
|
| 201 |
midVar = leftVar + rightVar;
|
| 202 |
|
| 203 |
if (checkConvergence() > 0){
|
| 204 |
minStatus = MINSTATUS_CONVERGE;
|
| 205 |
return;
|
| 206 |
}
|
| 207 |
|
| 208 |
}
|
| 209 |
|
| 210 |
|
| 211 |
minStatus = MINSTATUS_MAXITER;
|
| 212 |
return;
|
| 213 |
}
|
| 214 |
|
| 215 |
int BrentMinimizer::checkConvergence(){
|
| 216 |
|
| 217 |
if (fabs(minVar - midVar) < stepTol)
|
| 218 |
return 1;
|
| 219 |
else
|
| 220 |
return -1;
|
| 221 |
}
|
