| 1 |
tim |
1011 |
#include "ConjugateMinimizer.hpp"
|
| 2 |
|
|
bool ConjugateMinimizerBase::isSolvable(){
|
| 3 |
|
|
|
| 4 |
|
|
//conjuage gradient can only solve unconstrained nonlinear model
|
| 5 |
|
|
|
| 6 |
|
|
if (!model->hasConstraint())
|
| 7 |
|
|
return true;
|
| 8 |
|
|
else
|
| 9 |
|
|
return false;
|
| 10 |
|
|
}
|
| 11 |
|
|
|
| 12 |
|
|
void ConjugateMinimizerBase::Init(){
|
| 13 |
|
|
|
| 14 |
|
|
}
|
| 15 |
|
|
|
| 16 |
|
|
void ConjugateMinimizerBase::printMinizerInfo(){
|
| 17 |
|
|
|
| 18 |
|
|
}
|
| 19 |
|
|
|
| 20 |
|
|
void ConjugateMinimizerBase::Minimize(){
|
| 21 |
|
|
int maxIteration;
|
| 22 |
|
|
int nextRestIter;
|
| 23 |
|
|
int resetFrq;
|
| 24 |
|
|
int nextWriteIter;
|
| 25 |
|
|
int writeFrq;
|
| 26 |
|
|
|
| 27 |
|
|
if (!isSolvable()){
|
| 28 |
|
|
cout << "ConjugateMinimizerBase Error: This nonlinear model can not be solved by " << methodName <<endl;
|
| 29 |
|
|
|
| 30 |
|
|
exit(1);
|
| 31 |
|
|
}
|
| 32 |
|
|
|
| 33 |
|
|
printMinizerInfo();
|
| 34 |
|
|
|
| 35 |
|
|
resetFrq = paramSet->getResetFrq();
|
| 36 |
|
|
nextRestIter = resetFrq;
|
| 37 |
|
|
|
| 38 |
|
|
writeFrq = paramSet->getWriteFrq();
|
| 39 |
|
|
nextWriteIter = writeFrq;
|
| 40 |
|
|
|
| 41 |
|
|
prevGrad = model->calcGrad();
|
| 42 |
|
|
|
| 43 |
|
|
direction = preGrad;
|
| 44 |
|
|
|
| 45 |
|
|
maxIteration = paramSet->getMaxIteration();
|
| 46 |
|
|
|
| 47 |
|
|
for(currentIter = 0;currentIter < maxIteration; currentIter++){
|
| 48 |
|
|
|
| 49 |
|
|
// perform line search to minimize f(x + stepSize * direction) where stepSize > 0
|
| 50 |
|
|
lsMinimizer->minimize(direction, 0.0, 1.0);
|
| 51 |
|
|
|
| 52 |
|
|
lsStatus = lsMinimizer->getMinimizationStatus();
|
| 53 |
|
|
|
| 54 |
|
|
if(lsStatus ==MINSTATUS_ERROR){
|
| 55 |
|
|
minStatus = MINSTATUS_ERROR;
|
| 56 |
|
|
return;
|
| 57 |
|
|
}
|
| 58 |
|
|
|
| 59 |
|
|
prevMinX = minX;
|
| 60 |
|
|
minX = minX + lsMinimizer->getMinVar() * direction;
|
| 61 |
|
|
|
| 62 |
|
|
//calculate the gradient
|
| 63 |
|
|
prevGrad = gradient;
|
| 64 |
|
|
|
| 65 |
|
|
gradient = model->calcGrad();
|
| 66 |
|
|
|
| 67 |
|
|
// stop if converge
|
| 68 |
|
|
convStatus = checkConvergence();
|
| 69 |
|
|
if (convStatus == ){
|
| 70 |
|
|
writeOut(minX, currentIter);
|
| 71 |
|
|
|
| 72 |
|
|
minStatus = MINSTATUS_CONVERGE;
|
| 73 |
|
|
return;
|
| 74 |
|
|
}
|
| 75 |
|
|
|
| 76 |
|
|
|
| 77 |
|
|
//calculate the
|
| 78 |
|
|
gamma = calcGamma(grad, preGrad);
|
| 79 |
|
|
|
| 80 |
|
|
// update new direction
|
| 81 |
|
|
prevDirection = direction;
|
| 82 |
|
|
direction += gamma * direction;
|
| 83 |
|
|
|
| 84 |
|
|
//
|
| 85 |
|
|
if (currentIter == nextWriteIter){
|
| 86 |
|
|
nextWriteIter += writeFrq;
|
| 87 |
|
|
writeOut(minX, currentIter);
|
| 88 |
|
|
}
|
| 89 |
|
|
|
| 90 |
|
|
if (currentIter == nextResetIter){
|
| 91 |
|
|
reset();
|
| 92 |
|
|
nextResetIter += resetFrq;
|
| 93 |
|
|
}
|
| 94 |
|
|
|
| 95 |
|
|
}
|
| 96 |
|
|
|
| 97 |
|
|
// if writeFrq is not a multipiler of maxIteration, we need to write the final result
|
| 98 |
|
|
// otherwise, we already write it inside the loop, just skip it
|
| 99 |
|
|
if(currentIter != (nextWriteIter - writeFrq))
|
| 100 |
|
|
writeOut(minX, currentIter);
|
| 101 |
|
|
|
| 102 |
|
|
minStatus = MINSTATUS_MAXITER;
|
| 103 |
|
|
return;
|
| 104 |
|
|
}
|
| 105 |
|
|
|
| 106 |
|
|
int ConjugateMinimizerBase::checkConvergence(){
|
| 107 |
|
|
|
| 108 |
|
|
//test absolute gradient tolerance
|
| 109 |
|
|
|
| 110 |
|
|
if (norm2(gradient) < paramSet->absGradTol)
|
| 111 |
|
|
return 1;
|
| 112 |
|
|
else
|
| 113 |
|
|
return -1;
|
| 114 |
|
|
}
|
| 115 |
|
|
|
| 116 |
|
|
void ConjugateMinimizerBase::reset(){
|
| 117 |
|
|
|
| 118 |
|
|
}
|
| 119 |
|
|
|
| 120 |
|
|
double FRCGMinimizer::calcGamma(vector<double>& newGrad, vector<double>& oldGrad){
|
| 121 |
|
|
return norm2(newGrad) / norm2(oldGrad);
|
| 122 |
|
|
}
|
| 123 |
|
|
|
| 124 |
|
|
double PRCGMinimizer::calcGamma(vector<double>& newGrad, vector<double>& oldGrad){
|
| 125 |
|
|
double gamma;
|
| 126 |
|
|
vector<double> deltaGrad;
|
| 127 |
|
|
|
| 128 |
|
|
deltaGrad = newGrad - oldGrad;
|
| 129 |
|
|
|
| 130 |
|
|
return norm(deltaGrad, newGrad) / norm2(oldGrad);
|
| 131 |
|
|
|
| 132 |
|
|
}
|
