static double nloptObjectiveFunction(unsigned n, const double *x, double *grad, void *f_data)
{
GradientOptimizerContext *goc = (GradientOptimizerContext *) f_data;
nlopt_opt opt = (nlopt_opt) goc->extraData;
FitContext *fc = goc->fc;
assert(n == fc->numParam);
int mode = 0;
double fit = goc->solFun((double*) x, &mode);
if (grad) {
fc->iterations += 1;
if (goc->maxMajorIterations != -1 && fc->iterations >= goc->maxMajorIterations) {
nlopt_force_stop(opt);
}
}
if (grad && goc->verbose >= 2) {
mxLog("major iteration fit=%.12f", fit);
}
if (mode == -1) {
if (!goc->feasible) {
nlopt_force_stop(opt);
}
return nan("infeasible");
}
if (!grad) return fit;
Eigen::Map< Eigen::VectorXd > Epoint((double*) x, n);
Eigen::Map< Eigen::VectorXd > Egrad(grad, n);
if (fc->wanted & FF_COMPUTE_GRADIENT) {
Egrad = fc->grad;
} else if (fc->CI && fc->CI->varIndex >= 0) {
Egrad.setZero();
Egrad[fc->CI->varIndex] = fc->lowerBound? 1 : -1;
fc->grad = Egrad;
} else {
if (goc->verbose >= 3) mxLog("fd_gradient start");
fit_functional ff(*goc);
gradient_with_ref(goc->gradientAlgo, goc->gradientIterations, goc->gradientStepSize,
ff, fit, Epoint, Egrad);
fc->grad = Egrad;
}
if (goc->verbose >= 3) {
mxPrintMat("gradient", Egrad);
}
return fit;
}
void omxSD(GradientOptimizerContext &rf)
{
int maxIter = rf.maxMajorIterations;
if (maxIter == -1) maxIter = 50000;
Eigen::VectorXd currEst(rf.numFree);
rf.copyToOptimizer(currEst.data());
int iter = 0;
double priorSpeed = 1.0, shrinkage = 0.7;
rf.setupSimpleBounds();
rf.informOut = INFORM_UNINITIALIZED;
{
int mode = 0;
rf.solFun(currEst.data(), &mode);
if (mode == -1) {
rf.informOut = INFORM_STARTING_VALUES_INFEASIBLE;
return;
}
}
double refFit = rf.getFit();
rf.grad.resize(rf.numFree);
fit_functional ff(rf);
Eigen::VectorXd majorEst = currEst;
while(++iter < maxIter && !isErrorRaised()) {
gradient_with_ref(rf.gradientAlgo, 1, rf.gradientIterations, rf.gradientStepSize,
ff, refFit, majorEst, rf.grad);
if (rf.verbose >= 3) mxPrintMat("grad", rf.grad);
if(rf.grad.norm() == 0)
{
rf.informOut = INFORM_CONVERGED_OPTIMUM;
if(rf.verbose >= 2) mxLog("After %i iterations, gradient achieves zero!", iter);
break;
}
int retries = 300;
double speed = std::min(priorSpeed, 1.0);
double bestSpeed = speed;
bool foundBetter = false;
Eigen::VectorXd bestEst(majorEst.size());
Eigen::VectorXd prevEst(majorEst.size());
Eigen::VectorXd searchDir = rf.grad;
searchDir /= searchDir.norm();
prevEst.setConstant(nan("uninit"));
while (--retries > 0 && !isErrorRaised()){
Eigen::VectorXd nextEst = majorEst - speed * searchDir;
nextEst = nextEst.cwiseMax(rf.solLB).cwiseMin(rf.solUB);
if (nextEst == prevEst) break;
prevEst = nextEst;
rf.checkActiveBoxConstraints(nextEst);
int mode = 0;
double fit = rf.solFun(nextEst.data(), &mode);
if (fit < refFit) {
foundBetter = true;
refFit = rf.getFit();
bestSpeed = speed;
bestEst = nextEst;
break;
}
speed *= shrinkage;
}
if (false && foundBetter) {
// In some tests, this did not help so it is not enabled.
// It might be worth testing more.
mxLog("trying larger step size");
retries = 3;
while (--retries > 0 && !isErrorRaised()){
speed *= 1.01;
Eigen::VectorXd nextEst = majorEst - speed * searchDir;
nextEst = nextEst.cwiseMax(rf.solLB).cwiseMin(rf.solUB);
rf.checkActiveBoxConstraints(nextEst);
int mode = 0;
double fit = rf.solFun(nextEst.data(), &mode);
if (fit < refFit) {
foundBetter = true;
refFit = rf.getFit();
bestSpeed = speed;
bestEst = nextEst;
}
}
}
if (!foundBetter) {
rf.informOut = INFORM_CONVERGED_OPTIMUM;
if(rf.verbose >= 2) mxLog("After %i iterations, cannot find better estimation along the gradient direction", iter);
break;
}
if (rf.verbose >= 2) mxLog("major fit %f bestSpeed %g", refFit, bestSpeed);
majorEst = bestEst;
priorSpeed = bestSpeed * 1.1;
}
rf.est = majorEst;
if ((rf.grad.array().abs() > 0.1).any()) {
rf.informOut = INFORM_NOT_AT_OPTIMUM;
}
if (iter >= maxIter - 1) {
rf.informOut = INFORM_ITERATION_LIMIT;
if(rf.verbose >= 2) mxLog("Maximum iteration achieved!");
}
if(rf.verbose >= 1) mxLog("Status code : %i", rf.informOut);
}