void GPOptimizerRprop::maximize(GPRegression* gp, int const& n,
bool const& verbose) {
int const& n_dim_Y = gp->yDimension();
for (int d = 0; d < n_dim_Y; ++d) {
KernelFunction* Kf = gp->kernelFunction(d);
int const& n_dim_param = Kf->parameterDimension();
VecN delta = VecN::Ones(n_dim_param) * _p->_delta0;
VecN grad_old = VecN::Zero(n_dim_param);
VecN params = Kf->logParameters();
VecN best_params = params;
Scalar best = log(0);
for (int i = 0; i < n; ++i) {
VecN grad = -gp->logLikelihoodGrad(d);
grad_old = grad_old.cwiseProduct(grad);
for (int j = 0; j < grad_old.size(); ++j) {
if (grad_old(j) > 0) {
delta(j) = std::min(delta(j) * _p->_eta_plus, _p->_delta_max);
} else if (grad_old(j) < 0) {
delta(j) = std::max(delta(j) * _p->_eta_minus, _p->_delta_min);
grad(j) = 0;
}
params(j) += -_p->sign(grad(j)) * delta(j);
}
grad_old = grad;
if (grad_old.norm() < _p->_eps_stop) break;
Kf->initLogParameters(params);
Scalar lik = gp->logLikelihood(d);
if (verbose) std::cout << i << " " << -lik << std::endl;
if (lik > best) {
best = lik;
best_params = params;
}
}
Kf->initLogParameters(best_params);
}
}
void RpropSolver::internalSolve(Vector & x0,
const function_t & FunctionValue,
const gradient_t & FunctionGradient,
const hessian_t & FunctionHessian)
{
//helper function
auto sign = [](double x) -> double
{
if (x>0) return 1.0;
if (x<0) return -1.0;
return 0.0;
};
const size_t DIM = x0.rows();
Vector Delta = Eigen::VectorXd::Ones ( DIM ) * Delta0; //step size
Vector Delta_old = Eigen::VectorXd::Ones ( DIM ) * Delta0; //step size
Vector grad = Eigen::VectorXd::Zero ( DIM );
Vector grad_old = Eigen::VectorXd::Zero ( DIM );
Vector params = x0; // hyperparameter
//Vector params_old = x0;
Vector best_params = params;
double best = std::numeric_limits<double>::infinity();
size_t iter = 0;
do
{
FunctionGradient(params, grad); // calc gradient
auto FunctionValue_this = FunctionValue(params);
grad_old = grad_old.cwiseProduct ( grad );
auto params_new = params;
for ( int j = 0; j < grad_old.size(); ++j )
{
if ( grad_old ( j ) > 0 ) // no sign change
{
Delta ( j ) = std::min ( Delta ( j ) * etaplus, Deltamax );
}
else if ( grad_old ( j ) < 0 ) // sign change
{
Delta ( j ) = std::max ( Delta ( j ) * etaminus, Deltamin );
}
params_new ( j ) += -sign(grad (j)) * Delta (j);
}
params = std::move(params_new);
//std::cout << best << std::endl;
//std::cout << iter << " "<< FunctionValue_this << std::endl;
FunctionValueHistory.push_back(FunctionValue_this);
x0History.push_back(x0);
if ( FunctionValue_this < best )
{
best = FunctionValue_this;
best_params = params;
}
grad_old = grad;
//params_old = params;
Delta_old = Delta;
++iter;
}
//while((grad_old.lpNorm<Eigen::Infinity>() > settings.gradTol) && (iter < settings.maxIter)); // if ( grad_old.norm() < eps_stop ) break;
while((iter < settings.maxIter)); // if ( grad_old.norm() < eps_stop ) break;
// save best hyperparameters
double temp = FunctionValue(best_params);
}
