C++ exceptions are ignored when eliminating dimensions

[roelmatthysen]

Hi,

When the objective function throws an exception such as nlopt::forced_stop, optimization should be gracefully halted and an exception is thrown to the caller. However when bounds coincide for one or more variables in combination with a gradient-free algorithm, this exception is never rethrown.

When stepping through the code, it seems the force_stop flag is set in the original nlopt::opt, while the check is done for the elimdim wrapper, could this be the issue?

MWE:

double myvfunc(const std::vector<double> &x, std::vector<double> &grad,
               void *my_func_data) {
  throw nlopt::forced_stop();
  return x[0];
}

int main() {
  nlopt::opt opt("GN_CRS2_LM", 2);
  std::vector<double> lb{-10.0, 0};
  opt.set_lower_bounds(lb);
  std::vector<double> ub{10, 1};
  opt.set_upper_bounds(ub);
  opt.set_min_objective(myvfunc, NULL);
  opt.set_maxeval(1000);
  opt.set_xtol_rel(1e-4);
  std::vector<double> x{1.234, 0};
  double minf;
// This works as it should, catching and rethrowing the forced_stop exception
  try {
    opt.optimize(x, minf);
    std::cout << "found minimum at f(" << x[0] << "," << x[1]
              << ") = " << std::setprecision(10) << minf << std::endl;
  } catch (std::exception &e) {
    std::cout << "nlopt failed: " << e.what() << std::endl;
  }
  opt.set_upper_bounds({10, 0});
// This catches but never rethrows the exceptions and continues until maxeval is reached.
  try {
    opt.optimize(x, minf);
    std::cout << "found minimum at f(" << x[0] << "," << x[1]
              << ") = " << std::setprecision(10) << minf << std::endl;
  } catch (std::exception &e) {
    std::cout << "nlopt failed: " << e.what() << std::endl;
  }
}

This returns

nlopt failed: nlopt forced stop
found minimum at f(1.234,0) = inf