Hello,
I just faced a small usability issue using NLopt.jl. I believe the issue is common enough (e.g. forum post from Oct 2020 https://discourse.julialang.org/t/nlopt-forced-stop) that it may suggest a better error catching.
To give a bit of background, I had some prior experience with NLopt last year, then a long pause and then yesterday I was studying a fresh optimization problem from scratch (not starting from a working code template). So I was partly a NLopt newcomer. And I get stuck with a :FORCED_STOP return code.
I took me some hours of instrumenting my cost functions (with prints, logging of parameters values in a global array…) to find the silly cause of the error. Here is a simplified version of my problem which shows the issue: a seemingly innocent simple unconstrained derivative-free optimization problem:
using NLopt
function obj_badsig(x)
J = sum(x.^2)
return J
end
opt = Opt(:LN_NELDERMEAD, nx)
x0 = [1., 1., 1.]
xmax = [2., 2., 2.]
opt.lower_bounds = -xmax
opt.upper_bounds = xmax
opt.min_objective = obj_badsig
opt.xtol_rel = 1e-4
(minf,minx,ret) = optimize(opt, x0)
This yields (0.0, [1.0, 1.0, 1.0], :FORCED_STOP). And in fact the objective function is never called.
And the solution is to reread the NLopt.jl README with better eyes: a gradient parameter is required, even when working with gradient-free algorithms:
function obj(x::Vector, grad::Vector)
J = sum(x.^2)
return J
end
which works.
So my question is: is there a better way for NLopt.jl to catch wrong signatures of objective (and constraint) function to report a more informative error message than :FORCED_STOP for absent-minded users like me?