MethodError for opt.initial_step

[mzaffalon]

No method is defined when invoking the initial_step property:

julia> using NLopt; opt = Opt(:LN_NELDERMEAD, 2);
julia> opt.initial_step
ERROR: MethodError: no method matching initial_step(::Opt)

Would

initial_step(o::Opt) = initial_step(o, zeros(Float64, ndims(o)))

fix it?

PS: I thought

initial_step(o::Opt) = initial_step(o, Vector{Float64}(undef, ndims(o)))

would work too, but the values I get make no sense.

[stevengj]

We could define initial_step(o::Opt) = x -> initial_step(o, x) so that you could do opt.initial_step(x), I guess, but that's maybe a little too OOP for Julia. Or just remove initial_step from getproperty … it was an oversight, probably, since I forgot that initial_step didn't work like that.

[mzaffalon]

What is wrong with creating the temporary array in my proposal above?

By the way, I also see that you defined opt.inequality_constraint = ..., which sets the tolerance to 0. There are now two methods to define the constraints, using properties and the function call.

I find it a bit confusing having some methods being implemented as function calls only (e.g. ndims(opt)), some as get_property (e.g. opt.min_objective = ...) and some (e.g. inequality_constraints) exposing two different interfaces according to whether one passes the tolerance.

Yes, I know that I can call min_objective(opt, ...) as before.

[stevengj]

What is wrong with creating the temporary array in my proposal above?

Because the initial step in general depends on the initial guess vector, and there is no reason that initial guess vector should be the zero vector.

[mzaffalon]

there is no reason that initial guess vector should be the zero vector.

I forgot about that.