NLopt :FORCED_STOP

Specific Domains Optimization (Mathematical)
1

Hello,

I am trying to get NLopt to work but it just stops immediately and gives :FORCED_STOP. I tried the following but it looks like the algorithm isn’t even going into _ff. Any thoughts? Thank you!

function ff(x)
    return sum(x.^2)
    #return sum(abs.(SSresid(mod_pars,x[1],x[2],x[3])));
end
function _ff(x)
    try ff(x)
    catch 
        println("error")
    end
end
opt = Opt(:LN_SBPLX, 3);
opt.xtol_rel = 1e-1;
opt.min_objective = _ff;
(minf,minx,ret) = NLopt.optimize(opt, [1,25,0.1])
2

Read the tutorial: https://github.com/JuliaOpt/NLopt.jl#tutorial. You need to provide the gradient:

using ForwardDiff
using NLopt

opt = NLopt.Opt(:LN_SBPLX, 3);
f(x) = sum(x.^2)
function my_obj(x, g)
    if length(g) > 0
        ForwardDiff.gradient!(g, f, x)
    end
    return f(x)
end
opt.min_objective = my_obj
sol = NLopt.optimize(opt, [1.0, 25.0, 0.1])

You may be interested in https://github.com/JuliaNLSolvers/Optim.jl

2 Likes
3

Giving some additional details – may be helpful to other newbies like me:

I had similar errors and this page was one of the few that popped up. So it may be helpful to post my solution.

I was trying to use NLopt’s LN_BOBYQA to do a gradient free parameter search within box constraints.

I eventually figured it out but it was very obscure. One key thing that is an easy-to-miss requirement is that any function passed to NLopt has to take 2 inputs: (1) the parameter vector, and (2) a gradient function that modifies the input (e.g. g! rather than g).

However, the gradient function is pointless in a gradient-free search. But it still has to be there to get NLopt to work. Here was how I did it - YMMV but this worked for me:

# (Note: I have a pre-existing function, func_to_optimize, that takes 
# input parameter vector x, as well as a variety of constant inputs. 
# This could be anything, it just has to return a value depending on 
# the input parameters in x. (e.g. a -1*log-likelihood, so that by minimizing
# you conduct a Maximum Likelihood search).

# Reduce func_to_optimize() to a single-input function that just takes x
func = x -> func_to_optimize(x, parnames, inputs, p_Ds_v5; returnval="bgb_lnL")

# Create func2, which adds a dummy_gradient! input to make NLopt happy, but then just runs func()
function func2(pars, dummy_gradient!)
	return func(pars)
end # END function func2(pars, dummy_gradient!)

# Define starting guesses at the parameters; check that both func and func2 return log-likelihoods:
pars = [0.9, 0.9]
func(pars)
func2(pars, [])

# Limits on the parameter guesses
lower = [0.0, 0.0]
upper = [5.0, 5.0]

#######################################################
# Set up the optimization
#######################################################
using NLopt
opt = NLopt.Opt(:LN_BOBYQA, length(pars))
ndims(opt)
opt.algorithm
algorithm_name(opt::Opt)
opt.min_objective = func2

# Set the lower & upper bounds on the parameters
opt.lower_bounds = lower::Union{AbstractVector,Real}
opt.upper_bounds = upper::Union{AbstractVector,Real}
opt.lower_bounds
opt.upper_bounds
opt.ftol_abs = 0.00001 # tolerance on log-likelihood

# Run the optimization
(optf,optx,ret) = NLopt.optimize!(opt, pars)
3 Likes