returns :XTOL_REACHED but doesn't modify the input values

[giadasp]

I don't know what's wrong with my code.
I'm using the low-level nlopt wrapper because I need to define custom functions to be optimized on a vector of variables and JuMP doesn't allow it.
If I choose the algorithm LD_SLSQP I get ROUNDOFF error, instead, if I use LD_MMA I get Xtol and Ftol reached but the values of the variables don't change from the starting values.
Do you know which can be the problem?
I tested all the functions I implemented and all of them work outside the nlopt optimization.

Thank you in advance for your support

Giada

[mzaffalon]

Can you post your code?

As a side remark, unless you are reporting a bug, you should post on discourse, first.

[giadasp]

Since with the same inputs in works in JuMP 0.18.6 I guess it is a bug.
Anyway the code is the following:

X=rand(100,2)
sp=rand(100)
r2=rand(100)
nPar=2
function myf(x::Vector,grad::Vector)
  nPar=size(x,1)
  y=X*x
  if size(grad,1) > 0
    grad= r2 - (sp ./ (1 .+(exp.(.- y))))
    grad= X'* gradend
  end
  z=log.(1 .+exp.(y))
  return sum(r2 .* y - (sp.*z))
end
opt=NLopt.Opt(:LD_SLSQP,nPar)
opt.max_objective = myf
 opt.lower_bounds = bds.minPars
opt.upper_bounds = bds.maxPars
# opt.xtol_rel = 0.00000001
# #opt.maxtime = 10.0
# opt.ftol_rel= 0.00000001
#opt.constrtol_abs = 0.0
pars_i=zeros(nPar)
# #(minf,pars_i,ret) = NLopt.optimize(opt,pars_i)
(minf,pars_i,ret)=NLopt.optimize!(opt,pars_i)

[mzaffalon]

There are two mistakes in the gradient calculation:

• first, grad= r2 - (sp ./ (1 .+(exp.(.- y)))) is a 100 element vector (and so is probably the second assignment to grad), whereas grad should be a vector of 2 elements;
• second, you should assign to grad in place, as in the NLopt.jl tutorial:

grad[1] = ...
grad[2] = ...

What you were doing was creating a new grad variable: you can check that by running myf as follows

grad = zeros(2)
myf(some_value_for_x, grad)
grad

and verify that grad has not been changed.
I haven't used JuMP, but maybe its solvers do not make use of the gradient.

[mzaffalon]

Can this issue be closed?

[giadasp]

Yes. I was assigning values to the grad vector in a wrong way. Thank you for the support! Il mer 18 set 2019, 08:40 Michele Zaffalon <[email protected]> ha scritto:

…

Can this issue be closed? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#136>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AIBXOCRF6VMPLGO2Z4CJY2DQKHENFANCNFSM4IULHNYA> .

[mx-the-gray]

Can I open a related issue?
I am trying the following constrained minimisation problem.

function ps(x::Vector,grad::Vector)
if length(grad)>0
grad[1]=1
grad[2]=0
grad[3]=0
end
return x[1]
end

function ps_con(z::Vector,x::Vector,grad::Matrix,w::Vector)
if length(grad)>0
grad[1,1]=w[1]
grad[2,2]=20x[2]
grad[1,2]=-2x[2]
grad[1,3]=-1
grad[2,1]=w[2]
grad[2,3]=-0.1
end
z[1]=-(x[2]^2-1+x[3])+w[1]x[1]
z[2]=-(-10x[2]^2+0.1*x[3])+w[2]*x[1]
end

opt=Opt(:LD_MMA,3)
opt.min_objective=ps
jopt.lower_bounds=[0,-Inf,-Inf];
opt.upper_bounds=[1,Inf,Inf];

inequality_constraint!(opt, (z,x,g) -> ps_con(z,x,g,[1,1]), [1e-8,1e-8]::AbstractVector)
(minf,minx,ret) = optimize(opt, [1,1, 1])

but now I get:
(1.0, [1.0, 1.0, 1.0], :FORCED_STOP)

[mzaffalon]

There are two errors I can see:

1. the gradient in the constraint function is a 3x2 matrix, but the routine tries to access grad[1,3] and grad[2,3]
2. you are missing a * in the expression for z[1].

Could you please:

1. indent your code and quote it with 3 back ticks
2. post these questions in the Julia mailing list?

[odow]

Closing because this is not a bug in NLopt.

Please ask usage questions like this on the community forum: https://discourse.julialang.org/c/domain/opt/13