matlab optimization toolbox "solve" function. How to change solver?

Question

I want to solve the same basic nonlinear minimization using different solvers(e.g. quadprog, fmincon, fminunc)/algorithms with solve function in Matlab's Optimization Toolbox.

Here's what's happening: I try setting the solver and algorithm through an structure using optimoptions function. It goes like this:

options = optimoptions(@fminunc, 'Algorithm', 'quasi-newton')
[S, fval, exitflag] =  solve(nonlinprob, x0,'options', options)

but the function apparently only uses the quadprog solver with interior-point-convex algorithm no matter what. It seems to ignore the different options I'm setting. I'm running Matlab r2018b. The code section brings the optimization problem being handled with example solver and algorithm chosen and what the solve function call is returning

The function is overwriting the options, but I need to force it to do it in different ways. How can I accomplish that?

OptimizationProblem : 

    minimize :
       4*x^2 + 2*y^2 + 4*x*y + 2*y - 1

    subject to cons1:
       x + 6*y <= 2

    subject to cons2:
       -4*x + 2*y <= 0

    variable bounds:
       -10 <= x <= 10

       -10 <= y
x0 = 
    x: -3
    y: 3

x0 = 
    x: -3
    y: 3

options = 
  fminunc options:

   Options used by current Algorithm ('quasi-newton'):
   (Other available algorithms: 'trust-region')

   Set properties:
                   Algorithm: 'quasi-newton'

   Default properties:
              CheckGradients: 0
                     Display: 'final'
    FiniteDifferenceStepSize: 'sqrt(eps)'
        FiniteDifferenceType: 'forward'
      MaxFunctionEvaluations: '100*numberOfVariables'
               MaxIterations: 400
              ObjectiveLimit: -1.0000e+20
         OptimalityTolerance: 1.0000e-06
                   OutputFcn: []
                     PlotFcn: []
    SpecifyObjectiveGradient: 0
               StepTolerance: 1.0000e-06
                    TypicalX: 'ones(numberOfVariables,1)'

   Show options not used by current Algorithm ('quasi-newton')

Your Hessian is not symmetric. Resetting H=(H+H')/2.
Warning: You have passed FMINUNC options to QUADPROG. QUADPROG will use the common options and ignore the FMINUNC options that do not apply.
To avoid this warning, convert the FMINUNC options using OPTIMOPTIONS.
The interior-point-convex algorithm does not accept an initial point.
Ignoring X0.

Minimum found that satisfies the constraints.

Optimization completed because the objective function is non-decreasing in 
feasible directions, to within the selected value of the optimality tolerance,
and constraints are satisfied to within the default value of the constraint tolerance.

<stopping criteria details>
S = 
    x: 0.5000
    y: -1.0000

fval = -2.0000
exitflag = 
    OptimalSolution

I have not had any problems with MATLAB resetting the algorithm when I set it using optimoptions. Can you provide a minimum reproducible example? — tmcg0

obchardon obchardon · Accepted Answer · 2019-09-04T10:27:26

What happens ?

You use the function solve, the function solve decided to call the quadprog function in order to solve your system of equation. But now quadprog receive options intended for fminunc ! So of course quadprog cannot manage those options and select some default options instead.

You need to specifically choose the right solver (in your case fminunc):

options_fminunc     = optimoptions(@fminunc, 'Algorithm', 'quasi-newton')
[S, fval, exitflag] = fminunc(nonlinprob, x0,'options', options_fminunc)
%                        ↑
%                 should be fminunc

matlab optimization toolbox "solve" function. How to change solver?

1 Answers