Optimization toolbox: "lsqnonlin" stopped because it exceeded the function evaluation limit

Question

Mohammad Heidari 2018-11-20

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/431049-optimization-toolbox-lsqnonlin-stopped-because-it-exceeded-the-function-evaluation-limit

评论： has azz 2020-7-21

Hi everyone,

I am using the lsqnonlin from optimization toolbox for a problem. After 88 itterations I am getting my answeres and this msg:

lsqnonlin stopped because it exceeded the function evaluation limit

options.MaxFunEvals=800

As I am not satisfied with the answeres, I am guessing maybe changing the MaxFunEvals to a higher number can help. Could anyone let me know how I can set it in the toolbox? I am not seeing any parameter like this in the options.

Thank you

1 个评论
显示 -1更早的评论隐藏 -1更早的评论

Walter Roberson 2018-11-20

https://www.mathworks.com/help/optim/ug/current-and-legacy-option-name-tables.html

old name MaxFunEvals

current name MaxFunctionEvaluations

The name to use depends upon which options constructing routine you are using

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

John D'Errico 2018-11-20

2
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/431049-optimization-toolbox-lsqnonlin-stopped-because-it-exceeded-the-function-evaluation-limit#answer_348059

编辑：John D'Errico 2018-11-20

在 MATLAB Online 中打开

Note that I don't know which release of MATLAB you are using. But what did you try? This seems clear:

opts = optimoptions('lsqnonlin')
opts = 
  lsqnonlin options:
   Options used by current Algorithm ('trust-region-reflective'):
   (Other available algorithms: 'levenberg-marquardt')
   Set properties:
     No options set.
   Default properties:
                   Algorithm: 'trust-region-reflective'
              CheckGradients: 0
                     Display: 'final'
    FiniteDifferenceStepSize: 'sqrt(eps)'
        FiniteDifferenceType: 'forward'
           FunctionTolerance: 1e-06
         JacobianMultiplyFcn: []
      MaxFunctionEvaluations: '100*numberOfVariables'
               MaxIterations: 400
         OptimalityTolerance: 1e-06
                   OutputFcn: []
                     PlotFcn: []
    SpecifyObjectiveGradient: 0
               StepTolerance: 1e-06
         SubproblemAlgorithm: 'factorization'
                    TypicalX: 'ones(numberOfVariables,1)'
                 UseParallel: 0

Now, nothing stops you from setting:

opts.MaxFunctionEvaluations = 2000
opts = 
  lsqnonlin options:
   Options used by current Algorithm ('trust-region-reflective'):
   (Other available algorithms: 'levenberg-marquardt')
   Set properties:
      MaxFunctionEvaluations: 2000
   Default properties:
                   Algorithm: 'trust-region-reflective'
              CheckGradients: 0
                     Display: 'final'
    FiniteDifferenceStepSize: 'sqrt(eps)'
        FiniteDifferenceType: 'forward'
           FunctionTolerance: 1e-06
         JacobianMultiplyFcn: []
               MaxIterations: 400
         OptimalityTolerance: 1e-06
                   OutputFcn: []
                     PlotFcn: []
    SpecifyObjectiveGradient: 0
               StepTolerance: 1e-06
         SubproblemAlgorithm: 'factorization'
                    TypicalX: 'ones(numberOfVariables,1)'
                 UseParallel: 0

As you can see, opts now contains the desired value. Pass opts into lsqnonlin to use these settings.

As Walter points out, in older releases, the property name was MaxFunEvals. But the same sort of thing will still apply. You may just need to use the proper property name.

Having said that, frequently when you run out of function evals or iterations, it means your problem is not converging very well. Tht in turn either means that your starting values are particularly poor, or that your problem is fairly ill-posed. In either case, merely increasing the function evals will not really help that much. Essentially, it says that you really need to get better data or a better model or better starting values, each a far better choice than merely increasing the function evals allowed.