nonlinear minimization with fminunc

Question

Daniel 2014-7-24

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/143223-nonlinear-minimization-with-fminunc

评论： Shashank Prasanna 2014-7-24

Hi All:

I am doing parameterization by minimizing a nonlinear target function. However, after the iteration runs, it returns the following message. And returns with the initial values for the parameters that I set.

 *Iteration  Func-count       f(x)        Step-size       optimality
     0           9          1.46536                      7.77e+06
     1         144          1.39431    1.57417e-14       5.47e+06

Local minimum possible.

fminunc stopped because the size of the current step is less than the default value of the step size tolerance.*

There is no error in the code. What do you suggest to solve this issue?

Thank you!

0 个评论
显示 -2更早的评论隐藏 -2更早的评论

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Matt J 2014-7-24

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/143223-nonlinear-minimization-with-fminunc#answer_146313

在 MATLAB Online 中打开

Evaluate the gradient at the initial point and see if it is close to zero. Also, call fminunc with all of its output arguments,

[x,fval,exitflag,output,grad,hessian]= fminunc(...)

to get more diagnostic information.

2 个评论
显示无隐藏无

Daniel 2014-7-24

Thanks for your comments.

Gradient: 1.1230 -0.5735 -0.7339 -2.8473 -0.4665 -0.7594 -1.5631 -2.0650

Hessian: -0.3929 -0.0146 1.6484 1.8609 1.4881 2.0857 -0.6735 2.4585 -0.0146 1.9118 4.0316 3.6329 2.6916 2.6806 -0.9732 6.1937 1.6484 4.0316 0.3282 0.9016 1.4807 4.5981 -1.3104 3.2933 1.8609 3.6329 0.9016 3.0804 3.2111 2.4965 1.9743 0.3590 1.4881 2.6916 1.4807 3.2111 0.5078 1.3144 0.0943 1.4038 2.0857 2.6806 4.5981 2.4965 1.3144 2.8224 0.0054 1.7383 -0.6735 -0.9732 -1.3104 1.9743 0.0943 0.0054 1.2775 3.2253 2.4585 6.1937 3.2933 0.3590 1.4038 1.7383 3.2253 2.3437

How do you think?

Thanks!

Matt J 2014-7-24

编辑：Matt J 2014-7-24

I'm guessing your function may not be differentiable. A local minimum of a twice continuously differentiable function should have a positive semi-definite Hessian and gradient near zero. It appears you are far from either.

请先登录，再进行评论。

Answer 2

Shashank Prasanna 2014-7-24

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/143223-nonlinear-minimization-with-fminunc#answer_146326

编辑：Shashank Prasanna 2014-7-24

The optimization stopped because size of the current step is less than the default value. However you can change the defaults.

I suggest you read the following articles in the link below:

http://www.mathworks.com/help/optim/improving-results.html

When the Solver Fails
When the Solver Might Have Succeeded
When the Solver Succeeds

There are guidelines on what you can try in each of the situations.

2 个评论
显示无隐藏无

Daniel 2014-7-24

Thanks Shashank. I actually noticed that and fixed it with changing the default step size. But now, got another message:

fminunc stopped because it cannot decrease the objective function along the current search direction.

... looked online support, feels like there is really no solution for this...

Do you have any idea about this?

Thanks!

Shashank Prasanna 2014-7-24

fminunc is a derivative based optimizer. If you have discontinuous objective surface or have multiple optimums then fminunc becomes sensitive to initial start points. If you do have an exotic objective function I recommend trying multistart or patternsearch which does better at finding "global" optimum solutions.

Local vs Global:

http://www.mathworks.com/help/gads/what-is-global-optimization.html#bsbaljn

请先登录，再进行评论。