Conditional Optimization problem: inf and NaN

Hi guys, I am doing a conditional optimization procedure. My problem is that it says:

'fmincon cannot continue: user function is returning Inf or NaN values.'

Apparently my user function returns a Inf, or NaN! Is there a way to avoid this issue? I am running the optimization 1000 times at once, but it is pretty frustrating when it fails on the half-way!

I tried FunValCheck 'on' without luck.

The fmincon looks: fmincon(@(varrho) nlobj(varrho,...,...),...,...)

The objective function is nlobj. I need to minimize it to catch the 5x1 parameter vector "varrho". nlobj returns a scalar and the gradient of it returns a 5x1 vector. I believe these are NaN or Inf. Is there a way to convert these to, say nlobj=5 and gradient=[5 5 5 5 5]' ?

Or probably this not the best way of doing it ?

hope to hear from someone. Thank in advance and have a good day,

-Martin Bergholt

Answer 1

[Yao Li]

Try other initial values and adjust the tolerance by implementing optimoptions()

[Martin]

Hi Yao, my problem is that I need to try out enormous many initial values ! Tol does not work either. But thx

Answer 2

[Shashank Prasanna]

Don't let your objective function return nans and infs. Put a condition in the objective function to check for nans and infs isnan isinf

Or use a solver that takes care of this automatically:

http://www.mathworks.com/help/optim/ug/choosing-a-solver.html#bsbwxm7

interior-point or sqp

[Martin]

woups, message to you just below

[Shashank Prasanna]

How you wish you take care of inf and nan is up to you. It is your objective function and it depends on your problem and what you are trying to optimize.

If you don't know how, I'd recommend trying my second suggestion, changing the solver. Some solvers are capable of handling infs and nans.

[Shashank Prasanna]

Martin, you may have to provide us with your nlobj to understand how we can modify it to not return inf or nan.

Answer 3

[Martin]

Hi Shashank, Thanks for answer.

I agree that isinf() and isnan() binary identifies the inf and NaN, respectively. But how should such kind of condition be implemented in an objective function? I have pretty hard times in seeing it:) If you can elaborate it would be great

My system looks like

for for ....

[varrho, y] = fmincon(@(varrho) nlobj(varrho,...,...),...,...)

end end ....

where

ret = nlobj(varrho,...,...)

stuff

ret = xyz

It runs thousands of times until the damn message arrives: 'fmincon cannot continue: user function is returning Inf or NaN values.'

- best Martin Bergholt

[Shashank Prasanna]

message to you above

Answer 4

[Martin]

I tried all algorithms. It seems to, that I have to use to trust-region-reflective, because of my gradient function. And I do not apply an constraint function.

I tried all tolerances. I tried to return some numbers which I know is satisfied of the objective function, in case objective function is NaN or inf.

I am lost.

[Shashank Prasanna]

Martin, please reply as a comment to an existing answer instead of creating a new answer which is really a comment. This makes it easier to track the right answer.

[Martin]

Hi Shashank, I am noob here:), - sorry, wont happen again.