What are the signs that my parameter estimator is on the right track before convergence?

I am doing parameter estimation using parameter estimator app

and the optimization is taking hours

but i have noticed as shown in photo attached that EXP (Minimize ) is not converging and Shows NAN , and i gues this is a bad sign

So How to evaluate parameter estimator performance mid-optimization?

[Torsten]

What is EXP ?

[YASSER]

IT is my experiment plot

[John D'Errico]

Basic rule #1: If you are seeing NaNs reported in any place, you have a significant problem. You need to stop it and figure out where and why NaNs are arising.

[Torsten]

IT is my experiment plot

And why do you get NaN for experimental data ? Do the data you want to fit contain NaN values ?

[YASSER]

This is why iam asking about

My data is just numbers

But the optimization is not converging well .

Which means something to be considered or the algorithm is not powerful.

[Sam Chak]

If you attach the data here, someone will be able to check whether the data isNaN or not.

[YASSER]

And why do you get NaN for experimental data ? Do the data you want to fit contain NaN values ?

-------------------------------------------------------------------

Dear Torsten : the Exp (Minimize) your asking about is the name Experiment (see photo)

because when openning the paramter estimator app you name your eperiment using your data and parameters to be estimated

The data is numbers for sure

I guess the problem is more deep and it depends on something like optimization or configuration

To conclude : i have tried the same experiment but with same reduced data and i havent face this issue

Answer 1

[Sam Chak]

Hi @YASSER

In MATLAB, a 'NaN' result is primarily produced when there are indeterminate operations, such as Inf/Inf or 0/0. A secondary occurrence of NaN can result from operations involving NaN, such as 2*NaN or NaN^2. However, even if the data contain 'NaN', we can still plot the data. See the demo below.

Since you are certain that the data contains only numerical values, the third explanation for this occurrence is that extracting measured data from a spreadsheet with unmatched dimensions can also lead to 'NaN', specifically if empty cells are present. The time data comes from internal clock, and the measured data comes from the sensor.

If you have studied instrumentation and control before, you will know that 'no data' can sometimes happen due to malfunctioning sensors, momentary power outages, during hotline maintenance, or human error in data management. Typically, the device should indicate 'no data' at this timestamp.

Data in 'Book1' spreadsheet

Data loaded onto MATLAB

Plotting the data contain 'NaN'.

% time data
t   = linspace(0, 4000, 4001);

% artificially inject a NaN at a specific indexed location
t(2001) = NaN;

% output data
y   = 2300*exp(- ((t - 2000)/800).^2);

disp('Output y(t) at time, t = 2000'), disp(y(2001))

Output y(t) at time, t = 2000
   NaN

% plot 
plot(t, y), grid on
ylim([0, 2500])
xlabel('Time (seconds)'), ylabel('Amplitude')

●

[YASSER]

Well thank you Mr Sam Chak for enlightening me but looke at the data

no problems with data

Look at my data in the attached

[Sam Chak]

You're welcome, @YASSER. I simply explained the possible ways of encountering 'NaN' in MATLAB. You should continue with the investigation. From your attached 'Book.xlsx', could you demonstrate how you plotted the 'Exp' data?

[YASSER]

I have simply Get the product of Vmpp*Impp to get the Pmpp and then plot the time serie data Pmpp (Timemin)

To be noted that the curve is line drawing plot and not stair plot

[Sam Chak]

Hi @YASSER

I plotted the graph, and the product of indeed resembles the 'Exp' curve. If you are certain that the data in 'Exp' became corrupted with NaN after inputting the data into the Parameter Estimator app, please contact Technical Support, as there is nothing further we can improve from the user side.

[YASSER]

Dear sir

The curve you are talking about is experimental data that we have bring from datasheet

But the result that contains NaN is the estimation or the process of the parameter estimator that is based on several factors, among them i mention : the estimation options specified

the curves your are comaring are not same thing

[Torsten]

We don't make progress this way.

If you are allowed to do so, include your data and your MATLAB code so that we can reproduce your problem.

If you are not allowed to do so, you could send your data and your MATLAB code to MATLAB Support - the people there will treat your problem confidentially.

Here is a description on how to export the MATLAB code:

https://uk.mathworks.com/help/sldo/ug/generate-matlab-code-for-parameter-estimation-problems-gui.html

[YASSER]

Thank you for your reply

I do not mind to share my data and matlab simulink

However , it is not easy to test it , since it is computational heavy and take hours if not days

Due to lack in parameter estimator resources (Matwork resources are simple and just for tutorials )

What i need is more experienced person with this app.

File attached

[Torsten]

Getting NaN values in the course of your computation can have many reasons. Maybe your Simulink model gives back model values that are NaN for the parameters supplied by the fitting program. Or the fitting program itself approaches a set of NaN parameters.

I have no experience about coupling Simulink and normal MATLAB, but I guess that there should be a way to survey the parameter values and the model values during the optimization process.

[YASSER]

Thank you sir

If no one reply this question i will accept your answer

Best regards