Is it possible to run the neural netwok fitting app in matlab using a matlab code ?

Hello all,

I have been using the nnstart neural networking fitting app for my current research. I feed the data and sometimes i have to retrain the data to get a good fitting (regression close to 1). And then i test the obtained network using other sample points I have. According to the regression value i get, i decide whether i move on or train the network again. I did this with the application. I would like to know whether its possible to do the same through a matlab code, where i can set conditions, something like (final network regression >0.9) and if it is not met, then to retrain the network again. It could be of great help to know if this is possible as, I can make this complete process automated.

With Regards,

Aravind

Answer 1

[Steven Lord]

I just did a quick little experiment with release R2016b and all the apps that you can launch from the nnstart app allow you to generate MATLAB functions and/or scripts as the last steps in their operation. So you can set up your problem using the app and generate code then modify the code to adjust how the network is created, trained, and simulated.

[Aravind Sasikumar]

Hey Steven Lord,

Thanks for your reply. Yes, we can generate a matlab code for the network generated where i can give any input and get the output which is based on the network already created. I mean, the code generated already has the network defined and i cannot change the code manually to get another network which may give higher regression value.

[Steven Lord]

To see if I understand what you're trying to do, you want to do something like initially generate code that uses a feedfowardnet then modify the code to use a patternnet instead?

If that is a correct (or close to correct) description, that's not something you're going to be able to automate. You can manually modify the call in the generated code that creates the feedforwardnet object to a similar call that creates a patternnet instead. The signatures are close enough that it probably wouldn't be too difficult but different enough that some human intervention will probably be needed.