Using Neural Network for data interpolation

I am working on a research project that involves interpolating known Monte Carlo data to approximate data for unknown parameters. The known data is in the form of a 1750x3 array of independent variables (1750 sets of atomic number, electron energy, and depth) and a 1750x1 array containing 1750 samples of the charge deposited in the material. I used nftool to create and train a network using these data sets and the resulting network does not approximate the data at all.

In the image above, the blue line is a plot of charge deposition vs depth of the actual data, while the red line is the output of the neural network when the independent variables corresponding with the blue line data are input (so they should be exactly the same). I have tried training the network with a large array of hidden neurons. I'm a physics major, not compsci, so I know next to nothing about why this would happen/how to fix it.

Answer 1

[Greg Heath]

Use the command line approach

 help FITNET

 doc FITNET

Search the NEWSGROUP and ANSWERS using subsets of

 greg fitnet tutorial

Hope this helps.

Greg

[Shane]

I had actually checked the documentation for fitnet, and it doesn't give much troubleshooting help. I did a search for the various questions you had answered regarding neural networks and tried to adapt some of the advice given to my application and I still got the same linear result. Do you have any slightly more specific advice for this situation?

-Is it possible that a neural network interpolation simply won't work for this data set? Or is it just a matter of finding the proper configuration?

-Should I try to just keep adding hidden layers? I have tried up to a few hundred, would going even higher possibly fix the fit?

-What proportions would you use for training/validation/testing for this size of data set?

Thanks, Shane

[Greg Heath]

You made one or more mistakes.

How do you suggest we help you further?

[Greg Heath]

That was a rhetorical question.

 Post relevant code and error messages.

[Peter Hristov]

Why is this answer accepted?

[Greg Heath]

Good question.