Is it possible with Neural Network

Hi everyone,

I hope you're safe and well.

I was wondering if it's possible to use Neural Network for the following:

I've calculated the volumes (my y) for every row in X, is it possible to use this data to train my network and to predict the volumes for every row in X1?

I'm trying to do this instead of having to do calculations for X1, or at least have a good representation?

I tried using this file here:

https://uk.mathworks.com/matlabcentral/fileexchange/64247-simple-neural-network

But kept getting an error.

Is what I'm thinking, even possible?

Any suggestions or insight would be helpful.

Thank you!

[Mark Rzewnicki]

It looks like you want the neural net to learn to map 6 inputs to 1 output based on 5,000 training examples. Once the network is trained, you want to apply the learned mapping to 15,693 inputs and record the output predictions.

This is absolutely possible. What kind of problems are you having?

[John Doe]

Yes! That's exactly what I want!

I tried looking for ready scripts that I can work on and try to adjust it for my needs because I don't have the tool. So I found this link:

https://uk.mathworks.com/matlabcentral/fileexchange/64247-simple-neural-network

and I tried using it with the data that's attached but it kept giving me an error.

I guess you can say, I'm not sure where to start and when I tried, I worried that it might not work for me, so I need some guidance.

I would really appreciate your help mark!

[Mark Rzewnicki]

Can I ask what this data represents? Seems like there are some serious nonlinearities in here!

[Ameer Hamza]

@John, do you have an explicit mathematical model that can relate the thickness values with the production? If yes, you can use curve fitting to model your system.

[John Doe]

@ Ameer, sadly it isn't explicit

Answer 1

[Ameer Hamza]

If you have Deep learning toolbox, then you can do it in just in few lines of code

net = feedforwardnet([10 10]); % create a feedforward neural network with 2 hidden 
                               % layers 10 neurons in each layer

net = train(net, X', y'); % train the neural network
y_predicted = net(X'); % calculate the prediction on training dataset
mean_square_error = mse(y_predicted, y'); % accuracy on training dataset

% make prediction on unknown dataset
Y1 = net(X1');

[John Doe]

Hi Ameer, thank you for answering my question! However, I don't have the deep learning tool box :/.

Is there there any way around it?

[Ameer Hamza]

@John, I couldn't find an open-source library on feedforward neural network library for MATLAB on file exchange or anywhere else. The one you shared is limited to a classification problem, i.e., when matrix y only contains integers.

[John Doe]

@Ameer, thank for trying to help! It's really appreciated.

I found this link:

https://stackoverflow.com/questions/21998987/programming-a-basic-neural-network-from-scratch-in-matlab

Do you think it'll help?

[Ameer Hamza]

I guess that can help, but it requires implementing the neural network from scratch. Even after that, I am sure the optimizer will fail for a medium-size neural network because, in that question, the OP used fminsearch, which does not have explicit information about gradient about the objective function. The neural network libraries usually calculate an analytical gradient of the whole network, which is why they can train the network so fast. Any other MATLAB optimizer will not have that information. Theoretically, you can provide that information yourself but calculating the gradient of your network, but that is a quite complex task.