Model to predict surface movements using Neural Networks

Hello,

I have a 3-D surface that moves over time. The surface is defined in N (natural numbers) so it is not really a surface but a set of points. The domain of the function is: 1=<x<=300 1=<y<=300

I have collected many empirical data to build the model. I have (x,y,z) for t=1:200.

So a total of 300*300*200 = 18*10^6

I am interested in building a model to predict the movements of the surface, I have no idea of the relation between INPUTS (x,y) and OUTPUT (z).

I have 2 questions: 1) What would you try?

2) Considering that I think exist a sort of autocorrelation in z, I am thinking of using neural networks, but I don't know hot to pass the inputs/output variables.

I have tried to use the time series tool, building a cell array, with 200 cells containing a 300*300 array. However Matlab runs out of memory (I have 8 GB).

So how should I pass my inputs/outputs?

Answer 1

[Greg Heath]

If the motion is force free, x,y,z(x,y,t-2),z(x,y,t-1) as inputs and z(x,y,t) as output.

If the motion is forced you will need more delayed inputs. Maybe just z(x,y,t-3)

I would first try a 4-H-1 FITNET (unforced) or 5-H-1 FITNET (forced)

Then I would try a 2 input NARXNET with feedback delays but no input delays.

Reduce the size of (x,y) plane so that you can test the concept.

Then deal with size problems afterward.

Hope this helps.

Thank you for formally accepting my answwer.

Greg

[condor]

Sorry for my delay but I am trying to understand the problem... What do you mean with "force free" ??

[Greg Heath]

No external force is applied.