How create a time-delay RBF neural network???

Hello,

I am pretty new at the NN topic and trying to get into it. Is there any way to create a time-delay RBF network using the newrb() function? Or do I have to start with narxnet() and edit it proper? And how do I train it then, because for RBF is no backpropergation needed.

I want to use the RBF time-delay net to obtain a acceleration-signal from a position-signal. I do have the time depending data of both signals (input and target) from measurement data. The task is to get the acceleration-signal without any time-delay or even better, to predict the acceleration.

Do you think that time-delay RBF is good for that kind of task?

Answer 1

[Greg Heath]

If you are new at NNs, I recommend first trying to use what MATLAB has to offer. That, in itself, can be an exasperating experience.

Once you have solved that problem, then you can consider modifications.

 [ I N ] = size(input) = ?
 [ O N ] = size(target) = ?

 Significant delays of the autocorrelation functions for each of 
the O outputs = ? 

 Significant delays of the cross-correlation functions for each 
of the O*I output/input combinations = ?

Now you have enough information to determine, by trial and error, which minimal set of delays is sufficient and how few hidden nodes are sufficient for a stable design.

Good Luck.

Thank you for formally accepting my answer

Greg