Prev: mkl_def.dll not a valid image (r14, win2k)
Next: MPC
From: Mike Parham on 3 May 2006 06:24
I am using Matlab's newfftd function to create a time delay neural
network.

Does the network always use the 'current' (in time) input value
during training or do you have to explicitly specify a 0 entry in the
input delay vector to get it to do this?

The issue has not arisen before because I have been trying to learn a
current distinct target variable (at t=0) given separate current and
previous inputs (t = 0,1,2...).

The reason I am asking is that I now want to use past target values
as inputs as well as other current inputs and past values. I want to
make sure my network is not cheating!!!

Finally is it possible to specify different delay vectors for
different input variables?

Thanks for any help / pointers!

Mike
From: Mike Parham on 4 May 2006 05:35
Found some pointers on setting up the delays with newwtd to see
whether the current time step is used or not:

Neural Network toolbox doc has an example showing a Focused Time
Delay Neural Network set-up for predicting the intensity of a laser
in a chaotic state (Used in Santa Fe Time Series Competition). This
uses input delays of 1-8 (t-1,t-2,...t-8) to predict output at
timestep 0 (t=0).

Matlab Solution Number 1-178VI has another example. This one uses
delays 0-2 (i.e. t-0,t-1,t-2) and predicts output at the next
timestep i.e. t+1

I set both up and analysed the resulting NN structures input
layers/weights to ensure whether or not t = 0 is used and so verify
it's not cheating.

Hope this helps anyone else struggling.

I'd still be interested to know if you could have different delay
vectors for different components of the input?

Mike Parham wrote:
>
>
> I am using Matlab's newfftd function to create a time delay neural
> network.
>
> Does the network always use the 'current' (in time) input value
> during training or do you have to explicitly specify a 0 entry in
> the
> input delay vector to get it to do this?
>
> The issue has not arisen before because I have been trying to learn
> a
> current distinct target variable (at t=0) given separate current
> and
> previous inputs (t = 0,1,2...).
>
> The reason I am asking is that I now want to use past target values
> as inputs as well as other current inputs and past values. I want
> to
> make sure my network is not cheating!!!
>
> Finally is it possible to specify different delay vectors for
> different input variables?
>
> Thanks for any help / pointers!
>
> Mike
 | 
Pages: 1
Prev: mkl_def.dll not a valid image (r14, win2k)
Next: MPC