What is the difference between NAR and closed-NAR in time series analysis (in a conceptual level)

What is the difference between NAR and closed-NAR in time series analysis. I know that NAR is neural networks which can be trained to predict a time series from that series past values. but then how closed NAR differ (in a conceptual level) which one is more like the Elman and Jordan networks?

Answer 1

[Greg Heath]

OL NARXNET uses the known training target as an input. Obviously, it is only used for design.

The CL NARXNET replaces the known training target input with a feedback connection from output to input. Therefore, as long as there is an input, an output will be continuously estimated and feedback for future estimation, especially beyond the time that the target is known.

[Mohammed Habboub]

Great.. Therefore (just to be sure), The OL NARXNET is used to test the architecture of the ANN and can't be used for prediction?

But since "The CL NARXNET replaces the known training target input with a feedback connection from output to input." could you kindly clarify why the target (T) is used, in one of your tutorials , in the preparation for the close loop, as shown in the code below?

   [ netc Xci Aci ] = closeloop(net,Xsi,Asi);
   view(netc)
   [Xc,Xci,Aci,Tc] = preparets(netc,{},{},T);
   [ Yc Xcf Acf ] = netc(Xc,Xci,Aci);
   Ec = gsubtract(Tc,Yc);

[Greg Heath]

NOTE: MATLAB does not mention this!!!

I recommend testing the CL net on the data used to design the OL net.

If the CL net can not even perform well using the original input data, then how in the world is it expected to perform well on any other data?

If you read some of my posts carefully, you will find my recommendation that: if this result is not acceptable, then I recommend using train on the CL net. If this result is still unacceptable, then go back to square one and find another acceptable OL design.

Sometimes you just need another random weight initialization. Other times you may have to change the narxnet (ID,FD,H) inputs as well.

It is unfortunate that this is not mentioned in the documentation.

Greg

[Greg Heath]

WHOOPS!!! Your original question was about NARNET. However, I responded with an answer about NARXNET!

 I apologize.

However, the basic principles are the same.

To help clarify matters, compare the open and close loop diagrams of ALL of the timeseries nets.

CL nets are designed using OL nets as a first step because designing a single CL net from scratch takes a much, much, longer time. Therefore, if you have to design multiple nets to luck into a good set of initial random weights, OL leaves more time for you to brush your teeth before you go to bed.