Why does the output of "configure" not take time delays into account for Neural Networks with Delays?

I noticed that with networks that have time-delays, there are errors in the 'xmin' and 'xmax' fields of the networks 'processSettings' field.

This is problematic, because these fields are used by the input layer to map the input between -1 and 1, and if an element held in the delay field is outside of this range, it may lead to inaccurate training.

The example code snippet below illustrates this issue:

[X,T] = simpleseries_dataset;
X{1} = 1000;
net = timedelaynet(1:2,20);
[Xs,Xi,Ai,Ts] = preparets(net,X,T);
net = configure(net,Xs,Ts);
net.inputs{1}.processSettings{2}

ans = 

    struct with fields:

           name: 'mapminmax'
          xrows: 1
           xmax: 0.9706
           xmin: 0.0119
         xrange: 0.9587
          yrows: 1
           ymax: 1
           ymin: -1
         yrange: 2
           gain: 2.0862
        xoffset: 0.0119
      no_change: 0

Answer 1

[MathWorks Support Team]

As of R2017b, this issue is present. 

Note that this issue treats "Xs" as the input dataset, and "configure" does not have any options to treat the delays.

Note that padding the time-series so that extremal values that happen at the beginning are moved to the inside of the training is a viable solution.