Time variant net: Version #3
Mackey-Glass time variant net #3
As a cross check we delete the delay neurons from the net and keep just the decay
neurons:
Note:
If you do not want to do the net edits yourself you can also load the example net from file:
MackeyGlassTimeVariant3.mbn
This leads to the following results for training and validation:
As we can see the decay neurons significantly improve the training results compared
to the time invariant net. However, the delay neurons seem to be far more important.
This is due to the characteristics of the Mickey-Glass time series: The recent values
provided by the delay neurons are of much higher interest for the prediction of the future
values than the low pass filtered past value that the decay neurons can provide.
In other words the future value mostly depends on the few recent time values and not so
much on the gross time course of the function during the past.
We now want to see if we can further improve the version #1 net (the one with the delay
neurons).
Continue with time variant net version #4
Copyright © 2003 - 2007, Thomas Jetter