Time variant net: Version #4
Mackey-Glass time variant net #4
We will now try to improve the time variant net #1 by adding differential neurons. These
neurons build the difference between two source neurons. If the two source neurons are
an input and one of its delayed signals then the differential neuron will output a signal that
is proportional to the change ratio of the input neuron. I.e. the neuron will output the
differential signal of the time series.
See the MemBrain help file for more information about differential neurons.
In order to check if this approach helps with the Mackey-Glass problem we modify the net
version #1 to incorporate a differential neuron.
The following picture shows the time variant net #1 again.
Now we do the following.
The net now looks as follows.
Note the newly created differential neuron with the name '(MG(t) - MG(t) DLY 2)/2'. It
represents the difference between neuron MG(t) and its first delay neuron divided by 2.
We now connect this new differential neuron to the hidden and the output layer of the net.
After randomization the net looks like following.
We start the training and look at the results for training...
... and for validation:
Again it is difficult to tell if this has improved the results or not. At least it doesn't seem
to have significant influence.
Finally, we can conclude that the Mackey-Glass time series can be best predicted with
a neural net that incorporates the last few time values of the series derived from delay
neurons.
This example has not been optimized so far. If you like you can try to further optimize the
net, e.g. by adding more delay neurons that reach further into the past or by adding more
neurons in the hidden layer, more layers etc.
<End of tutorial>
Copyright © 2003 - 2007, Thomas Jetter