Time variant feed forward

Advanced Tutorial: Time Series Prediction


This tutorial is intended for users who already worked through the Short Beginner's Tutorial
and who want to get in touch with some more advanced features of MemBrain on a time
series prediction example.


Note that this tutorial does not always fully describe every single step that has to be performed
since it assumes that the user already has basic knowledge of handling neural nets and lesson
data in MemBrain.


In this tutorial the goal is to build up and train a neural net to predict the so called Mackey-Glass

time series three time steps ahead.


The Mackey-Glass is a mathematical series that is chaotic in its long term progression but which is
predictable in the short term.


The following picture shows the Mackey-Glass series for the time values 0 through 302, i.e. the first
303 values of the series.


For the training of the net we will use the values for t = 0 through t = 198 of the Mackey-Glass series where
the input data is the Mackey-Glass time series MG(t) itself and the output is the Mackey-Glass time series value
three times later in time, i.e. MG(t+3).


For validation we use a lesson with the Mackey-Glass time series values for t = 199 to t = 299 as input, again
the output being validated against MG(t+3).


The corresponding lessons can be found in the files


MackeyGlassTrain.mbl and MackeyGlassValidate.mbl


Go to beginning of the tutorial

Copyright © 2003 - 2007, Thomas Jetter