A simple demonstration

Here we demonstrate the effectiveness of the neural network with some "fake" data:

A successful neural network prediction of mock data in which the target in the data is a nonlinear function of the input vector, plus some noise. (Points on the diagonal are a perfect fit).	The singular value decomposition prediction, which shows the best possible linear prediction.

History of the mean square error. The fact that the error in the verification data is only a bit larger than the error in the training data indicates the network' has been trained to the signal rather than the noist. The coefficient alpha used in the minimization scheme is also shown. (The change in the "weights" are proportional to alpha).

Free source code v0.52 nnet.tar.gz (1 mb, includes compiled binaries that may work for you!) is available here for using the backpropagation algorithm to fit a neural network to pairs of input vectors and target vectors. Requirements are a F90 compiler, and numerical recipes. The numerical recipes are needed only for the singular value decomposition prediction, not the neural network. But the numerical recipes software may cost you approximately $100. So grab your advisor's credit card, buy the software, download it, unzip it, untar it, and assign NRROOT in the Makefile to the recipes_f-90 directory. That is all there is to it.

A fortran compiler just might be free for you, if you can grab Intel Fortran Compiler. The usual Unix (Linux for me) software gnuplot, ghostview, gnumake and perl are also needed.

Here is an example of how to use the software:

tar -xvf nnet.tar
cd nnet
ls
cd perl
chmod u+x *.pl
cd ../bin
make nn
make svd

now add nnet/exe and nnet/perl to your path, preferably in your .cshrc

cd ../mock
chmod u+x dothis
dothis 

• Read all the "Readme" files, especially the one in the mock directory.

• The perl utilities have a little help function, e.g.:mseplot.pl -help

• Try hashplot.pl without the -xy option. This is useful only when the scatter plot has lots of overlapping data, which happens quite frequently when making scatter plots with meteorological data (temperature vs. wind direction, for example). The "hash" capabilities are not really needed with this mock data because the points don't overlap.

• The neural network program is designed to work with data scaled between -1 and 1.

You also may want to investigate a neural network AMOS for SFO surface air temperature. To learn more about neural networks, see neural network FAQ. You can also grab my seminar Neural Networks For Fun and Prophet. Comments? bfiedler@ou.edu