Other Options

If we want to use algorithms other than the default, we can specify them with the -W, -P, and -B parameters and their associated options. Each algorithm we define can be assigned to any number of targets. To run the tutorial experiment with algorithms we specify, we can run SNoW as follows:

> snow -train -I tutorial/traindata.snow -F tutorial/test.net
-W 1.5,0.8,4.0,0.5:0
-P 0.1,4.0,0.20:1

Here, a Winnow algorithm was used for target 0 and a Perceptron was used for target 1.

We could have also executed SNoW like this:

> snow -train -I tutorial/traindata.snow -F tutorial/test.net
-W 1.5,0.8,4.0,0.5:0-1
-P 0.1,4.0,0.20:0-1

This will use both a Winnow and Perceptron on each target, combining the results of the algorithms to calculate a single activation for each target.

Command line parameters can also be specified in an architecture file. This file is specified with the -A parameter, and contains parameters which don't need to be changed frequently. For example, when running an experiment over many datasets, the only parameters which change from dataset to dataset will usually be the -I inputfile and the -F network file parameters. Therefore, we can use an architecture file to save typing most of our command line parameters. Our architecture file might be named ``archfile'', and it might read:

-W 1.5,0.8,4.0,0.5:0-1
-P 0.1,4.0,0.20:0-1
-e count:1
-r 4

Using this architecture file, we have defined two algorithms, set the eligibility threshold to 1, and set SNoW to run through the training data four times. Now we can use these parameters with any data files. We can execute SNoW with the architecture file as follows:

> snow -train -I tutorial/traindata.snow -F tutorial/test.net -A archfile