NAME

Math::Gann - Gann design philosophy (package Neural(::.*)*)

AUTHOR

Tuomas J. Lukka (Tuomas.Lukka@Helsinki.FI)

DESCRIPTION

This manpage is an attempt to describe the philosophy behind Gann (Neural). While very much incomplete at the moment, I hope it gives some idea.

For instructions on using Gann, see Neural, Neural::TD, Neural::Gui, and Neural::Gui::Examples.

Classes

An attempt has been made to separate the network topology from the actual data. This way, the network behaves logically as a function of many variables (the weights and the inputs etc.) The primary advantage is that debugging is fairly easy and that the weights and the internal state of the network are easily accessible.

The classes and their relationships are as follows:

Neural::Vector

A small vector class, designed for speed and compactness.

Neural::Net

The class that holds the network topology. Note that this class does not contain the weights, which are given to the methods of this class in an external vector.

Neural::Pars

A convenient set of vectors made to a suitable size for a class. Contains all the other necessary vectors except Input, Output and Target.

Neural::Examples

Stores a set of examples for teaching the network. Also does the evaluation of the error function for a set of weights given as a parameter.

Neural::Minimizer

A minimizer. Calls a given function to evaluate a function and its derivative and tries to minimize the function. Contains some subobjects that define the actual minimization type; see Neural

GUI classes

The graphical user interface of GANN is composed of several different types of objects that can be put together in a variety of interesting ways. In the division into objects an attempt has been made to make everything that might have to be replaced replaceable.

The classes are detailed in Neural::Gui and Neural::Examples.

Network building

The interface for building networks in the C code is as follows: first, an empty network with n inputs is created.

        $net = new Math:Neural::Net($n)

The low-level network code

The low-level network code is in gann_gen.c and gann_gen.h. It is possible to easily add new neuron types to design different kinds of networks.

For efficiency, the following design decisions have been made: In Gann, the abstraction level of the neurons are Groups. A group has a set of weights. Inside the group, any arrangement of weight sharing is possible.

Defining new group types

Group types are defined by writing a definition such as those in the file grouptypes and compiling it.