havBpNet-J Introduction

havBpNet:J

Intro    Network Classes Overview    Data Classes Overview    Packages

havBpNet:J Home Page

This documentation assumes that the user is familiar with neural networks in general and with the Backpropagation paradigm in particular.

Introduction

havBpNet allows user applications to work at the layer level. User applications are not required to manage and manipulate individual nodes in the network; however, node level manipulation methods are available. havBpNet makes it easy to build networks ranging from simple 3-layer models to more complicated composite models consisting of any number of standard and/or Recurrent Backpropagation networks.

In havBpNet, networks are built by defining several layers and connecting these layers together.

Things like the transfer-function, learning rate, and momentum are defined for each layer individually, thereby allowing maximum control over network operation.

Typical network training consists of

presenting an input pattern (training pattern) to the input layer (or layers),
feeding data forward through the network’s layers,
presenting the “expected” values to the output layer (or layers), and
calculating and propagating errors backward through the layers.

Typical network consultation consists of

presenting input patterns to the input layer(s),
feeding the data forward through the layers, and
retrieving the resulting network response at the output layer(s).

havBpNet allows error calculation, based on presented expected values, even while training is disabled, thus allowing a convenient way of determining how well the network is functioning in consultation mode for known patterns.

havBpNet consists of three main base classes:

havBpLayer
havBpNode

havBpNodeList

The havBpLayer class, along with a few methods from the havBpNodeList class, provides the main application interface. The typical applications developer can do all that is needed using this interface to create, train, save, restore, and consult networks. Only those users who wish to extend or otherwise modify the library will need to become familiar with the havBpNode class and its sub-classes, and the remaining methods of the havBpNodeList class.

The havBpNodeList class’ main purpose is to provide a generic mechanism for resolving references to node instances. It is through an instance of havBpNodeList that layers and nodes accomplish the referencing needed to connect the network. This mechanism also provides the ability to save and later restore a network without concern for memory address resolution of pointers to node instances.

The sub-classes of the havBpNode class are the main workhorses of havBpNet. Most processing takes place in the nodes of a network. A network, once created, consists as a hierarchical lattice of nodes with layers providing only a convenient and efficient application interface for controlling the creation and processing of the lattice. Although user applications are not required to deal with nodes directly, havBpNet does allow direct user manipulation of nodes by use of user callable access methods.

Auxiliary Data-handling Classes: havTdata & havTpat

Several additional classes are supplied with havBpNet. These classes - havTdata and havTpat - are designed to help users handle input and expected data used to train or consult networks built using havBpNet. These data handling classes are not required and are provided as a convenience only.

Intro    Network Classes Overview    Data Classes Overview    Packages