Package dlife.nn

Classes and interfaces that support the creation of neural networks.

See:
Description

Interface Summary

Neuron	Implemented by classes representing the Neurons that make up a NeuralNetwork.
NNDataFilter	Implemented by objects that can be used to filter the inputs or outputs of a NeuralNetwork.

Class Summary

BipolarNeuron	Implementation of a bipolar transfer function for a Neuron.
BoundFilter	An NNDataFilter that bounds its input values to a a specified range.
CMACNeuralNetwork	Implementation of a CMAC (Cerebellar Model Articulation Controller) Neural Network.
ElmanNeuralNetwork	Implementation of a fully connected multi-layer Elman neural network.
FFBPNeuralNetwork	Implementation of a fully connected multi-layer feed forward neural network with back propagation training.
FFNeuralNetwork	Implementation of a fully connected multi-layer feed forward neural network.
NeuralNetwork	Abstract base class that provides common functionality for a variety of neural networks.
NNDataSet	A set of input/output pairs for evaluating or training a NeuralNetwork.
NNDataSetPair	Represents an input/output pair for a NNDataSet.
NNResults	The results of either training or evaluating a NeuralNetwork on a specific NNDataSet.
ShiftAndScaleFilter	A NNDataFilter that shifts and then scales a data value.
SigmoidNeuron	Implementation of a Sigmoid transfer function for a Neuron.
UnipolarNeuron	Implementation of a unipolar transfer function for a Neuron.

Package dlife.nn Description

Classes and interfaces that support the creation of neural networks. This package provides classes that are sufficient for many basic neural network applications. It has also been designed so that it can be easily extended for more advanced applications. The text below describes the overall structure of the dlife.nn package and how its classes and interfaces interact. The examples.dlife.nn.adder package contains a working example that uses the dlife.nn package.

Overview:

The core of the dlife.nn package is formed by one class and one interface:

• Neuron: The Neuron class is an abstract base class for all of the types of neurons that can be used to construct a neural network. The dlife.nn package provides a number of implementations of the Neuron interface: BipolarNeuron, SigmoidNeuron, UnipolarNeuron.
• NeuralNetwork: The NeuralNetwork class is an abstract base class for a variety of different types of neural networks. This base class provides a variety of methods that are common to many types of neural networks. There are methods for assigning filters (e.g. for scaling) to the network inputs and outputs and for evaluating the performance of the network (see below). The details about the structure of the network, the connection weights, how input values are propagated from the inputs to the outputs and whether or not the network can be trained are specified by concrete subclasses of NeuralNetwork. The dlife.nn package provides concrete implementations of the NeuralNetwork class for several types of neural networks: FFNeuralNetwork, FFBPNeuralNetwork, CMACNeuralNetwork, ElmanNeuralNetwork.

Filtering Inputs and Outputs:

The NNDataFilter interface is implemented by classes that filter the inputs to a network before they are propagated through the neurons and/or filter the outputs of the network before they are returned to the client code. For example, a filter on the output can be used to change a sigmoid output in the range [0...1] to the range [-1...1] that may be more suitable for a particular application. The dlife.nn package provides several implementations of the NNDataFilter interface: BoundFilter, ShiftAndScaleFilter.

Data Sets:

The NNDataSet class provide a convenient means for specifying a set of input/output pairs for evaluating or training a NeuralNetwork. The data for a NNDataSet can be read from a file or can be constructed manually using the methods of that class.

Evaluating a Network:

The evaluate method in the NeuralNetwork class provides a convenient way to evaluate the performance of the network. Given a NNDataSet this method will (optionally) display detailed information about the network's performance on each input/output pair and will return a NNResults object containing summary information about its overall performance on the data set.