dlife.nn
Class ElmanNeuralNetwork

java.lang.Object
  dlife.sys.SerializationBase
      dlife.nn.NeuralNetwork
          dlife.nn.FFNeuralNetwork
              dlife.nn.ElmanNeuralNetwork

All Implemented Interfaces:

Serializable

public class ElmanNeuralNetwork
extends FFNeuralNetwork

Implementation of a fully connected multi-layer Elman neural network. The Elman neural network contains a set of recurrent connections from the last hidden layer to the input layer. These connections allow the network to develop a sense of time, where the output is determined not only by the current input but by past inputs and outputs.

Care must be taken in training an Elman neural network as the sequence of training inputs is relevant. The actual inputs to the network for each training pattern include those in the training pattern as well as the outputs of the last hidden layer as computed for the previous training pattern.

For fuller descriptions of an Elman neural network see:

• Elman, J. L. (1990) Finding structure in time. Cognitive Science, 14(2):179--211.
• Elman, J. L. (1991) Distributed representations, simple recurrent networks, and grammatical structure. Machine Learning, 7:195--224.

Version:

Mar 2, 2009

Author:

Lewis Flanagan

See Also:

Serialized Form

Field Summary

protected int

numInputs

Fields inherited from class dlife.nn.FFNeuralNetwork

ffweights, layerInputs, layerOutputs, nType

Fields inherited from class dlife.nn.NeuralNetwork

filteredInputs, filteredOutputs, rawInputs, rawOutputs

Constructor Summary

ElmanNeuralNetwork(double[][][] ffweights, Neuron nType)
Construct a new ElmanNeuralNetwork by providing the weight matrix for the network.

ElmanNeuralNetwork(int[] layers, double[] weights, Neuron nType)
Construct a new ElmanNeuralNetwork using the provided 1d array of weights.

ElmanNeuralNetwork(int[] layers, Neuron nType)
Construct a new ElmanNeuralNetwork with the specified number of neurons in the layers and using the specified type of Neuron.

Method Summary

void	clearContext() This method clears the data stored in the context units by setting them to 0.
int	numInputs() Get the number of inputs that there are in this ElmanNeuralNetowrk.
void	propagate() Propagate the current filtered input values through the network.

Methods inherited from class dlife.nn.FFNeuralNetwork

getConnections, getFFWeights, getFFWeights1d, getLayerInputs, getLayerOutputs, getWeights, makeLayerInputs, makeLayerOutputs, numOutputs, randomizeWeights, read, setInputs

Methods inherited from class dlife.nn.NeuralNetwork

evaluate, evaluate, getFilteredInputs, getFilteredOutputs, getInputFilter, getInputFilter, getOutputFilter, getOutputFilter, getRawInputs, getRawOutputs, setInputFilter, setInputFilter, setOutputFilter, setOutputFilter, setOutputs

Methods inherited from class dlife.sys.SerializationBase

write

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Field Detail

numInputs

protected int numInputs

Constructor Detail

ElmanNeuralNetwork

public ElmanNeuralNetwork(int[] layers,
                          Neuron nType)
                   throws IllegalArgumentException

Construct a new ElmanNeuralNetwork with the specified number of neurons in the layers and using the specified type of Neuron.

Parameters:

layers - an array of ints specifying the number of neurons in each layer of the network.

nType - the type of neuron in the network.

Throws:

IllegalArgumenException - if the network does not have at least one hidden layer.

IllegalArgumentException

ElmanNeuralNetwork

public ElmanNeuralNetwork(double[][][] ffweights,
                          Neuron nType)
                   throws IllegalArgumentException

Construct a new ElmanNeuralNetwork by providing the weight matrix for the network. The format of the weight matrix for a ElmanNeuralNetwork is:

• Layers are numbered left to right starting with 0.
• Neurons are numbered top to bottom starting with 0.
• If weights is accessed with [i][j][k] then:
  • i is the layer of weights.
  • j is the neuron receiving the activation.
  • k is the neuron whose output is being used.

Therfore:

• Weights[i] holds a 2d array of the input weights for the neurons in the i+1st layer.
• weights[i][j] holds a 1d array of the input weights for neuron j in the i+1st layer. These weights include a bias weight. Thus, there will be one more weight than there are neurons in layer i.
• weights[i][j][k] holds the weight between output of the k-1st neuron in the ith layer and the input of the jth neuron in the i+1st layer.

Note: This constructor assumes that the context units have already been created and are included in the weights matrix.

Parameters:

ffweights - the weights matrix.

nType - the type of Neuron to use for this network.

Throws:

IllegalArgumentException - if the ffweights do not specify a fully connected network.

IllegalArgumenException - if the network does not have at least one hidden layer.

ElmanNeuralNetwork

public ElmanNeuralNetwork(int[] layers,
                          double[] weights,
                          Neuron nType)

Construct a new ElmanNeuralNetwork using the provided 1d array of weights. The order of the weights in the 1d array is expected to be in the same order as that returned by the getFFWeights1d method.

Parameters:

layers - an array of ints specifying the number of neurons in each layer of the network.

weights - the weights for this network.

nType - the type of Neuron to use for this network.

Throws:

illegalArgumentException - if ffweights does not contain the correct number of weights.

IllegalArgumenException - if the network does not have at least one hidden layer.

Method Detail

numInputs

public int numInputs()

Get the number of inputs that there are in this ElmanNeuralNetowrk. This does not include context units.

Overrides:

numInputs in class FFNeuralNetwork

Returns:

the number of inputs.

propagate

public void propagate()
               throws IllegalStateException

Propagate the current filtered input values through the network. A call to this method causes the raw network outputs to be computed. Those raw outputs are then also filtered using the output filters to compute the filtered outputs. Note that the outputs of the last hidden layer from the last call to propagate will become the additional inputs to the current call.

Overrides:

propagate in class FFNeuralNetwork

Throws:

IllegalStateException - if this method is called before the inputs have been set by a call to setInputs.

clearContext

public void clearContext()

This method clears the data stored in the context units by setting them to 0.