jordan
Create and train a Jordan network
Jordan networks are partially recurrent networks and similar to Elman
networks (see elman). Partially recurrent networks are useful
when working with time series data. I.e., when the output of the network not
only should depend on the current pattern, but also on the patterns presented
before.
Usage
jordan(x, ...)# S3 method for default
jordan(x, y, size = c(5), maxit = 100,
initFunc = "JE_Weights", initFuncParams = c(1, -1, 0.3, 1, 0.5),
learnFunc = "JE_BP", learnFuncParams = c(0.2),
updateFunc = "JE_Order", updateFuncParams = c(0),
shufflePatterns = FALSE, linOut = TRUE, inputsTest = NULL,
targetsTest = NULL, ...)
Arguments
- x
a matrix with training inputs for the network
- ...
additional function parameters (currently not used)
- y
the corresponding targets values
- size
number of units in the hidden layer(s)
- maxit
maximum of iterations to learn
- initFunc
the initialization function to use
- initFuncParams
the parameters for the initialization function
- learnFunc
the learning function to use
- learnFuncParams
the parameters for the learning function
- updateFunc
the update function to use
- updateFuncParams
the parameters for the update function
- shufflePatterns
should the patterns be shuffled?
- linOut
sets the activation function of the output units to linear or logistic
- inputsTest
a matrix with inputs to test the network
- targetsTest
the corresponding targets for the test input
Details
Learning on Jordan networks: Backpropagation algorithms for feed-forward networks can be adapted for their use with this type of networks. In SNNS, there exist adapted versions of several backpropagation-type algorithms for Jordan and Elman networks.
Network architecture: A Jordan network can be seen as a feed-forward network with additional context units in the input layer. These context units take input from themselves (direct feedback), and from the output units. The context units save the current state of the net. In a Jordan net, the number of context units and output units has to be the same.
Initialization of Jordan and Elman nets should be done with the default init
function JE_Weights, which has five parameters. The first two
parameters define an interval from which the forward connections are randomly
chosen. The third parameter gives the self-excitation weights of the context
units. The fourth parameter gives the weights of context units between them,
and the fifth parameter gives the initial activation of context units.
Learning functions are JE_BP, JE_BP_Momentum,
JE_Quickprop, and JE_Rprop, which are all adapted versions of
their standard-procedure counterparts. Update functions that can be used are
JE_Order and JE_Special.
A detailed description of the theory and the parameters is available, as always, from the SNNS documentation and the other referenced literature.
Value
an rsnns object.
References
Jordan, M. I. (1986), 'Serial Order: A Parallel, Distributed Processing Approach', Advances in Connectionist Theory Speech 121(ICS-8604), 471-495.
Zell, A. et al. (1998), 'SNNS Stuttgart Neural Network Simulator User Manual, Version 4.2', IPVR, University of Stuttgart and WSI, University of T<U+00FC>bingen. http://www.ra.cs.uni-tuebingen.de/SNNS/welcome.html
Zell, A. (1994), Simulation Neuronaler Netze, Addison-Wesley. (in German)
See Also
Examples
# NOT RUN {
demo(iris)
# }
# NOT RUN {
demo(laser)
# }
# NOT RUN {
demo(eight_elman)
# }
# NOT RUN {
demo(eight_elmanSnnsR)
# }
# NOT RUN {
data(snnsData)
inputs <- snnsData$laser_1000.pat[,inputColumns(snnsData$laser_1000.pat)]
outputs <- snnsData$laser_1000.pat[,outputColumns(snnsData$laser_1000.pat)]
patterns <- splitForTrainingAndTest(inputs, outputs, ratio=0.15)
modelJordan <- jordan(patterns$inputsTrain, patterns$targetsTrain,
size=c(8), learnFuncParams=c(0.1), maxit=100,
inputsTest=patterns$inputsTest,
targetsTest=patterns$targetsTest, linOut=FALSE)
names(modelJordan)
par(mfrow=c(3,3))
plotIterativeError(modelJordan)
plotRegressionError(patterns$targetsTrain, modelJordan$fitted.values)
plotRegressionError(patterns$targetsTest, modelJordan$fittedTestValues)
hist(modelJordan$fitted.values - patterns$targetsTrain, col="lightblue")
plot(inputs, type="l")
plot(inputs[1:100], type="l")
lines(outputs[1:100], col="red")
lines(modelJordan$fitted.values[1:100], col="green")
# }