rbftrain: RBF neural network

Description

A simple RBF neural network which suitable for approximation.

Usage

rbftrain(inp,neurons,out,weight=c(),dist=c(),alfa=0.2,it=40,err=0,
        sigma=NaN,online=TRUE,permute=TRUE,visual=TRUE, ...)

Arguments

inp

a matrix that contains one input data in each row.

neurons

the number of neurons in the hidden layer.

out

a matrix that contains one output data in each row.

weight

the starting weights of the network.

dist

the starting distortions of the network.

alfa

the learning-rate parameter of the back-propagation algorithm.

the maximum number of training iterations.

err

the average error at the studying points,if the average error anytime lower than this value,the algorithm will stop.

sigma

the width of the Gauss functions.

online

if TRUE the algorithm will operate in sequential mode of back-propagation,if FALSE the algorithm will operate in batch mode of back-propagation.

permute

if TRUE the algorithm will use a random permutation of the input data in each epoch.

visual

a logical value, that switches on/off the graphical user interface.

...

currently not used...

Value

list with 4 argument
weightthe weights of the network.
distthe distortion of the network.
neuronsa numeric vector with length equals to the number of layers in the network, and the ith layer will contains neurons[i] neuron.
sigmathe width of the Gauss functions.

Details

The function creates an RBF neural network on the basis of the function parameters. After the creation of the network the function trains it using the back-propagation algorithm using the inp and out parameter. This two parameters row number must be the same, else the function will stop with an error message. If you use the weight or dist argument, than that variables won't be determined by random. This could be useful if you want to retrain your network. In that case use both of this two arguments in the same time. The function works with normalized Gauss-functions, which width parameter will be the sigma argument. If you want to give the values, this argument should be a matrix, with rows equal the number of neurons in the first layer, and columns equal the number of neurons in the second layer. If the sigma argument is NaN, then the width of each Gauss function will be the half of the distance between the two nearest training samples times 1,1. If the sigma argument is exactly one number, then all sigma value will be that exact number. The function has a graphical user interface that can be switched on and off, with the visual argument. If the graphical user interface is on, then the function could show the result of the approximation in a co-ordinate system, if it's a function with one parameter. The result of the function is the parameters of the trained RBF neural network. Use the rbf function for information recall.

Examples

Run this code

x<-t(matrix(-5:10*24,1,16));
	y<-t(matrix(sin(pi/180*(-5:10*24)),1,16));
	neurons<-8;
	data<-rbftrain(x,neurons,y,sigma=NaN)
	rbf(x,data$weight,data$dist,data$neurons,data$sigma)