R/perceptrain.R

In freestats: Statistical algorithms used in common data mining course

#' @title An original perceptron algorithm
#' @description Train data with perceptron algorithm
#' @export perceptrain
#' @return \item{z}{Normal vector of a hyperplane}
#' \item{Z_history}{Trajactory of normal vector of a hyperplane}
#' \item{NumofIteration}{Number of iterations for algorithm}
#' @author Xiaoyao Yang
#' @details 
#' S is especially designed for perceptron.
#' 
#' For more information \code{\link{fakedata}}
#' @param S Each row represents a data points with last column equal to 1; S=[X,1]
#' @param y Class label for data points in S
#' @param alpha_k The speed of converge
#' @param endcost The termination condition of cost function
#' @examples
#' 
#' set.seed(1024)
#' z <- runif(n=3)
#' mydata <- fakedata(w=z,n=100)
#' r <- perceptrain(S=mydata$S,y=mydata$y,alpha_k=1,endcost=0)
#' r
#' 



perceptrain<-function(S,y,alpha_k=1,endcost=0){
    d<-dim(S)[2]-1;n<-dim(S)[1]
    rbind(1,t(S)[1:d,])->x.matrix
    z<-x.matrix[,2]
    Z_history<-matrix(0,0,ncol=d+1)
    
    NumofIteration=0
    Cost.gradient=10000
    while (sum(Cost.gradient^2) > endcost){
        index1 <- classify.pti(S,z)!=y
        x.matrix -> cost.temp
        for (i in 1:n){
            cost.temp[,i] <- index1[i]*x.matrix[,i]*(-y[i])
        }
        Cost.gradient <- apply(cost.temp,MARGIN=1,sum)
        Z_history <- rbind(Z_history,z)
        z <- z - alpha_k*Cost.gradient
        NumofIteration <- NumofIteration + 1
    }
    res <- list(z=z,Z_history=Z_history,NumofIteration=NumofIteration)
    class(res) <- 'pt'
    return(res)
}

classify.pti<-function(S,z){
    #z<-c(-c,Vh) x<-c(x,1)
    d<-length(z)-1;
    yy<-as.vector(sign(z%*%rbind(1,t(S)[1:d,])))
    return(yy)
}