library(RColorBrewer)

## Nearest neighbour classifier

nclass <- 4

mu1 <- rnorm(nclass, mean = 0, sd = 10)
mu2 <- rnorm(nclass, mean = 0, sd = 10)
sigma <- 1 / rchisq(nclass, df = 1)

n <- 1000
true.class <- sample(1:nclass, n, replace = TRUE)

x <- rnorm(n, mean = mu1[true.class], sd = sigma[true.class])
y <- rnorm(n, mean = mu2[true.class], sd = sigma[true.class])

d <- cbind(x, y)

plot(d, col = true.class)

## Number of clusters in model

K <- 2

## Initial guesses

m <- d[sample(1:n, K), , drop = FALSE]

## Loop

devAskNewPage(TRUE)
change <- TRUE
old.class <- pred.class <- numeric(n)

while(change)
{
    min.dist <- rep(Inf, n)
    for (k in 1:K) 
    {
        kdist <- sqrt(colSums((t(d) - m[k,])^2))
        pred.class[kdist < min.dist] <- k
        min.dist <- pmin(min.dist, kdist)
    }
    plot(d, col = brewer.pal(8, "Dark2")[pred.class])
    points(m, pch = 16)
    ## update m
    for (k in 1:K) 
    {
        m[k,] <- colMeans(d[pred.class == k, , drop = FALSE])
    }
    change <- !identical(old.class, pred.class)
    old.class <- pred.class
}

devAskNewPage(FALSE)