algorithm - 期望最大化算法的数值示例

Question

<分区>

Answer 1

这个怎么样: http://en.wikibooks.org/wiki/Data_Mining_Algorithms_In_R/Clustering/Expectation_Maximization_(EM)#A_simple_example

一年前我也在 (edit)R 中写了一个简单的例子，不幸的是我无法找到它。稍后我会再次尝试找到它。

编辑:这是 -

EM <- function() 
{
    ### Read file, get necessary cols 
    dataFile <- read.csv("wine.csv", head = FALSE, sep = ",")
    sl <- dataFile[, 2]
    #sw <- dataFile[, 3]
    #pl <- dataFile[, 3]
    #pw <- dataFile[, 4]
    class <- dataFile[, 5]
    N <- length(sl)
    pi1 <- 0.5
        ### Init ### 
    rand1 <- floor(runif(1) * N) 
    rand2 <- floor(runif(1) * N) 
    mu1 <- sl[rand1]
    mu2 <- sl[rand2] 
    mean1 <- sum(sl)/N
    sigma1 <- sum(  (sl - mean1) ** 2)   / N 
    sigma2 <- sigma1
    print(mu1)
    print(mu2)
    print(sigma1)
    print(sigma2)
    COUNTLIM <- 10
    count <- 1 
    prevmu1 <- 0.0; 
    prevmu2 <- 0.0; 
    prevsigma1 <- 0.0; 
    prevsigma2 <- 0.0; 
    gamma <- array(0, length(sl)) 
    while (count <= COUNTLIM) 
    { 
        gamma <- pi1 * dnorm(sl, mu2, sigma2)/ ( (1 - pi1) * dnorm(sl, mu1, sigma1) + pi1 * dnorm(sl, mu2, sigma2))
        mu1 <- sum((1 - gamma) * sl) / sum(1 - gamma)

        mu2 <- sum((gamma) * sl) / sum(gamma)

        sigma1 <- sum((1 - gamma) * (sl - mu1) ** 2)/sum(1 - gamma) 
        sigma2 <- sum((gamma) * (sl - mu2) ** 2)/sum(gamma) 
        pi1  <- sum(gamma)/N
        print(c(mu1, mu2, sigma1, sigma2, pi1))
        if (count == 1) 
        {
            prevmu1 <- mu1; 
            prevmu2 <- mu2; 
            prevsigma1 <- sigma1; 
            prevsigma2 <- sigma2; 
        } 
        else 
        { 
            val <- ((prevmu1 - mu1)*2 + (prevmu2 - mu2)*2 + (prevsigma1 - sigma1)*2 + (prevsigma2 - sigma2)*2) ** 0.5; 
            print(c("val: " , val))
            if (val <= 1) 
            {
                break; 
            }
        } 
        count <- count + 1
    } 
    print(mu1)
    print(mu2)
    print(sigma1)
    print(sigma2)
}