Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                
Scribd
Upload
3K views

Em Algorithm

The document describes using the Expectation-Maximization (EM) algorithm to fit a mixture of two normal distributions to simulated data where some data points were drawn from N(1,1) and others from N(7,1). The EM algorithm iteratively estimates the latent class assignments (E-step) and distribution parameters (M-step) until convergence. It demonstrates the EM algorithm converging to the correct parameters over iterations on two examples of simulated data.

Uploaded by

api-285777244
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
3K views

Em Algorithm

The document describes using the Expectation-Maximization (EM) algorithm to fit a mixture of two normal distributions to simulated data where some data points were drawn from N(1,1) and others from N(7,1). The EM algorithm iteratively estimates the latent class assignments (E-step) and distribution parameters (M-step) until convergence. It demonstrates the EM algorithm converging to the correct parameters over iterations on two examples of simulated data.

Uploaded by

api-285777244
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 4

EM Algorithm

YIK LUN, KEI


set.seed(123)
tau_1_true <- 0.25
x <- y <- rep(0,1000)
for( i in 1:1000 ) {
if( runif(1) < tau_1_true ) {
x[i] <- rnorm(1, mean=1);y[i] <- "heads"
} else {
x[i] <- rnorm(1, mean=7);y[i] <- "tails"
}
}
library(lattice)
densityplot( ~x, par.settings = list(plot.symbol = list(col=as.factor(y))))

0.25

Density

0.20

0.15

0.10

0.05

0.00
0

x
##initial guesses for the distribution parameters
mu_1 <- 0
mu_2 <- 1
##latent variable parameters
tau_1 <- 0.5
tau_2 <- 0.5

10

for( i in 1:10 ) {
## Given the observed data and distribution parameters, what are the latent variables?
T_1 <- tau_1 * dnorm( x, mu_1 )
T_2 <- tau_2 * dnorm( x, mu_2 )
P_1 <- T_1 / (T_1 + T_2)
P_2 <- T_2 / (T_1 + T_2) ## note: P_2 = 1 - P_1
tau_1 <- mean(P_1)
tau_2 <- mean(P_2)
## Given the observed data, as well as the latent variables, what are the population parameters?
mu_1 <- sum( P_1 * x ) / sum(P_1)
mu_2 <- sum( P_2 * x ) / sum(P_2)
print( c(mu_1, mu_2, mean(P_1)) )
}
##
##
##
##
##
##
##
##
##
##

[1]
[1]
[1]
[1]
[1]
[1]
[1]
[1]
[1]
[1]

0.5045618
0.8546336
0.9732251
0.9853947
0.9864849
0.9865811
0.9865895
0.9865903
0.9865903
0.9865904

6.1011529
6.9403680
7.0006108
7.0054109
7.0058260
7.0058624
7.0058656
7.0058659
7.0058660
7.0058660

0.1002794
0.2301181
0.2423406
0.2434347
0.2435309
0.2435394
0.2435401
0.2435402
0.2435402
0.2435402

set.seed(123)
tau_true <- 0.25
x <- y <- rep(0,1000)
for( i in 1:1000 ) {
if( runif(1) < tau_true ) {
x[i] <- rnorm(1, mean=1);y[i] <- "heads"
} else {
x[i] <- rnorm(1, mean=4);y[i] <- "tails"
}
}
densityplot( ~x, par.settings = list( plot.symbol=list( col=as.factor(y) ) ) )

0.25

Density

0.20

0.15

0.10

0.05

0.00
2

x
mu_1 <- 0
mu_2 <- 1
tau_1 <- 0.5
tau_2 <- 0.5
for( i in 1:30 ) {
## Given the observed data and the distribution parameters, what are the latent variables?
T_1 <- tau_1 * dnorm( x, mu_1 )
T_2 <- tau_2 * dnorm( x, mu_2 )
P_1 <- T_1 / (T_1 + T_2)
P_2 <- T_2 / (T_1 + T_2) ## note: P_2 = 1 - P_1
tau_1 <- mean(P_1)
tau_2 <- mean(P_2)
## Given the observed data and the latent variables, what are the population parameters?
mu_1 <- sum( P_1 * x ) / sum(P_1)
mu_2 <- sum( P_2 * x ) / sum(P_2)
print( c(mu_1, mu_2, mean(P_1)) )

}
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##

[1]
[1]
[1]
[1]
[1]
[1]
[1]
[1]
[1]
[1]
[1]
[1]
[1]
[1]
[1]
[1]
[1]
[1]
[1]
[1]
[1]
[1]
[1]
[1]
[1]
[1]
[1]
[1]
[1]
[1]

1.0835357
0.6797230
0.7320122
0.7910984
0.8298998
0.8545108
0.8701122
0.8800221
0.8863270
0.8903429
0.8929026
0.8945350
0.8955764
0.8962408
0.8966648
0.8969354
0.8971081
0.8972184
0.8972887
0.8973336
0.8973623
0.8973806
0.8973922
0.8973997
0.8974045
0.8974075
0.8974094
0.8974107
0.8974115
0.8974120

3.6048714
3.8663167
3.9306341
3.9574819
3.9730967
3.9827182
3.9887344
3.9925240
3.9949222
3.9964445
3.9974127
3.9980293
3.9984223
3.9986729
3.9988327
3.9989347
3.9989998
3.9990414
3.9990679
3.9990848
3.9990956
3.9991025
3.9991069
3.9991097
3.9991115
3.9991126
3.9991134
3.9991138
3.9991141
3.9991143

0.1320495
0.1865272
0.2059336
0.2165093
0.2230743
0.2272189
0.2298464
0.2315159
0.2325783
0.2332551
0.2336866
0.2339618
0.2341373
0.2342493
0.2343208
0.2343664
0.2343955
0.2344141
0.2344260
0.2344335
0.2344384
0.2344414
0.2344434
0.2344447
0.2344455
0.2344460
0.2344463
0.2344465
0.2344466
0.2344467

myEM <- normalmixEM( x, mu = c(0,1), sigma=c(1,1), sd.constr=c(1,1) )

## number of iterations= 21
myEM$mu ## the means of the two distributions

## [1] 0.8974058 3.9991120


myEM$lambda ## the mixing probabilities

## [1] 0.2344461 0.7655539

You might also like