EMGMM |
Expectation-Maximization Algorithm for Gaussian mixture model.
Synopsis:
model = emgmm(X)
model = emgmm(X,options)
model = emgmm(X,options,init_model)
Description:
This function implements the Expectation-Maximization algorithm
(EM) [Schles68][DLR77] which computes the maximum-likelihood
estimate of the paramaters of the Gaussian mixture model (GMM).
The EM algorithm is an iterative procedure which monotonically
increases log-likelihood of the current estimate until it reaches
a local optimum.
The number of components of the GMM is given in options.ncomp
(default 2).
The following three stopping are condition used:
1. Improvement of the log-likelihood is less than given
threshold
logL(t+1) - logL(t) < options.eps_logL
2. Change of the squared differences of a estimated posteriory
probabilities is less than given threshold
||alpha(t+1) - alpha(t)||^2 < options.eps_alpha
3. Number of iterations exceeds given threshold.
t >= options.tmax
The type of estimated covariance matrices is optional:
options.cov_type = 'full' full covariance matrix (default)
options.cov_type = 'diag' diagonal covarinace matrix
cov_options.type = 'spherical' spherical covariance matrix
The initial model (estimate) is selected:
1. randomly (options.init = 'random')
2. using C-means (options.init = 'cmeans')
3. using the user specified init_model.
Input:
X [dim x num_data] Data sample.
options [struct] Control paramaters:
.ncomp [1x1] Number of components of GMM (default 2).
.tmax [1x1] Maximal number of iterations (default inf).
.eps_logL [1x1] Minimal improvement in log-likelihood (default 0).
.eps_alpha [1x1] Minimal change of Alphas (default 0).
.cov_type [1x1] Type of estimated covarince matrices (see above).
.init [string] 'random' use random initial model (default);
'cmeans' use K-means to find initial model.
.verb [1x1] If 1 then info is displayed (default 0).
init_model [struct] Initial model:
.Mean [dim x ncomp] Mean vectors.
.Cov [dim x dim x ncomp] Covariance matrices.
.Priors [1 x ncomp] Weights of mixture components.
.Alpha [ncomp x num_data] (optional) Distribution of hidden state.
.t [1x1] (optional) Counter of iterations.
Output:
model [struct] Estimated Gaussian mixture model:
.Mean [dim x ncomp] Mean vectors.
.Cov [dim x dim x ncomp] Covariance matrices.
.Prior [1 x ncomp] Weights of mixture components.
.t [1x1] Number iterations.
.options [struct] Copy of used options.
.exitflag [int] 0 ... maximal number of iterations was exceeded.
1 or 2 ... EM has converged; indicates which stopping
was used (see above).
Example:
Note: if EM algorithm does not converge run it again from different
initial model.
EM is used to estimate parameters of mixture of 2 Guassians:
true_model = struct('Mean',[-2 2],'Cov',[1 0.5],'Prior',[0.4 0.6]);
sample = gmmsamp(true_model, 100);
estimated_model = emgmm(sample.X,struct('ncomp',2,'verb',1));
figure; ppatterns(sample.X);
h1=pgmm(true_model,struct('color','r'));
h2=pgmm(estimated_model,struct('color','b'));
legend([h1(1) h2(1)],'Ground truth', 'ML estimation');
figure; hold on; xlabel('iterations'); ylabel('log-likelihood');
plot( estimated_model.logL );
See also
MLCGMM, MMGAUSS, PDFGMM, GMMSAMP.
About: Statistical Patte7rn Recognition Toolbox
(C) 1999-2003, Written by Vojtech Franc and Vaclav Hlavac
Czech Technical University Prague
Faculty of Electrical Engineering
Center for Machine Perception
Modifications:
26-jul-07, VF, inconsistent parameter names 'ncomp and 'num_gauss' removed
26-may-2004, VF, initialization by K-means added
1-may-2004, VF
19-sep-2003, VF
16-mar-2003, VF