How to initialize the parameter mu to zero of a Gaussian Mixture model ?

Question

A. P. B. 2017-6-28

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/346538-how-to-initialize-the-parameter-mu-to-zero-of-a-gaussian-mixture-model

回答： Richard Marveldoss 2017-7-12

X = mat_cell;

        [counts,binLocations] = imhist(X);
        stem(binLocations, counts, 'MarkerSize', 1 );
        xlim([-1 1]); 
        % inital kmeans step used to initialize EM
        K = 2;               % number of mixtures/clusters
        rng('default');
        cInd = kmeans(X(:), K,'MaxIter', 75536);
        % fit a GMM model
        options = statset('MaxIter', 75536); 
        gmm = fitgmdist(X(:), K,'Start',cInd,'CovarianceType','diagonal','Regularize',1e-5,'Options',options);

Here X is the input matrix and I would like to initialize only the sigma parameter of the GMM to zero. How can i proceed?

0 个评论
显示 -2更早的评论隐藏 -2更早的评论

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Richard Marveldoss 2017-7-7

1
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/346538-how-to-initialize-the-parameter-mu-to-zero-of-a-gaussian-mixture-model#answer_273286

在 MATLAB Online 中打开

I see that you have used the 'Start' argument to initialize your Gaussian Mixture model. Instead of using 'Start' with an array that contains mapping of the observation to its label in which case you cannot specify mu and sigma as they are computed from the result of kmeans operation. You can try to explicitly specify the mu and sigma parameter by using a structure whose field elements are mu and sigma. Also, the documentation for the fitgmdist contains an example where the mu and sigma are initialized manually. Link to the documentation: https://www.mathworks.com/help/stats/fitgmdist.html. Additionally, I have attached an example code based on your code to get you started where I have assumed a random dataset for X.

dataLength = 5000;
muData = [5 30];
stdData = [4 10];
X = [muData(1) + stdData(1)*randn(dataLength/2,1);  muData(2) + stdData(2)*randn(dataLength/2,1)];
[counts,binLocations] = imhist(X);
stem(binLocations, counts, 'MarkerSize', 1 );
xlim([-1 1]); 
% inital kmeans step used to initialize EM
K = 2;               % number of mixtures/clusters
rng('default');
[~,cInd.mu] = kmeans(X(:), K,'MaxIter', 75536);
% fit a GMM model
gmInitialVariance = 0.1;
initialSigma = cat(3,gmInitialVariance,gmInitialVariance)
% cInd.mu=[0;0];            %%if you want to initialize the mu parameter to 0.
cInd.Sigma=initialSigma;
options = statset('MaxIter', 75536); 
gmm = fitgmdist(X(:), K,'Start',cInd,'CovarianceType','diagonal','Regularize',1e-5,'Options',options);

1 个评论
显示 -1更早的评论隐藏 -1更早的评论

A. P. B. 2017-7-8

This is for univariate gaussian distribution?? How can it be done for multivariate gaussian distribution

请先登录，再进行评论。

Answer 2

Richard Marveldoss 2017-7-12

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/346538-how-to-initialize-the-parameter-mu-to-zero-of-a-gaussian-mixture-model#answer_273898

For a multivariate Gaussian distribution, the manner in which you construct your input datasets will vary. For example ,the below documentation contains a example where a dataset is generated using two bivariate Gaussian distribution. https://www.mathworks.com/help/stats/fitgmdist.html#examples