From the help page (for fitgmdist in R2014b which has replaced gmdistribution.fit):
You can fit a GMM with defined initial values as in the example given:
Mu = [1 1; 2 2; 3 3];
Sigma(:,:,1) = [1 1; 1 2];
Sigma(:,:,2) = 2*[1 1; 1 2];
Sigma(:,:,3) = 3*[1 1; 1 2];
PComponents = [1/2,1/4,1/4];
S = struct('mu',Mu,'Sigma',Sigma,'ComponentProportion',PComponents);
GMModel3 = fitgmdist(X,3,'Start',S);
By default a k-means algorithm is used which includes a random element and is thus not deterministic.
