AIC in a Gaussian Mixture Regression

Hello everybody, I am trying to fit a gaussian mixture model to a set of predictor variables. I'm not using the built-in functions of matlab. I have six predictor variables to one response value. Each predictor variable is described by 100 observations I want to determine the number of Guassians (clusters) to fit the model. My script uses first an initialization using K-means and then the EM algorithm to calculate the model parameters ( means, covariance and mixing proportions).

How can I calculate the AIC to determine the numbers of Gaussians that bettwer fit my model?

Joaquim

[Sujit Dahal]

Hello Joaquim,

Did you calculate the AIC for your problem. I also have similar problem to yours. If you have the solution could you please provide it.

Thank you

Sujit

Answer 1

[Adam Danz]

This computes the AIC and the AICc using the following defined inputs.

% RSS: Vector; residual sum of squares between your data and the fit for each model
% N: number of data points
% Np: number of model parameters (must be same size as RSS)
AIC=N*(log(RSS/N)+1)+2*(Np+1);
AICc=AIC+2*(Np+1).*(Np+2)./(N-Np-2);

Based on

Hurvich, C. M., & Tsai, C. L. (1989). Regression and time series model selection in small samples. Biometrika, 76(2), 297-307.