probability - Using matlab function "pdf" -


I have found a Gaussian mix of 64 dimension distribution object obj and want to keep it in < / P>

However, when I type the pdf function to type pdf (obj, obj.mu (1, :)) This generates very high probability (i.e. 2.4845 a + 06 9) to examine the object

and this does not make sense, because the probability is between zero to one needed.

Is there a problem with my MattelBab?

ps

pdf (obj, obj.mu (1, :) + obj sigma (1,1) * rand ()) create a high probability ( 2.1682e + 06 9) First things first: A probability density function does not always evaluate for 1, it only integrates to 1 on its domain

Besides, what you are seeing (see page 434, figure 9.7) when Gaussian mixing model fittings. Some components collapsing at a data point necessarily lead to variance to go to 0 and PDF to explode. It often appears in the Gaussian mixing model because it is not log-convex and the maximum maxim is local maxime. We try to find a well-behaved local maxim that works well, and singularity is particularly bad case.

When you see it, you can run the algorithm again with different starting points or reduce the number of components you are using, the particular component in the above book. Resetting to a different value is recommended.

Another method would be to use the Bayesian approach by adopting a pre-or regularization period for your criteria, which are foreign standards such as 0 sigma parameters

you indirectly You can control the first part by using different initial values ​​in gmdistribution.fit . For the second part, you can use the regularize argument:

Comments

Popular posts from this blog

excel vba - How to delete Solver(SOLVER.XLAM) code -

github - Teamcity & Git - PR merge builds - anyway to get HEAD commit hash? -

ios - Replace text in UITextView run slowly -