21 Sep 94, 12:00, WeH 1327
Maximum likelihood estimation and the EM algorithm
[See the 1976 Dempster, Laird, and Rubin paper "Maximum Likelihood from
Incomplete Data via the EM Algorithm" for more information. You don't
need to have read this paper beforehand.]
I'll start off by defining maximum likelihood estimation and giving two
simple examples: the sample mean is the MLE of the population mean when
we have normal errors; but if the errors follow a different
distribution with heavier tails, the sample median is the MLE.
In these simple cases, we can find a MLE analytically; but in more
complicated cases, we must resort to numerical methods such as Newton
iteration, Fisher scoring, or an EM algorithm. I will define these
methods, then give three examples of EM algorithms: analyzing survival
data, k-means clustering, and linear regression with missing values.