An issue related to the expectation maximization algorithm for a coin toss experiment

85 Views Asked by At

I just read a very nicely written introduction paper for the expectation maximisation algorithm published in Nature biotechnology by Do and Batzoglou (http://www.nature.com/nbt/journal/v26/n8/full/nbt1406.html).

Regarding the example shown in the paper, which is a coin toss problem, I have simple question, what if I start with $\hat{\theta}_{A}^{(0)}=0.5$ and $\hat{\theta}_{B}^{(0)}=0.5$? Looks like the algorithm doesn't work since the next parameter will remain same.

Am I wrong with such an initial value issue?

or Is this one of the critical issue in the expectation maximisation algorithm?

1

There are 1 best solutions below

0
On BEST ANSWER

You're right. There's an inherent symmetry in the problem that needs to be spontaneously broken, and if you start with a symmetric guess the symmetric update procedure necessarily reproduces the symmetric guess. This not a problem, though, as long as the procedure converges to the same asymmetric solution as long as you choose asymmetric initial values.