It there any best practice to select initial values when doing maximum likelihood estimation?
For the time being I'm generating series of initial values at random using for example a normal distribution mean=0 and sd = 10, and pick the initial values with the highest maximum likelihood. A question might be how many series to generate to guarantee the global maximum. Yet this approach might be very time consuming.
I recently came across the following paper that suggest some test for global maximum oflikelihood function. But I have not yet implemented it. http://web.eecs.umich.edu/~hero/Preprints/blatt_hero_icassp05.pdf