I don't think that they would be all that difficult as long as you have a large enough problem.
EM methods for discrete problems like HMM's as well as the closely related variational Bayesian methods depend mostly on counting instances. Indeed, Gibbs sampling on hidden variable techniques depend on the same sort of thing. A good example is the Buntine and Jakulin paper on DCA. Map-reduce is famously good at this sort of counting problem. In general for methods analogous to EM, you will have a map-reduce step for the estimation phase and one for the maximization phase. Both steps are very much like word counting except that it just takes a bit of math to figure out which words you think you are counting. Just like with word counting, if you are doing a tiny example, MR will be much slower. If you working on a very large problem, though, it can be much larger. On 2/2/08 3:48 AM, "edward yoon" <[EMAIL PROTECTED]> wrote: > I thought of Hidden Markov Models (HMM) as absolutely impossible on MR model. > If anyone have some information, please let me know. > > Thanks. > > On 2/2/08, edward yoon <[EMAIL PROTECTED]> wrote: >> I'm newbie in here. :) >> >> -- >> B. Regards, >> Edward yoon @ NHN, corp. >> >
