Unsupervised Learning f or Finite Mixture Models Via Modif ied Gibbs Sampl ing
L Iu · 2009
Si nce t he co nventio nal al go rit hm can not deal wit h t he variable paramet er di mensio n i n t he unsupervi sed lear ni ng of fi nit e mi xt ure mo del s ( FMM) , an unsupervi sed lear ni ng al go rit hm based o n t he mo dified Gi bbs sampli ng scheme i s p ropo sed. The key fo r t he p ropo sed al go rit hm i s to adop t t he co mpo nent management t echniques t hat i ncl ude co mpo nent co mbi natio n and eli mi na2 tio n af ter each co mplet e it erative st ep . The 22no r m of t he differences in t he mean and covariance are used fo r t he co mpo nent co mbi natio n r ule , and t he co mpo nent eli mi natio n r ule i s t hat t he co m2 po nent t hat has t he lea st weight and i s less t han cert ai n t hreshol d will be di scar ded. Si mulatio n result s show t hat t he p ropo sed al go rit hm i s ro bust fo r t he paramet er i nitializatio n and requires f e2 wer p rio r i nfo r matio n fo r t he number of co mpo nent s. The p ropo sed al go rit hm can deal wit h t he variable di mensio n and avoid t he calculatio n of t he j ump p ro babilit y. Mo reo ver , it can estimate t he number of t he co mpo nent s and paramet er s eff ectively.