Probabilistic Multi-State Split-Merge Algorithm for Coupling Parameter Estimates
Juan K. Lin · 2006
A new approach to finding good local maxima of the likelihood function based on synthesizing information from two local maxima is presented. We investigate the coupled EM algorithm (CoEM) for coupling local maxima solutions from two separate EM runs for the multinomial mixture model. The CoEM algorithm probabilistically splits and merges multiple latent states based on conditional independence assumptions and is numerically shown to significantly improve on uncoupled EM or deterministic annealing (DAEM) parameter estimates.