On the Convergence of Iterated Random Maps with Applications to the MCEM Algorithm
Gersende Fort, Éric Moulines, Philippe Soulier · COMPSTAT · 1998
Optimization via simulation is a promising technique to solve maximum likelihood problems in incomplete data models. Among the techniques proposed to date to solve this problem, the MCEM algorithm proposed by Wei & Tanner (1991) plays a preeminent role. Perhaps surprisingly, very little is known on the convergence of this algorithm and on the strategies to monitor this convergence. A particular emphasis is given on the stability issue (which is not guaranteed in the original proposal by Wei & Tanner, 1991). A random truncation strategy, inspired by Chen’s truncation method for stochastic approximation algorithms, is proposed and analysed.