Nonparametric maximum likelihood (NPML) estimation for exponential families with unspecified dispersion parameter φ suffers from computational instability, which can lead to highly fluctuating EM trajectories and suboptimal solutions, in particular when φ is allowed to vary over mixture components. In this paper, a damped version of the EM algorithm is proposed to cope with these problems.

Return to the NUI Galway Mathematics Preprints