A0327
Title: Mixtures of generalised normal distribution with constraints
Authors: Pierdomenico Duttilo - University G. d'Annunzio of Chieti-Pescara (Italy) [presenting]
Stefano Antonio Gattone - University G. d'Annunzio of Chieti-Pescara (Italy)
Alfred Kume - University of Kent (United Kingdom)
Abstract: A family of univariate mixtures of generalised normal distribution with constrained parameters (CMGND) is proposed. Specifically, the location, scale and shape parameters are constrained to be equal across any subset of mixture components. In this way, it is possible to obtain more parsimonious mixture models and to soften the well-known problem of log-likelihood unboundedness. Additionally, an estimation approach is proposed based on the expectation conditional maximization (ECM) algorithm and the iterative Newton-Raphson method used to handle the non-linear iteration equations of the parameters. A simulation study is performed to assess the estimation performance of a two-component CMGND. Findings show that the estimation accuracy of the constrained mixture is higher than the unconstrained mixture model.
