An alternative to the RJMCMC algorithm
Résumé
This paper studies a new approach for non-constant dimension problems, such as, for example, mixture deconvolution. In a Bayesian framework, Monte Carlo Markov chain (MCMC) algorithms provides an ecient way to optimize the problem. A particular class of these algorithms have been developed since classical MCMC methods (e.g. Metropolis-Hastings or Gibbs) cannot deal with systems whose order may change. The most famous of these algorithms is the so-called reversible jump MCMC which is recalled in this article. An alternative approach to reversible jump is also proposed: it consists in working with a constant dimension model and introducing a variable coding the occurrences of the objects to estimate. It provides an interesting method whose main advantage is to have a faster dynamic behavior than reversible jump MCMC.