Estimation bayésienne approximative par échantillonnage préférentiel

For numerous models, it is impossible to conduct an exact Bayesian inference. There are many cases where the derivation of the posterior distribution leads to intractable calculations (due to the fact that this generally involves intractable integrations). The Bayesian computational literature has been dominated by the development of simulation approximations based on Markov chains, the famous Markov Chains Monte Carlo (MCMC) methods [Robert and Ca- sella, 1999].However, the more complicated the model, (especially if there are many parameters and/or hyperparameters), the more expensive theMCMC approaches become, in terms of time and storage. Moreover, it is difficult to monitor the convergence of Monte Carlo schemes. As an alternative to the use of the MCMC approximations, Cappé et al. [2004] have shown that the notion of importance sampling [Rubinstein, 1981, Ripley, 1987] can actually be greatly generalized to encompassmuchmore adaptive and local schemes than thought previously. This leads to the Population Monte Carlo (PMC) algorithm. We illustrate here the good properties of this scheme on a switching ARMA model [Hamilton, 1988, 1989] for which the MCMC approximations are less satisfactory.