Population Stochastic approximation Monte Carlo Approximate Bayesian Computation

Approximate Bayesian Computation ABC methods aim at simulating from posterior distributions whose likelihood function is computationally intractable and provided that a sample can be practically drawn from the sampling distribution. Several commonly used ABC versions such as ABC-MCMC, suffer from local trapping problems in regions of low probability which are caused when the tolerance in the distance function is not properly adjusted. As a result there is an undesirable mismatch between the generated sample and the posterior distribution of interest. This project aims at using Stochastic approximation Monte Carlo ideas to overcome such problems, and population ideas in order to enhance stability of the proposed procedure.