Generalized multiple-point Metropolis algorithms for approximate Bayesian computation

Abstract

It is well known that the approximate Bayesian computation algorithm based on Markov chain Monte Carlo methods suffers from the sensitivity to the choice of starting values, inefficiency and a low acceptance rate. To overcome these problems, this study proposes a generalization of the multiple-point Metropolis algorithm, which proceeds by generating multiple-dependent proposals and then by selecting a candidate among the set of proposals on the basis of weights that can be chosen arbitrarily. The performance of the proposed algorithm is illustrated by using both simulated and real data.

Keywords:

• approximate Bayesian computation
• generalized lambda distribution
• likelihood free method
• Markov chain Monte Carlo
• multiple-point Metropolis