Thomas Prescott
Mathematical Institute, University of Oxford, UK
Multifidelity approximate Bayesian computation (MF-ABC) is a likelihood-free technique for parameter inference that exploits model approximations and low-fidelity simulations to significantly increase the speed of ABC algorithms. We first introduce the multifidelity framework in the simplistic context of rejection sampling ABC, before discussing how to integrate the multifidelity approach with the ABC sequential Monte Carlo (ABC-SMC) algorithm into a novel MF-ABC-SMC algorithm. We show that the improvements generated by each of ABC-SMC and MF-ABC to the efficiency of generating Monte Carlo samples and estimates from the ABC posterior are amplified when the two techniques are used together.