I have presented my work with Ruth Baker on multifidelity ABC in the Stochastic Systems Biology: Theory and Simulation minisymposium at SMB 2021.
Learning a multifidelity simulation strategy for likelihood-free Bayesian inference.
Abstract: Likelihood-free Bayesian inference is a popular approach to calibrating complex mathematical models typical of biological systems, where likelihoods are often intractable. However, being reliant on repeated model simulation, the complexity that prohibits the likelihood calculation can also cause these methods to suffer from a significant computational burden. Multifidelity inference methods have been shown to reduce this burden by exploiting approximate simulations, such as coarser numerics or lower-dimensional models. By incorporating both high- and low-fidelity simulations, computational savings can be achieved without introducing any further bias in the resulting likelihood-free posterior. Instead, these approaches are forced to trade between reducing computational burden and increasing estimator variance. This trade-off is balanced by optimally assigning a simulation budget between the models at different fidelities. We will discuss how the optimal multifidelity simulation strategy can be learned in parallel with the posterior, and the multifidelity algorithm thus adaptively tuned as the posterior is uncovered.