ABC World Seminar: Ivis Kerama and Richard Everitt

on the April 23, 2020

at 11:30 am [UK time]
For this second session of the ABC World Seminar, Ivis Kerama from the University of Reading and Richard Everitt from the University of Warwick will talk about "Rare event ABC-SMC^2"

Inspired by the "One World Probability Seminar", we decided to run The ABC World Seminar, a weekly/fortnightly series of seminars that will take place on Blackboard Collaborate on Thursdays at 11.30am [UK time]. The idea is to gather members and disseminate results and innovation during these weeks and months under lockdown.


Ivis Kerama and Richard Everitt 

Abstract
We propose a new algorithm based on a combination of rare event methods for estimating the ABC likelihood in the case of high-dimensional data (Prangle et al, 2018) and ABC-SMC for exploring the parameter space. The new method has a similar structure to SMC2 (Chopin et al, 2012). We demonstrate an improvement in performance (in terms of accuracy and e?ciency) over the rare-event MCMC algorithm in (Prangle et al 2018) for the same (still high dimensional problem) problems. Moreover, we show that the SMC2 algorithm retains all the advantages of regular SMC over a general MCMC, while performing extremely well in higher dimensional problems than the ones initially attacked by the rare-event based method. Furthermore, we show a marked increase in e?ciency compared to ABC-SMC (and adaptive ABC-SMC) without a large additional computational load due to the nested SMC (the “internal”particle SMC can be ?xed to a low number of particles with very little e?ect on the overall accuracy). We also proposed an adaptive scheme for both the sequence of ABC tolerances in the SMC, and also for the MCMC rejuvenation steps of the external parameter space particle ?lter. We demonstrate the above on some toy models and a real world example of an epidemic model. This is joint work with Tom Thorne.
 

 

Published on May 6, 2020