Discussion Lead: Alexandre Bouchard-Côté (UBC)
Topic: How to choose an annealing algorithm
Link: TBD
Abstract: Over the years, several algorithms have been developed to tackle normalization constant estimation. A handful of those have passed the test of time thanks to their capacity to beat the curse of dimensionality in many realistic scenarios: on one hand, Annealed Importance Sampling (AIS) and Sequential Monte Carlo (SMC) methods, and on the other, Parallel Tempering (PT) and Simulated Tempering (ST) algorithms. Indeed many recent developments can be contextualized as members of one of these two families of meta-algorithms.
A priori, these two families of algorithms, AIS/SMC versus PT/ST, appear quite distinct and indeed these communities are largely silos. This leads to an important practical question: for a given problem, which annealing algorithm should be recommended? I will present our work toward tackling this question.