Bayesian Interpretation of Meta-Heuristic Optimisation Algorithms

ABSTRACT: In this talk I will discuss the work from my PhD. There exist a large number of meta-heuristic optimisation methods in the literature: genetic algorithms, particle swam optimisers, simulated annealing, etc. These methods are frequently employed to solve high-dimensional, multimodal, non-linear search problems, and have demonstrated impressive results in various settings. However, understanding the behaviour of these optimisation algorithms is far from easy. In what ways are they distinct? How do their various hyper parameters influence their behaviours? Which optimiser is best? What exactly does understanding the behaviour of an optimiser even really mean? I will examine these questions, and in particular I will discuss the No-Free-Lunch theorems for optimisation, and how, through a Bayesian interpretation, all optimisers can be understood as points in a shared representation space. I’ll also talk about stochasticity in optimisation, and the possibilities of learning an optimiser.

BIO: Dr Thomas Joyce did his undergraduate in Mathemathics at the University of Bath, followed by a MSc in machine learning at The University of Edinburgh. He stayed on at Edinburgh to do his PhD under Professor Michael Herrmann at The School of  Informatics. He is currently a post-doc researcher under Dr Sotirios Tsafataris at IDCOM, focusing on applying deep learning to medical image data, with a particular focus on cardiac data.