Machine learning applied to prediction, control and planning from dynamic epidemiological models

We are ever aware of the global impact of infectious disease transmission in shaping the reality of the world around us. For much of human existence transmitted infections have been the principle cause for loss of human life. Through their very nature, the transient threats of pandemics or endemic disease transmission remain a constant challenge to epidemiologists and are also at the forefront of shaping public policy decisions. In order to bring the benefits of Machine learning (ML) to tasks such as optimal prediction, control and planning from epidemiological models of disease transmission, we need to define how principles, advancements and knowledge can be directed towards real-world progress. This work outlines steps towards making better decisions based on learn- ing from epidemiological models under uncertainty. It considers how we can automate and scale epidemiological model value using Machine learn- ing. The thesis investigates a range of methods from Multi-Armed Bandit approaches to sequential decision making using Reinforcement learning (RL) based on simulation. We use Gaussian Process Regression for sample efficiency in learning value-functions, improving model predictions via Bayesian optimisation of epidemiological model parameters and develop surrogate model descriptions for sharing, abstraction and combination of models as learning environments. Finally, the necessary computational infrastructure for sharing and running models at scale is considered.