Σεμινάριο: "Adaptive Frameworks for Epidemic Dynamics: From Bayesian Retrospective Assessment to Reinforcement Learning and Forecasting"

ΚΥΚΛΟΣ ΣΕΜΙΝΑΡΙΩΝ ΣΤΑΤΙΣΤΙΚΗΣ 2025-2026

Ομιλητής: Petros Barmpounakis, Department of Oncology, University of Cambridge, Cambridge, United Kingdom

Αίθουσα: Θα ανακοινωθεί

ΠΕΡΙΛΗΨΗ

The recent global COVID-19 pandemic has rekindled interest in models that can capture infectious disease dynamics under ever-changing conditions and further highlighted the necessity for flexible, data-driven frameworks to guide policy decisions under uncertainty. This presentation explores three recent methods that employ Bayesian inference and machine learning to different stages of epidemic modeling, ranging from retrospective analysis to prospective policy planning. We discuss hierarchical stochastic epidemic models with piece-wise constant reproduction number, suitable for capturing distinct epidemic phases. The time changes are inferred from data, while the number of phases is allowed to vary as a Poisson point processes or a Dirichlet process. We combine these multiphasic epidemic models with sequential Bayesian inference and reinforcement learning (RL) controllers that adaptively choose intervention levels to balance disease burden, such as Intensive Care Unit (ICU) load, against socio-economic costs. We develop a context-specific cost estimation method via empirical experiments and expert feedback. Two RL policies are implemented: an ICU-threshold rule carried out via Monte Carlo grid search and a policy involving a posterior-averaged Q-learning agent. Moving from discrete phases to continuous forecasting, we employ exchangeable multi-task Gaussian Processes for sharing temporal information and accounting for cross-correlations among different population groups or regions. We analyse distinct types of outbreak data from recent epidemics and find that the proposed methodologies result in improved predictive ability against competing alternatives, demonstrating the power of integrating Bayesian sequential learning with RL for optimisation of epidemic control policies.