Optimal Learning Using Monte Carlo Tree Search for Epidemic Control in the Meningitis Belt

Abstract

This thesis explores various decision-making policies that are applicable to the medical field, particularly for infectious disease control. Taking a specific look at meningococcal disease in Nigeria, this thesis first provides a sequential decision model from the point of view of a single entity looking to treat as many people as possible, while taking into account time lost due to travel.

Meningococcal disease is a strain of meningitis that is particularly harmful due to its high fatality rate and ability to morph, such that with each new outbreak, the current vaccines may or may not be effective. Meningococcal disease also tends to sweep an entire region on a cyclic basis, as the seasons change. This thesis develops a basic disease transmission model to cover key features of meningococcal epidemics and to explore the sequential decision problem of combating meningococcal outbreaks in Nigeria, a hotspot for epidemics as a populous country in the African meningitis belt.

This thesis then evaluates Monte Carlo Tree Search (MCTS) as a policy for making optimal decisions in an epidemic. While MCTS is known for its applications in computer-player board games, we will show how MCTS is particularly applicable to infectious disease outbreaks. By evaluating MCTS’ performance in comparison to realistic benchmark policies, this thesis will show how MCTS performs under different conditions, and how its performance can be optimized.