The public PhD defense Guiding the mitigation of epidemics with reinforcement learning will take place online on Wednesday, April 15, 2020 at 13:00 CET. It will be broadcasted live on YouTube. Click here to access the stream.
Please note that if anything would go wrong when broadcasting, the URL will be updated on this page. After the presentation and the question session by the jury, the public will be invited to ask questions. These questions can be sent to this email address (firstname.lastname@example.org), and a selection of the questions will be read by the president Prof. Wolfgang De Meuter to Pieter.
The PhD dissertation can be found here.
Abstract of the PhD research:
Epidemics of infectious diseases are an important threat to public health and global economies. To properly understand epidemic processes, and to study emergency scenarios, epidemiological models are necessary. Yet, the development of prevention strategies, which need to fulfil distinct criteria, remains challenging. For this reason, it is important to study how optimization techniques can be used to support decision makers. In this thesis we explored an approach based on reinforcement learning, a field within artificial intelligence, to automatically learn prevention strategies.
We investigated two main lines of research. Firstly, we studied the decision problem where a number of prevention strategies has been defined, and the decision makers need to determine which of the strategies is most efficient, by evaluating the models in a complex and computationally demanding epidemiological model. To perform this evaluation efficiently, we investigate the use of algorithms in the field of reinforcement learning that are grounded in the Bayesian uncertainty framework. This approach enables us to learn faster and to quantify the uncertainty of the decisions. To make this possible, we adapt existing algorithms and created new algorithms. Furthermore, we provide theoretical insights in how these algorithms operate. Secondly, we extend this approach such that we can learn adaptive strategies in an epidemiological model. This means that, rather than comparing preventive strategies, we attempt to learn which subsequent steps are necessary to act optimally over time, i.e., by considering the state of the epidemic. Since the state space of the epidemiological models that are necessary to investigate versatile prevention strategies is huge, we need to represent this space such that the reinforcement learner can use this in its learning process. To do this, we use deep reinforcement learning.
We evaluate both research trajectories in the context of pandemic influenza, a pathogen that has made many victims in the past. Our experiments show that our first research trajectory is very useful to evaluate prevention strategies. Furthermore, we show that these techniques are also useful to support other complex decision problems that involve computationally demanding models. In the experiments to validate the second research trajectory, we created a specific model to investigate school closure policies in case an influenza pandemic emerges. Through experiments, we show that our learning technique approximates the optimal strategy. Finally, we investigate whether there is a collaborative advantage when designing school closure policies. We formulate this research question as a multi-agent problem and solve it using deep multi- agent reinforcement learning techniques.
I am a PhD student at the department of computer science of the Vrije Universiteit Brussel (Brussels, Belgium), under the supervision of Prof. Ann Nowé and co-supervision of Prof. Philippe Lemey (KU Leuven—University of Leuven) and Prof. Bernard Manderick.
In my research, I investigate the use of epidemiological models to study epidemics of persistent infectious diseases (e.g. HIV, human papilloma virus, hepatitis C virus). I will apply machine-learning techniques to find optimal preventive strategies in such models. This research will support public health efforts to reduce the prevalence of pathogens causing persistent infectious diseases