Reinforcement learning for sequential decision making in population research

Abstract

AbstractReinforcement learning (RL) algorithms have been long recognized as powerful tools for optimal sequential decision making. The framework is concerned with a decision maker, the agent, that learns how to behave in an unknown environment by making decisions and seeing their associated outcome. The goal of the RL agent is to infer, through repeated experience, an optimal decision-making policy, i.e., a sequence of action rules that would lead to the highest, typically long-term, expected utility. Today, a wide range of domains, from economics to education and healthcare, have embraced the use of RL to address specific problems. To illustrate, we used an RL-based algorithm to design a text-messaging system that delivers personalized real-time behavioural recommendations to promote physical activity and manage depression. Motivated by the recent call of the UNECE for government-wide actions to adapt to population ageing, in this work, we argue that the RL framework may provide a set of compelling strategies for supporting population research and informing population policies. After introducing the RL framework, we discuss its potential in three population-study applications: international migration, public health, and fertility.