Date: 04 May 2022
Time: 12:45 – 13:15. Hours
Room: RA1501 & online
Speaker: Anne Zander (UT-MOR)
Title: "A healthcare example to showcase solution methods for sequential decision-making”
In his new book on reinforcement learning and stochastic optimization, Warren B. Powell presents his work on a unified framework for sequential decision-making problems, including a unified way of modeling and four meta-classes of policies, i.e., ways to decide how to act in a given system state. Taking this unified perspective helps to avoid choosing one’s favorite modeling and solution method and instead find the most appropriate approach to a given problem. In this talk, I will present an example problem to showcase Powell’s unified framework. More precisely, I will talk about physician panel management, where a panel is defined as the set of patients who regularly visit the same physician, e.g., a family doctor. The physician then needs to decide which requesting new patients to take into the panel to achieve a long-term balance between her capacity and the panel’s demand. Although we will cover all four policies, most results will be based on a direct lookahead approximation (my favorite method), which in this case is given by a deterministic integer linear program (ILP) together with a method to combine solutions of this ILP to decide on accepting or rejecting new patient requests in a stochastic environment in real-time. Using a stochastic discrete-event simulation with parameters based on real-world data shows that taking decisions while looking ahead several periods instead of only focusing on the current situation significantly lowers the mismatch between panel demand and physician capacity over time.