Milos Hauskrecht

Title: Modeling and optimizing patient-management processes with POMDPs.

Abstract:
Partially observable Markov decision processes (POMDPs) offer an elegant
mathematical framework for modeling and solving decision making problems
with action-outcome uncertainty and  imperfect state information. The
focus of this talk is on POMDPs suitable for modeling and solving
patient-management processes.  First, I present a POMDP model we built for 
the management of patients with the ischemic heart disease. The model
uses a hierarchical Bayesian belief network to represent the disease
dynamics, and a factored cost model to represent payoffs associated with
treatment choices and intermediate patient outcomes. Second, I briefly
describe approximation  methods that take advantage of the structure of
the model and let us solve the model efficiently.  Finally, I discuss a
number of practical issues we encountered while building, solving and
evaluating the model.