A machine learning approach to improving dynamic decision making

Decision strategies in dynamic environments do not always succeed in producing desired outcomes, particularly in complex, ill-structured domains. Information systems often capture large amounts of data about such environments. We propose a domain-independent, iterative approach that (a) applies data mining classification techniques to the collected data in order to discover the conditions under which dynamic decision-making strategies produce undesired or suboptimal outcomes and (b) uses this information to improve the decision strategy under these conditions. In this paper, we formally develop this approach and illustrate it by providing detailed examples of its application to a chronic disease care problem in a healthcare management organization, specifically the treatment of patients with type 2 diabetes mellitus. In particular, the proposed iterative approach is used to improve treatment strategies by predicting and eliminating treatment failures, i.e., insufficient or excessive treatment actions, based on information that is available in electronic medical record systems. We also apply the proposed approach to a manufacturing task, resulting in substantial decision strategy improvements, which further demonstrates the generality and flexibility of the proposed approach.