Deterministic Benchmarks to Inform Sequential Decisions

Information Systems and Operations Management 

Intervenant: Itai Gurvich (Kellogg)

Salle Bernard Ramanantsoa

Abstract:

Dynamic programming is a canonical tool for solving complex sequential decision problems in operations. Yet, because it suffers from the curse of dimensionality, one often must rely on approximations. Among these, deterministic—or “fluid”—approximations have long served as tractable benchmarks that reveal key structural properties of optimal or near-optimal policies in dynamic resource allocation problems across service operations and revenue management.

While such fluid approximations have been widely—and often ad hoc—applied, this talk presents a systematic framework for leveraging them to design high-quality control policies through what we call fluid lookahead. I will illustrate the approach in a family of finite-horizon revenue management problems, showing how fluid lookahead captures key structural properties of the true optimal policies and, perhaps surprisingly, achieves near-optimal performance with only a few lookahead steps. I will also discuss how the same principles extend naturally to infinite-horizon settings.