Data-Driven Decision Processes - Job Talk Practice

Title: Sequential Fair Allocation: Achieving the Optimal Envy-Efficiency Tradeoff Curve

Abstract: Optimizing the operations of complex systems often involves making tradeoffs between objectives such as efficiency, revenue, and fairness.  How these criteria interact is often not well understood, and current approaches focus on maximizing a convex combination which provides little operational insights. In this talk we investigate the tradeoff between fairness and efficiency in online resource allocation motivated by a partnership with the Food Bank of the Southern Tier.  We start by establishing an uncertainty principle: a lower bound exactly characterizing the envy and efficiency Pareto frontier.  We complement this by showing how to leverage the principle of algorithmic guardrails, artificial constraints imposed on the set of actions, which allows algorithms to exactly match the uncertainty principle.  These techniques extend to a variety of settings including perishable resources, evolving budgets, and a wide range of user preference models.

This work falls under a broader range of questions in designing practical sequential decision making algorithms for uncertain environments.  Such questions include optimizing multiple objectives, as discussed above, but also in designing algorithms which computationally and statistically scale to real-world systems.  Time permitting, I will highlight this by briefly discussing my work designing algorithms which leverage information relaxation in problems with exogenous dynamics.  The algorithm appeals to existing computational solutions for business problems for solving large-scale deterministic optimization problems. This algorithm design is currently in deployment on the Microsoft Azure platform.