Online and Offline Learning Algorithms in Operations Management

Abstract

The primary focus of this dissertation is to develop both online and offline algorithmic frameworks for data-driven decision-making in the context of operations management problems including supply chain management and revenue management. In a data-poor environment, online learning algorithms can be developed to utilize streaming data to help decision-making sequentially balancing exploration and exploitation. On the other hand, when there is already massive logged data available, offline learning algorithms can be developed for stochastic optimization and policy making.

Many real-world operations management problems have complex system dynamics, abundant operational constraints, as well as varying qualities of accessible data (e.g., missing or censored data). These features, unique to operations management problems, become major challenges of utilizing data in the process of optimization and better decision-making, despite the existence of numerous learning frameworks developed by researchers from other disciplines such as computer science. To overcome these challenges, we carry out research on representative problems arising in the context of operations management. First, we formulate mathematical models capturing the main trade-offs based on specific domain knowledge. Second, we derive structural properties of the optimal policies that provide a foundation for the design of algorithms. Third, we propose a learning algorithm to solve the incomplete information problem, along with a theoretical analysis of its performance guarantee. Finally, we carry out extensive numerical or empirical studies for validation of the method and the discovery of managerial insights.

This dissertation first investigates the online algorithm for a multi-variate optimization problem within a multi-product system with general upgrading. Then still focusing on online learning algorithms, the dissertation proposes two distinct two-layer methodological frameworks designed to solve joint optimization problems that encompass two distinct decision variables. One is based on the setting where the underlying dynamics form a Markov chain in a dual-sourcing system. A learning algorithm combining Successive Elimination and Sample Average Approximation is proposed and demonstrates an optimal convergence rate of regret. The other one is designed for scenarios in two-sided markets where the decision maker makes no parametric assumptions on underlying functions. By integrating Bisection Search with the Upper Confidence Bound (UCB) algorithm in bandit control, the proposed framework guides the sequential decision-making incurring regret with a provably tight upper bound, which is optimal for any learning algorithm. Finally, the dissertation studies the offline learning algorithm for the problem of feature-based pricing with an offline dataset containing information on historical decisions, covariates, and censored outcomes. An offline algorithm incorporating supervised learning techniques and survival analysis in the language of causal inference is proposed, whose profit is proven to converge to the optimum as the sample size goes to infinity.