Efficient Real-time Policies for Revenue Management Problems

Abstract

This dissertation studies the development of provably near-optimal real-time prescriptive analytics solutions that are easily implementable in a dynamic business environment. We consider several stochastic control problems that are motivated by different applications of the practice of pricing and revenue management. Due to high dimensionality and the need for real-time decision making, it is computationally prohibitive to characterize the optimal controls for these problems. Therefore, we develop heuristic controls with simple decision rules that can be deployed in real-time at large scale, and then show theirs good theoretical and empirical performances. In particular, the first chapter studies the joint dynamic pricing and order fulfillment problem in the context of online retail, where a retailer sells multiple products to customers from different locations and fulfills orders through multiple fulfillment centers. The objective is to maximize the total expected profits, defined as the revenue minus the shipping cost. We propose heuristics where the real-time computations of pricing and fulfillment decisions are partially decoupled, and show their good performances compared to reasonable benchmarks. The second chapter studies a dynamic pricing problem where a firm faces price-sensitive customers arriving stochastically over time. Each customer consumes one unit of resource for a deterministic amount of time, after which the resource can be immediately used to serve new customers. We develop two heuristic controls and show that both are asymptotically optimal in the regime with large demand and supply. We further generalize both of the heuristic controls to the settings with multiple service types requiring different service times and with advance reservation. Lastly, the third chapter considers a general class of single-product dynamic pricing problems with inventory constraints, where the price-dependent demand function is unknown to the firm. We develop nonparametric dynamic pricing algorithms that do not assume any functional form of the demand model and show that, for one of the algorithm, its revenue loss compared to a clairvoyant matches the theoretic lower bound in asymptotic regime. In particular, the proposed algorithms generalize the classic bisection search method to a constrained setting with noisy observations.