Online Network Revenue Management Using Thompson Sampling

by Kris Johnson Ferreira, David Simchi-Levi & He Wang

Overview — Kris Johnson Ferreira and colleagues develop a machine learning algorithm that changes product prices in order to learn consumer demand and maximize total revenue in the presence of limited inventory.

Author Abstract

We consider a network revenue management problem where an online retailer aims to maximize revenue from multiple products with limited inventory. As common in practice, the retailer does not know the expected demand at each price and must learn the demand information from sales data. We propose an efficient and effective dynamic pricing algorithm, which builds upon the Thompson sampling algorithm used for multi-armed bandit problems by incorporating inventory constraints into the pricing decisions. Our algorithm proves to have both strong theoretical performance guarantees as well as promising numerical performance results when compared to other algorithms developed for the same setting. More broadly, our paper contributes to the literature on the multi-armed bandit problem with resource constraints, since our algorithm applies directly to this setting when the inventory constraints are interpreted as general resource constraints.

Paper Information