Pricing in Network Revenue Management Systems with Reusable Resources

Abstract

This thesis focuses on the problem of pricing reusable products in the network revenue management setting. In a nutshell, dynamic pricing problem concerns pricing and selling a finite inventory of products within a given time horizon so as to maximize the total revenue. Most of the existing literature studies the setting with perishable products in which the products sold are permanently removed from inventory. In this thesis, we tackle a different and arguably more challenging problem with reusable products wherein the products are returned back to the seller upon serving a customer and can be used to serve another customer. This class of problems finds a broad range of applications including hotel management, cloud computing, workforce management, call center service, and car rental management.

In the first chapter of the thesis, we address the pricing of reusable resources with advance reservation when the demand function is known as a function of price and the demand follows a Poisson point process. We demonstrate that a simple static pricing policy is asymptotically optimal when demand and capacity are scaled without bound. The performance of the policy is measured as a ratio with respect to the policy that does not exhibit any blocking. We also show that the static policy becomes optimal at a rate close to 1 over the square root of n, where n is a scaling factor. Simulation results show the asymptotic behavior but additionally, it shows that for small-scaled systems, the static pricing policy performs very well relative to the no-blocking policy.

In the second chapter of the thesis, we consider the learning variant of the same problem in which the customer’s response to selling price and the demand distribution are not known a priori. Connecting this problem to multi-armed bandits (MAB), we propose a variant of the upper confidence bounds (UCB) algorithm. The setting is different from literature in that capacity constraints exist and booking profile is dynamically updated. We solve an LP in every period where the UCB estimates guides the right-hand side parameter and outputs a distribution over the finite pricing actions. We demonstrate that for some large scaling factor n, with high probability, the seller will always choose the optimum after the testing phase and will not exhibit any blocking.

In the third and final chapter of the thesis, we employ unsupervised learning methods to tackle the pricing policy from a practical point-of-view. In particular, model-free reinforcement learning method is used to implicitly learn the transition dynamics that governs the reward process to maximize revenue. Deep neural networks are used to parametrize the action policy and value function and through a simulated environment. We show that the generated pricing policy, using purely data, achieved good performance with respect to traffic, revenue, and blocking.