CORMSIS Seminar Event

Abstract:

It has been proved that choice-based Revenue Management can result in significant increases in revenue. Estimating customer-choice models is difficult, especially in situations where it is not possible to observe customers who enter the system but do not purchase a product. As these are live sales systems, it is important that the system continues to earn revenues while the parameters are being estimated. We use a Multi-Armed Bandit experiment to find the optimal trade-off between learning and earning in choice-based Revenue Management. We use the Multinomial Logit model to describe customer preferences, and a variation of the Expectation-Maximization (EM) algorithm is used to jointly estimate the parameters of this model and the customer arrival rate. This constitutes the learning phase. Approximate dynamic programming is used to find the set of optimal prices at each time step as part of the earning phase. The talk will describe the complete methodology that we have implemented and provide some numerical results.