A General Framework for Resource Constrained Revenue Management with Demand Learning and Large Action Space

This paper proposes a general framework/meta-policy to solve Revenue Management (RM) problems with demand learning and potentially large action space, constrained by initial unreplenishable resources. This framework combines the technique of primal-dual method in optimization and Upper-Confidence-Bound (UCB) algorithm in learning. Three important RM applications are discussed in this paper to illustrate the general framework: network revenue management (NRM), dynamic assortment selection with multinomial- logit (MNL) choice model, and joint pricing and assortment optimization problems. This paper demonstrates how to adapt the framework to each application by designing application-specific subroutines. Theoretical results show that the adapted algorithms in all three applications are effective in achieving low regrets, and numerical experiments using the example of dynamic assortment selection demonstrate that our algorithm has great empirical performance and it is computationally efficient.