Chapter 15. Online learning in inventory and pricing optimization

A key for efficient inventory management is the understanding of future customer demand. However, in real world applications that information may not be available. In this chapter, we discuss solving inventory control problems using online learning when little or limited prior demand information is available. The chapter starts by defining the objective function called regret, which is the cost difference between the designed algorithm and that of the clairvoyant solution when complete demand information is available, then reviews some common approaches in online learning. Due to the complexity of system dynamics in different inventory problems, such as inventory carryover, lead times, and censored demand observation, these methods may not be directly applicable and tailored solutions need to be developed. We first focus on models of pure inventory management, including periodic-review inventory systems with censored demand, perishable inventory systems, lost-sales inventory systems with positive lead time, multiple-product systems with substitution or warehouse capacity constraints, systems with fixed ordering cost, and dual sourcing inventory systems. Then we discuss joint inventory control and pricing optimization problems including periodic review inventory problem with backlog, lost-sales problem with censored demand, and the case that the number of price changes is constrained. The chapter concludes with some discussions on future research directions.