Online Learning

In the models that we have studied so far, we have assumed that the demand model and its parameters are all known. In practice, demand models need to be estimated before dynamic pricing, assortment optimization, and revenue management can be effectively done. In some instances, there is enough data over a long period of time to calibrate different demand models, do model selection, and update parameter estimates. At the other extreme, we may be pricing for products for which we have little or no information. In this case, demand learning needs to be done on the fly. This is particularly true for online retailing of new products. In this chapter, we address the problem of online demand learning. We study the expected loss in revenue of a learning-and-earning policy relative to an optimal clairvoyant policy that knows the expected demand function. We consider both the case of ample and constrained capacity and measure how the regret grows as the length of the sales horizon increases. We present only the strongest available results for both the case of ample and the case of constrained capacity. In Sect. 10.2, we consider the case with ample capacity, whereas in Sect. 10.3, we consider the case with constrained capacity.