Online Resource Allocation under Partially Learnable Demand

E-commerce platforms host markets for perishable resources in various industry sectors ranging from airline to Internet advertisement to online retail. In these markets, the demand realizes sequentially and the firm needs to make online (irrevocable) decisions regarding how (and at what price) to allocate resources to arriving demand without precise knowledge of the future demand. The success of any online allocation algorithm crucially depends on the firm’s ability to estimate the future demand. If the demand can be predicted or even learned, then under some mild conditions on the amount of available resources, there is little loss incurred for making online decisions (as shown in [1] among others). However, in many markets, the demand cannot be perfectly learned due to unpredictable components such as traffic spikes and competitor’s change of strategy. In such cases the firm can take a completely robust approach and assume that the demand is not predictable, but that usually results in strategies that are too conservative (as studied in [2] and others). Instead, the firm may wish to employ online polices that try to learn the demand, but at the same time do not overfit to the observed data with the caution that it could embody an unpredictable component. This paper aims to investigate to what extent the above goal is achievable: we propose a new demand model that is only partially learnable and we show how the firm can still make use of the limited information that the data reveals and improve upon a completely conservative approach. We study a basic online resource allocation problem, known as the single-resource revenue management with 2 fare classes, where stochastic information about demand is unknown a priori, and it can only be partially learned. In our demand model, an adversary determines a sequence of customers to be revealed to the online algorithms. However, a random subset of customers does not follow this prescribed order, and instead, arrives at uniformly random times. We call this group of customers the “predictable group”. Even though we cannot identify which customers belong to the predictable group, we can still (partially) learn the future demand because this group is almost uniformly spread in the time horizon. We use this to design online algorithms (adaptive and non-adaptive) with competitive ratios significantly higher than that of algorithms designed for the adversarial customer arrival model. Customers that do not belong to the predictable group