Measuring Policy Performance in Online Pricing with Offline Data

Abstract

For the problems of online pricing with offline data, or other similar general online learning with offline data problems, there are two classical approaches to measure the performance of an online policy: the worst-case approach and the Bayesian approach. We argue that the worst-case approach may not be able to appropriately handle offline data, particularly in evaluating the impact of offline data on online policy performance. We construct counterexamples to support our argument on the worst-case approach. Alternatively, we show that the Bayesian approach may be preferred because of its capacity in appropriately measuring the impact of offline data, as well as other properties regarding finding and analyzing optimal policies.