Anmerkungen:
Nach Informationen von SSRN wurde die ursprüngliche Fassung des Dokuments January 8, 2023 erstellt
Beschreibung:
We consider the problem of online pricing with offline data, where the decision maker has in hand pre-existed offline data and then takes online sequential actions to maximize expected cumulative revenue as well as learning the optimal pricing. We focus on a simple and specific problem setting where the distribution of the offline data is "polluted", meaning that their distribution can be different from the online data to be obtained. In this work, we provide explicit answers to two questions. First, if the decision maker does not know that the offline data are polluted, and applies a policy that should be optimal if the offline data are not polluted, how serious is the effect of pollution on the online learning performance? We find the critical pollution level, above which the ``optimal policy'' with polluted data behaves worse than another policy that does not utilize offline data at all. Second, if the decision maker knows the magnitude of how the offline data are polluted, how to utilize such polluted offline data to better enhance online learning performances? We find a better policy by assigning a smaller weight to each offline data point