Meta Dynamic Pricing: Transfer Learning Across Experiments

We study the problem of learning shared structure across a sequence dynamic pricing experiments for related products. consider practical formulation in which unknown demand parameters each product come from an distribution (prior) that is then propose meta algorithm learns this prior online while solving Thompson sampling (each with horizon T) N different Our addresses two challenges: (i) balancing need to learn (meta-exploration) leverage estimated achieve good performance (meta-exploitation) and (ii) accounting uncertainty by appropriately “widening” as function its estimation error. introduce novel alignment technique analyze regret misspecified prior, may be independent interest. Unlike prior-independent approaches, our algorithm’s grows sublinearly N, demonstrating price can negligible experiment-rich environments (large N). Numerical on synthetic real auto loan data demonstrate significantly speeds up compared algorithms. This paper was accepted George J. Shanthikumar, Management Science Special Section Data-Driven Prescriptive Analytics.