Policy Optimization Using Semiparametric Models for Dynamic Pricing

In this article, we study the contextual dynamic pricing problem where market value of a product is linear in its observed features plus some noise. Products are sold one at time, and only binary response indicating success or failure sale observed. Our model setting similar to work by? except that expand demand curve semiparametric learn dynamically both parametric nonparametric components. We propose statistical learning decision making policy minimizes regret (maximizes revenue) by combining estimation for generalized with unknown link online making. Under mild conditions, noise cdf F(·) mth order derivative ( m≥2), our achieves upper bound O˜d(T2m+14m−1), T time horizon O˜d hiding logarithmic terms feature dimension d. The further reduced O˜d(T) if F super smooth. These bounds close Ω(T), lower belongs class. generalize these results case dependent under strong mixing condition. Supplementary materials article available online.