A1958
Title: Non-stationary stochastic optimization: Adaptive algorithm and applications in operations
Authors: Jinzhi Bu - The Hong Kong Polytechnic University (Hong Kong) [presenting]
Siyi Wang - The Hong Kong Polytechnic University (Hong Kong)
Sen Yang - New York University (United States)
Abstract: Online Stochastic convex optimization is studied where the distribution of the random environment evolves over time under a variation-budget constraint. The state-of-the-art restarting procedure attains a minimax-optimal dynamic regret rate but requires prior knowledge of the variation budget. This crucial assumption is relaxed by developing an Adaptive Stochastic Gradient Descent (AdaSGD) algorithm that adapts to unknown non-stationarity without using information about the variation budget. AdaSGD achieves the same dynamic regret bound as in the case where the variation budget is known. The framework is further applied to three operations management problems and AdaSGD is extended to (i) the setting with strongly convex cost functions and (ii) the setting with time-dependent cost functions.
