Data Driven Optimization: Theory and Applications in Supply Chain Systems

[摘要] Supply chain optimization plays a critical role in many business enterprises. In a data driven environment, rather than pre-specifying the underlying demand distribution and then optimizing the system’s objective, it is much more robust to have a nonparametric approach directly leveraging the past observed data. In the supply chain context, we propose and design online learning algorithms that make adaptive decisions based on historical sales (a.k.a. censored demand). We measure the performance of an online learning algorithm by cumulative regret or simply regret, which is defined as the cost difference between the proposed algorithm and the clairvoyant optimal one.In the supply chain context, to design efficient learning algorithms, we typically face two majorchallenges. First, we need to identify a suitable recurrent state that decouples system dynamics into cycles with good properties: (1) smoothness and rich feedback information necessary to apply the zeroth order optimization method effectively; (2) convexity and gradient information essential for the first order methods. Second, we require the learning algorithms to be adaptive to the physical constraints, e.g., positive inventory carry-over, warehouse capacity constraint, ordering/production capacity constraint, and these constraints limit the policy search space in a dynamic fashion. To design efficient and provably-good data driven supply chain algorithms, we zoom into the detailed structure of each system, and carefully trade off between exploration and exploitation.

[发布机构] University of Michigan

[效力级别] Online learning

[关键词] Supply Chain;Online learning;Reinforcement learning;Industrial and Operations Engineering;Engineering;Industrial & Operations Engineering