4 présentations

• 
  14h45 - 15h07

  Online Rack Placement in Large-Scale Data Centers

  • Kayla Cummings, Microsoft
  • Alexandre Jacquillat, prés., Massachusetts Institute of Technology
  • Sean Lo, Massachusetts Institute of Technology
  • Konstantina Mellou, Microsoft Research
  • Ishai Menache, Microsoft Research
  • Marco Molinaro, Microsoft Research

  This paper optimizes data center configurations toward cost-effective, reliable and sustainable cloud supply chains. We formulate an integer optimization model that optimizes the placement of racks of servers to maximize demand coverage given space, power and cooling restrictions. We define a tractable single-sample online approximation (SSOA) approach to multi-stage stochastic optimization, which approximates unknown parameters with a single realization and re-optimizes decisions dynamically. Theoretical results provide strong performance guarantees of SSOA in the canonical online generalized assignment and online bin packing settings. Computational results using real-world data show that our optimization approach can enhance utilization and reduce power stranding in data centers. Our algorithm has been packaged into a software solution deployed in Microsoft's data centers worldwide, contributing an interactive decision-making process at the human-machine interface. Deployment data indicate a significant increase in adoption, leading to improved power utilization, multi-million-dollar annual cost savings, and concomitant savings in greenhouse gas emissions.

• 
  15h07 - 15h29

  Learning Optimal Inventory Policies by Sequential Neural Networks

  • Zhen Yang, prés., University of Texas at Austin
  • Yizhe Huang, University of Texas at Austin
  • Rui Gao, University of Texas at Austin
  • Shuang Li, The Chinese University of Hong Kong (Shenzhen)

  The inventory management problem is crucial for many businesses and has been extensively studied in previous works. Despite these efforts, some models remain challenging due to the unknown structure of the optimal inventory policy. We aim to approximate the policy at each stage by a neural network which is a universal approximator, and establish the statistical and optimality guarantees.

• 
  15h29 - 15h51

  Mobile parcel locker scheduling with customer choices under uncertain demand

  • Can Yin, University of Minnesota
  • Yiling Zhang, University of Minnesota
  • Can Yin,
  • Yiling Zhang, prés., University of Minnesota

  The movable unit equipped with a set of parcel lockers has been recently developed as a new mode to improve the efficiency of last-mile delivery. Mobile parcel lockers, different from traditional facility location problems placing permanent facilities in a specific region, can be relocated at any time, driven by the changing demand rate from location to location over time. This talk will present a scheduling problem of mobile parcel lockers under uncertain demand and customer choices. The problem is formulated as a distributionally robust two-stage stochastic programming model, where customer choices are captured by multinomial logit (MNL) models. The probability distribution of uncertain demand, although unknown for this new service, is characterized by the first moments and the support. Our models flexibly adapt different prior beliefs of spatial structures and pricing strategies. We obtain exact mixed-integer linear programming reformulations and derive valid inequalities used in decomposition algorithms.

• 
  15h51 - 16h13

  Online learning and pricing for service systems with reusable resources

  • Huiwen Jia, prés., UC Berkeley

  We consider a price-based revenue management problem with finite reusable resources over a finite time horizon T. Customers arrive following a price-dependent process and each customer requests one unit of the reusable resources. If there is an available unit, the customer gets served within a price-dependent random service time; otherwise, the customer waits in a queue until the next available unit. We assume that the firm does not know how the arrival and service processes depend on posted prices, and thus it makes adaptive pricing decisions in each period based only on past observations to maximize the cumulative revenue. Given a discrete candidate price set with cardinality P, we propose a batched online learning algorithm. The major challenge lies in the complicated system dynamics induced by customer arrival and departure processes, and we prove that the cumulative regret upper bound matches the regret lower bound up to a logarithmic factor.