Adaptive stochastic lookahead policies for dynamic multi-period purchasing and inventory routing / Daniel Cuellar-Usaquén, Martin W. Ulmer, Camilo Gomez, David Álvarez-Martínez
VerfasserCuellar-Usaquén, Daniel ; Ulmer, Martin W. ; Gomez, Camillo ; Álvarez-Martínez, David
ErschienenMagdeburg, Germany : Otto-von-Guericke-Universität Magdeburg, Fakultät für Wirtschaftswissenschaft, 2023
Umfang1 Online Ressource (50 Seiten, 1,52 MB) : Diagramme
SerieWorking paper series ; 2023, no. 4
 An einem Rechner der Bibliothek verfügbar
We present a problem motivated by discussions with Colombian e-commerce platforms for agri-food products. In regular time intervals (periods) the platforms collect groceries from local farmers and stores them at a warehouse to distribute them to local customers. The supply quantities and prices per farmer and the cumulated customer demand can change from period to period. Thus there is value in purchasing more than needed in one period to exploit cheap prices and consolidation opportunities to hedge against future uncertainty and to save routing cost in future periods. A careful balance between too much and not enough inventory needs to be found especially since inventory perishes over time. The resulting optimization problem is a stochastic dynamic multi-period routing problem with inventory and purchasing decisions. The decision space of the problem is vast as it combines purchasing inventory and routing decisions. Further the value of a decisions is unknown since it depends on future developments and decisions. We propose solving the problem with a stochastic lookahead method. In every state the method samples a set of future realizations and solves the resulting two-stage stochastic program. To cope with the complex decision space in first and second stage we propose a “soft” decomposition where the inventory and purchasing decision are fully considered but the routing decisions are simplified and their cost is approximated via a cost function approximation. As the routing cost also depends on future decisions the approximated cost are learned iteratively via repeated simulation and adaption of the lookahead. We show that our method outperforms a large number of benchmark policies for a variety of instances. We further analyze the functionality of our method and investigate variation in the problem dimensions in a comprehensive analysis.