Developing New Inventory Segmentation Methods for Large-Scale Multi-Echelon Inventory Systems

Alireza Sheikh Zadeh, University of Arkansas, Fayetteville


A unified objective of this dissertation is to increase the efficiency of large-scale inventory policy algorithms (i.e. including thousands of spare parts) by developing segmentation and aggregation methods while controlling the performance penalty associated with the aggregation. This dissertation is designed based on three articles. In the first and second articles, we present a near-optimal heuristic segmentation approach in the context of the VARI-METRIC and multi-echelon (r,q) problems, which are performance-based nonlinear stocking policy models. The performance-based classification method is developed based on the stocking policies, in contrast to conventional methods that rely on segmentation based on item operational parameters such as demand, annual dollar usage, or item unit cost. This research proposes the artificial stocking policy (ASP) as new inventory classification criteria. Moreover, a non-subjective linear programming scoring method is developed for ranking the criticality of the items. We also proposed performance-based partitioning techniques such as Pareto threshold cutoff-points (PTC), equal partitioning (EP), and simulated annealing (SA), which are tested successfully versus the alternative classification, clustering, and complete enumeration techniques. For instance, the performance difference between using the ASP and SA versus the optimal lower bound is about 1.2%.

Finally, in the third article, we design an aggregation and disaggregation framework for solving large-scale, multi-echelon inventory policy problems in a more efficient way. The aggregation process assigns a similar group policy (aggregated solution) for all items belonging to the same group, and the disaggregation process calibrates the group policy into the individual policies using marginal analysis greedy adjustment (MAGA). This adjustment is exceptionally fast in practice and indicates improvement in the performance versus an aggregated solution. The numerical results show that MAGA can approach the original optimal solution in a range between 1-3% penalty cost while reducing the execution time in a range between 90-99.5%.