A methodology for predicting single-stage lot-sizing performance: Analysis and experiments

This article seeks to structure the lot-sizing selection process by isolating conditions under which certain lot-sizing heuristics dominate others. To accomplish this requires as a first step some analytical comparisons of the heuristics, followed by extensive experimental tests of the hypotheses developed through comparative analyses. Building upon our studies of lot-sizing performance, this research evaluates four new heuristics recently proposed by Boe and Yilmaz, Freeland and Colley, Gaither, and Groff and compares them to existing methods such as Silver-Meal and the Wagner-Whitin algorithm. The decision sequence of each heuristic is carefully examined, yileding some hypotheses about when a particular method should be most effective. The analytical study indicates that the Freeland-Colley heuristic (and its equivalent, Boe-Yilmaz) should excel in extremely lumpy demand situations, as is often found in lower levels of the product structure in an MRP system. The Groff heuristic, although very similar to Silver-Meal, tends to produce larger orders than Silver-Meal in lumpy demand situations. Also, it is demonstrated that, using the same set-up and holding cost parameters, the Gaither and Freeland-Colley methods always specify larger orders than Groff or Silver-Meal for any given sequence of demands. These analytical findings were supported in an extensive set of simulation experiments, designed to compare the cost performance of these algorithms under a wide range of cost and demand conditions. As expected, no single heuristic dominated under all conditions. Groff and Silver-Meal, however, tend to outperform the others; Groff and Silver-Meal are similar to each other in cost performance with Groff being slightly better in the experiments. Freeland-Colley, Boe-Yilmaz, and Gaither become more cost competitive as demand variability and demand lumpiness increase.