A machine learning approach for predicting the best heuristic for a large scaled Capacitated Lotsizing Problem

Abstract For some NP-hard lotsizing problems, many different heuristics exist, but they have different solution qualities and computation times depending on the characteristics of the instance. The computation times of the individual heuristics increase significantly with the problem size, so that testing all available heuristics for large instances requires extensive time. Therefore, it is necessary to develop a method that allows a prediction of the best heuristic for the respective instance without testing all available heuristics. The Capacitated Lotsizing Problem (CLSP) is chosen as the problem to be solved, since it is a fundamental model in the field of lotsizing, well researched and several different heuristics exist for it. The CLSP addresses the problem of determining lotsizes on a production line given limited capacity, product-dependent setup costs, and deterministic, dynamic demand for multiple products. The objective is to minimize setup and inventory holding costs. Two different forecasting methods are presented. One of them is a two-layer neural network called CLSP-Net. It is trained on small CLSP instances, which can be solved very fast with the considered heuristics. Due to the use of a fixed number of wisely chosen features that are relative, relevant, and computationally efficient, and which leverage problem-specific knowledge, CLSP-Net is also capable of predicting the most suitable heuristic for large instances.