Supply chains and production environments, especially in high wage countries like Germany, are constantly becoming more dynamic and complex for several decades already. This development finds its most recent peak in the advent of Industrie 4.0, in which it is intended to enable production systems to be reconfigurable and flexible, manufacturing products in vastly varying amounts and constantly changing specifications. Such changes of course influence the planning model that creates the base for decision-making. For example, the introduction of a new laser cutting machine could require new planning constraints for representing that the machine can operate with different energy levels, influencing its cutting speed and energy consumption.
However, commonly used heuristic-based methods are formulated in a problem-specific manner to exploit domain knowledge about the solution space. The changes in the underlying problem structure and model might, therefore, require changes in the heuristic methods as well, e.g. operators composition and parameters. Additionally, the heuristics perform with significant variance concerning the solution quality and the required execution time depending on the models’ parametrization. Adapting and fine-tuning heuristic-based methods to these changing conditions is a time-consuming process, which are usually performed in a trial-and-error manner. Thus, the objective is to take advantage of recent developments of Automated Algorithm Design to devise an approach (which could be used in the context of Production Planning Systems) that generates heuristic-based methods for different configurations of planning problems.
This talk presents preliminary results of an algorithm-generation approach meant to automate this process, exemplified on the Multi-Level Capacitated Lot-Sizing Problem (MLCLSP). The MLCLSP is an NP-complete problem from the production-planning domain used to determine optimal lot-sizes in a make-to-stock production setting. Several experiments were carried out to evaluate the ability of the proposed method to generate competitive algorithms for benchmark instances, under consideration of different algorithm components and cutoff times. Results indicate that the method is able to generate heuristic algorithms that find high-quality solutions significantly faster than the compared human-designed algorithm.