« A Parallel-Machine Learning framework to tune metaheuristics for advanced manufacturing scheduling problems« , Hanser Jiménez (Postdoc, European project MODAPTO, CRAN/UL)
Abstract:
Meta-heuristics (MH) have become a de facto approach to find approximate solutions for complex scheduling problems. However, since the quality of solutions provided by these methods is highly sensitive to the value of their parameters, tuning parameters is a key and challenging step to guarantee a good performance. Tunning MHs is not trivial since it is in turn dependent on the complexity of the problem at hand and the available time to perform such procedure. In the context of real-world manufacturing processes, the specific characteristics of such processes give place to complex scheduling cases, which turn MHs into expensive-to-evaluate functions for candidate settings needing to be tested. Such characteristics include the high interaction of semi-finished goods and operations in their respective bills of material, as well as the specific constraints of manufacturing operations that need to be balanced.
In this talk, we propose a Bayesian-Optimization (BO)-based framework supported by parallel computing techniques to perform MH’s tuning for manufacturing processes. The proposed framework captures the specificity of manufacturing processes in a training phase by learning the impact of MH’s parameters on the business key performance indicators. By doing so, the framework can be used to find a near-optimal parameter setting able to produce efficient schedules for new cases once it is trained. The proposed framework configures as a managerial tool that can be integrated with existing Advanced-Planning-and-Scheduling software that use MHs as their underlying models.