International Conference on Operations Research - OR 2022
Validation of a meta-model approach for setting order release parameters in engineer-to-order systems
Ralf Gössinger, Jana Philips
In engineer-to-order systems, the additional uncertainties of specification time and capacity requirements are relevant for releasing orders. A MILP model for order release has been proposed, in which these stochastic variables are estimated using chance-constraints. This requires probability thresholds that need to be specified in advance as parameters. In order to obtain clues for favorable parameter settings, we propose a new data-driven multi-criteria approach that takes cost and robustness measures into account. The approach consists of four steps: (1) Data generation: A database of problem instances and related solutions is generated and aggregate performance values are calculated. (2) Meta-model estimation: The model considers the impact of both, planning situation and parameter values, on the aggregate performance. (3) Parameter recommendation: Favorable parameter values are derived by algebraically analyzing the meta-model. (4) Validation of recommendation quality: Model fit and predictive ability are verified by statistical means. The suitability of this approach is demonstrated by a comprehensive numerical study. Systematically generated problem instances are solved with regard to the cost objective. For each combination of planning situation and parameter values, the minimum cost levels and observed robustness values of the solutions are aggregated to normalized performance values. A second-order polynomial that accounts for two-factor interactions is used to estimate the meta-model. The ability of the approach to predict favorable parameter values is assessed by a Monte Carlo cross-validation.