33. A data-driven make-to-order system under operational uncertainty
Invited abstract in session MC-49: Analytics and the link with stochastic dynamics 1, stream Analytics.
Monday, 12:30-14:00Room: Parkinson B10
Authors (first author is the speaker)
| 1. | Shu-Jung Sunny Yang
|
| School for Business and Society, University of York | |
| 2. | Yingying Huang
|
| School for Business and Society, University of York |
Abstract
Manufacturing sectors with make-to-order systems often need long production planning horizons for product delivery. Despite the predetermined demand planning, there remains an inherent challenge in striking a balance when ordering raw materials due to the operational costs tied to procurement, alongside the risks of over-ordering and under-ordering, especially when these costs are not apparent during the decision-making phase for production and inventory. Traditionally, the “Predict-then-Optimise” (PO) approach has attempted to mitigate these issues by separating the tasks of prediction and optimisation. However, this methodology tends to falter, especially with scarce or volatile historical data. This study employs the emerging “Smart Predict-then-Optimise” (SPO) framework, which harnesses limited data to derive optimal decisions to develop a data-driven optimisation model that integrates machine learning and mixed-integer linear programming. A proximal SPO (PSPO) loss function is proposed to improve convergence and performance. We perform numerical experiments on both real-world and synthetic data to empirically verify the success of our data-driven optimisation model in comparison to the standard PO and SPO approaches. We show that the SPO-based solutions tend to diversify the production schedule to mitigate the inherent risks of cost fluctuations, while the PO-based solutions tend to produce more in the periods with predicted low operational costs.
Keywords
- Production and Inventory Systems
- Analytics and Data Science
- Machine Learning
Status: accepted
Back to the list of papers