My research is about high-mix, low-volume (HMLV) manufacturers, which produce a
wide variety of customized products in small batches. In this manufacturing
environments variable demand, heterogeneous routings and uncertain processing
times cause the product mix on the shop floor to change continuously, so that
bottlenecks shift over time. These unpredictable shifts cause frequent tardiness and
unreliable delivery dates. In practice, the two main shop-floor decisions, order
release and dispatching, are controlled by heuristic rules whose parameters are set
in advance and do not respond to the current shop-floor state. Moreover, no single
heuristic rule performs best under all conditions and both decisions are mostly
studied in isolation, although the workload at stations results from their interaction.
My research focuses on using learning-based approaches, which can adapt these
decisions dynamically, but they face open challenges, such as expensive training
data generation, robustness of predictions under shifting product mixes, delayed
effects of decisions and interpretability of deep-learning policies. In addition, most
approaches are validated only on synthetic instances. This PhD therefore develops
interpretable, learning-based methods that jointly adapt order release and
dispatching, integrate planners’ experience, and are validated on real production
data from the industry partner Thales NL.