Production scheduling is a key factor for manufacturing companies. It determines how efficiently staff, machines and materials are used, how reliably delivery dates can be met, and how quickly companies can respond to disruptions. In practice, scheduling is still often done manually or based on rigid rule-based systems. With increasing product variety, fluctuating demand and more dynamic supply chains, these approaches are reaching their limits.

AI-based methods offer considerable potential for improvement. However, in practice they are often limited by a lack of training data, insufficient explainability and the challenge of integrating them into existing IT systems. This is the motivation for the KI-HEAP4.0 research project.

The project focuses on developing a reusable toolkit of models and methods for order optimization in manufacturing. It consists of three closely connected components:

1. Digital twin as a simulation and optimization environment
A realistic representation of the production process, allowing workflows, deviations and failure scenarios to be simulated at an early stage, before sufficient real-world data has been collected.

2. Multidimensional mathematical optimization
A transparent and explainable approach that breaks down the overall problem into solvable subproblems and provides traceable decision recommendations.

3. Reinforcement-learning-based optimization
Software agents are trained exploratively in the digital twin, learning the stochastic properties and dynamics of the production system.

Based on these components, KI-HEAP4.0 develops a hybrid approach that brings together the strengths of mathematical optimization and learning-based methods. Mathematical optimization provides transparency, explainability and reliable solution properties, while learning-based methods enable adaptability and robust decision-making under uncertainty.

F. X. MEILLER Fahrzeug- und Maschinenfabrik in Munich is the project’s application partner. Its highly variant hydraulic production and historically grown factory structures are representative of many medium-sized manufacturing companies in Germany. This provides a realistic setting for testing and evaluating the methods developed in the project.