Unsere Studie "Flow-Shop Scheduling with Demand Plans and Multiple Lines Using Reinforcement Learning" wurde bei der Conference of Technology and Information Systems zur Präsentation angenommen.

Unsere Studie "Flow-Shop Scheduling with Demand Plans and Multiple Lines Using Reinforcement Learning" wurde bei der Conference of Technology and Information Systems (http://cist2019.conferency.com/) zur Präsentation angenommen. In unsere Studie adressieren wir das Flow Shop Problem mit mehreren Produktionslinien mithilfe von Reinforcement Learning. Wir zeigen, dass unser Ansatz nicht nur bessere Lösungen findet als aktuelle Baselineverfahren (z.B. Gurobi), sondern auch dynamisch auf spontane Anpassungen in der Produktion reagieren kann.

Autoren: Janis Brammer, Bernhard Lutz, Dirk Neumann.

Abstract:
We study the permutation flow-shop problem in which a given number of jobs must be assigned to a production sequence in order to minimize the total makespan. In contrast to previous studies, we consider flow shops with demand plans and multiple lines. It can be shown that finding the optimal solution for large problem instances with more than two machines is NP-hard. Existing heuristic approaches solve the multi-line problem by simplifying the problem structure. In this paper, we present a reinforcement learning-based approach with the goal of learning a scheduling policy that considers the full problem characteristic. Extensive computational results with more than 1,000 problem instances demonstrate that our approach outperforms other solution methods in non-standard problem congurations. Our approach is also able to generate sequences for preassigned production lines and can react to short-term disturbances without relearning the policy.