With AI-supported multi-agent reinforcement learning, inventories in industrial supply chains are to be managed more flexibly, robustly and economically.
A challenge for traditional processes
"In complex distribution networks, such as those we experience every day in the company, we already use powerful algorithms that reliably deliver optimal solutions. Reinforcement learning (RL) offers exciting opportunities: it could complement our tried-and-tested processes, further increase efficiency and address particularly challenging problems where traditional algorithms reach their limits. In the transfer project, we want to find out to what extent RL can actually realise this potential," says Katharina Klerx from OPTANO, describing the current situation. It is precisely when uncertainties and networking in the system increase that traditional methods reach their limits. There is often a lack of efficient mathematical solutions for particularly complex warehouse and delivery structures.
Multi-agent reinforcement learning as a solution
The ML4SAFETYSTOCK project, which is being carried out jointly at the SICP - Software Innovation Campus Paderborn by member company OPTANO and the Business Informatics Data Analytics research group at Paderborn University under the direction of Prof Dr Oliver Müller, Director of the Artificial Intelligence competence area at the SICP, is therefore testing a new approach: in multi-agent reinforcement learning, virtual agents control the safety stocks at the individual warehouse locations and react independently to changes such as delivery delays, fluctuations in demand or short-term bottlenecks. With the help of a specially developed simulation environment, the agents "learn" in numerous experiments how to achieve a balance between low storage costs and high supply security - and dynamically adjust the stocks to current risks in the entire network. Marcel Meyer from Paderborn University categorises the results: "The RL approach opens up promising possibilities for adjusting safety stocks in complex networks in a targeted and dynamic manner. Further development is required to achieve optimal results, but the potential of this method is already evident compared to conventional methods."
Experience remains key
The experience and professional assessment of uncertainties by the experts involved is crucial to the performance of AI. Only on this basis can the learning agents find viable solutions - pure data models are not enough. The simulations show: RL opens up new ways of managing industrial supply chains economically and stably, even in the face of uncertainty.
