Open Access

Stefan Sarawinsky, Kai Schaller, Kajan Kandiah

• Improving the sustainability of production systems requires decision support methods that can solve the problem of optimizing the combination of resources under ecological, operational, and compatibility constraints. This challenge arises in agriculture and manufacturing, particularly in Circular Industrial Production Systems (CIPS), where the creation of value depends on the coordinated use of materials, processes and resources, while minimizing waste and environmental impact. As production environments become more complex and dynamic, traditional rule-based or static optimization approaches struggle to cope with combinatorial decision spaces and evolving system constraints. To address this issue, the authors use intercropping as an example proxy domain with high complexity that mirrors these combinatorial challenges and compatibility constraints. This paper therefore presents a Deep Q-Network (DQN)-based Reinforcement Learning (RL) approach embedded in a domain-specific simulation environment for planning intercropping sequences. The RL-based Decision Support System (DSS) learns to maximize an aggregated reward comprising agronomic quality indicators and binary user preference feedback. Results from a simulation demonstrate a consistent convergence and structured long-term policy learning in complex combinatorial decision spaces. Beyond agriculture, this framework shows how RL-based sequencing policies can help CIPS facing similar optimization problems involving material combinations, process compatibility, and sustainability trade-offs.