Towards Real-World Deployment of Reinforcement Learning

Reinforcement Learning (RL) has gained significant research interest over the past decades. However, its application to industrial problems is rare, as many studies focus on simplified problems or RL-specific benchmarks that are not directly transferable to real-world scenarios.In this thesis, we address this research gap by developing RL solutions for three concrete real-world problems. The first use case involves applying RL to control a traffic light system in Lemgo, Germany. We developed a framework, which includes a realistic simulation model and a safety layer to ensure compliance with all legal requirements. Furthermore, we integrate the safety layer more deeply into RL algorithms to achieve faster training and better performance. Additionally, we explore methods to bridge the reality gap caused by discrepancies between the simulation model and its real-world counterpart. For the second and third use cases, we developed RL-based solutions to tackle real production scheduling problems in the household appliance and automotive industries. The second use case applies RL as a constructive heuristic, meaning that the solution is built up progressively. To make the problem solvable by RL, we integrate domain knowledge using techniques such as action masking and curriculum learning. In the third case, we explore how RL can be used as an improvement heuristic, where a suboptimal initial solution is improved by RL through iterative small adjustments. Through these case studies, we demonstrate how RL can be adapted and integrated with domain knowledge to address the specific needs and constraints of real-world environments.