A Software Architecture for the Application of Chaos Engineering in Reinforcement Learning-Based Production Scheduling

Resilient production scheduling is essential for industrial production systems to maintain performance under dynamic and disruptive conditions. Reinforcement learning offers a promising solution by enabling systems to learn reactive scheduling strategies through interaction with the environment. However, current reinforcement learning-based scheduling implementations predominantly focus on reactive behaviour and lack mechanisms for anticipating and proactively preparing for disruptions. To address this limitation, this paper investigates the integration of Chaos Engineering into the reinforcement learning training phase to improve system resilience. Chaos Engineering introduces synthetic disruptions in a controlled manner, allowing the reinforcement learning agent to learn from varied failure scenarios and enhance its reactivity in real-world operations. Building on the concepts of reinforcement learning, Chaos Engineering, and Advanced Planning and Scheduling system architectures, a microservice software architecture is developed. The architecture defines dedicated services for optimization, disruption management, consistency checking, and user interaction. Each service is linked to systematically derived requirements and supported by resilience-oriented data structures. The architecture is validated using the Scenario-based Architecture Analysis Method, with a focus on modifiability and future extensibility. The results demonstrate high modularity and suitability for implementation in industrial environments. This work establishes a foundation for integrating resilienceenhancing methods.