Abstract
The classical Job Shop Scheduling Problem (JSSP) focuses on optimizingmakespan under deterministic constraints. Real-world production environmentsintroduce additional complexities that cause traditional scheduling approachesto be less effective. Reinforcement learning (RL) holds potential in addressingthese challenges, as it allows agents to learn adaptive scheduling strategies.However, there is a lack of a comprehensive, general-purpose frameworks foreffectively training and evaluating RL agents under real-world constraints. Toaddress this gap, we propose a modular framework that extends classical JSSPformulations by incorporating key \mbox{real-world} constraints inherent to theshopfloor, including transport logistics, buffer management, machinebreakdowns, setup times, and stochastic processing conditions, while alsosupporting multi-objective optimization. The framework is a customizablesolution that offers flexibility in defining problem instances and configuringsimulation parameters, enabling adaptation to diverse production scenarios. Astandardized interface ensures compatibility with various RL approaches,providing a robust environment for training RL agents and facilitating thestandardized comparison of different scheduling methods under dynamic anduncertain conditions. We release JobShopLab as an open-source tool for bothresearch and industrial applications, accessible at:https://github.com/proto-lab-ro/jobshoplab