Learning Dispatching Rules for Flexible Assembly Scheduling Problems Using Deep Reinforcement Learning

The Job Scheduling Problem and its extensions have been extensively investigated in the past decade, utilizing various solving methods such as heuristics, metaheuristics, and even deep reinforcement learning (DeepRL) and Graph Neural Networks (GNN). Unfortunately, real-world scheduling problems are difficult to solve efficiently, leading to significant time consumption when attempting to solve large industrial instances. While heuristic-based dispatching methods offer a way to compute schedules within a reasonable time frame, creating effective heuristics that produce satisfactory solutions is challenging and laborious. In this paper, we analyze the effectiveness of DeepRL approaches when using to solve a class of flexible Assembly Scheduling Problems (ASP) inspired by real-world industrial scenarios. The starting point of our work is the Schlably framework (https://github.com/tmdt-buw/schlably). We aimed to extend its capabilities to accommodate in-tree precedence relations characteristic of ASP. Additionally, to further enrich the framework's versatility, we implemented ASP-specific dispatching rules and analyzed the influence of the observation space on the performance of DeepRL expressed through the schedule makespan.