Scheduling has been a complex problem across most industries, especially in the context of online scheduling, where information availability can often be limited or uncertain. Companies are therefore, tasked with making appropriate scheduling decisions to enhance their business value. This thesis begins with evaluating the performance of widely employed scheduling methods, then examine whether production aggregation could improve efficiency through iterative simulations. The results of the experiments suggest that production aggregations require more holding costs and do not reduce the machine changeover time as we expected. Nonetheless, the overall reward does not show any sign of differences compared to the priority heuristic. Beyond analyzing the scheduling method, we created DRL environments based on the heuristics and trained them with DQN and PPO. Notably, DQN yielded significant performance improvements, especially in the production aggregation environment. Unfortunately, the PPO algorithm does not properly converge due to its sensitive nature in model design and hyperparameter tuning. In conclusion, this thesis provides valuable insights on online scheduling, stochastic manufacturing systems, production aggregation, and shed the light of primary application of DRL in online scheduling.     «

Scheduling has been a complex problem across most industries, especially in the context of online scheduling, where information availability can often be limited or uncertain. Companies are therefore, tasked with making appropriate scheduling decisions to enhance their business value. This thesis begins with evaluating the performance of widely employed scheduling methods, then examine whether production aggregation could improve efficiency through iterative simulations. The results...     »