| dc.description.abstract | The Job Shop Scheduling Problem (JSSP) consists of allocating various tasks to distinct machines, each of which has a different sequence of operations. This thesis investigates the application of Reinforcement Learning (RL) algorithms in addressing the JSSP, focusing primarily on instances from the Lawrence, Dermikol, and Taillard datasets. Particularly, the study evaluates popular RL algorithms, including Proximal Policy Optimization (PPO), Policy Gradient (PG), Advantage Actor-Critic (A2C), and Asynchronous Advantage Actor-Critic (A3C), against traditional dispatching rules and other state-of-the-art methods. The findings highlight the superior performance of the PPO approach, which consistently outperforms alternative RL algorithms and dispatching rules across various instance sizes. PPO demonstrates robustness and adaptability in navigating dynamic job-shop scheduling landscapes, positioning it as a versatile and potent solution for learning complex scheduling strategies. Insights from the training dynamics of RL agents underscore their ability to improve performance over time by learning from the environment. The increasing reward values and decreasing makespan observed across all datasets signify the adaptive nature of RL
agents in optimizing scheduling policies. Remarkably, the PPO-based approach demonstrates a 6-9 times lower optimality gap compared to traditional scheduling algorithms and achieves a 2-3 times lower optimality gap than state-of-the-art approaches in all three datasets, highlighting its superiority in addressing the JSSP. This thesis contributes to the understanding of the efficacy of RL algorithms in scheduling optimization, suggesting their potential to significantly enhance operational efficiency across diverse industrial sectors. | |
