A reinforcement learning algorithm for Dynamic Job Shop Scheduling

In the era of Industry 4.0, optimizing production processes has become increasingly critical due to the high demand for efficiency, flexibility, and customization in manufacturing. The Job Shop Scheduling Problem (JSSP), a prominent NP-hard problem, plays a pivotal role in this context, requiring the scheduling of jobs with multiple operations on specific machines in a predetermined order. Effective solutions to JSSP are essential for minimizing production time, reducing costs, and enhancing overall productivity. This thesis presents the development and evaluation of a single-agent reinforcement learning algorithm designed to address both the JSSP and its dynamic variant (DJSSP). The primary objective of this research is to test the efficiency and adaptability of reinforcement learning algorithm for scheduling solutions in both deterministic and dynamic environments characterized by variability and uncertainty. The proposed reinforcement learning approach autonomously learns optimal scheduling policies through iterative interactions with the scheduling environment, dynamically adapting to changes and unexpected disruptions. The algorithm's performance is rigorously benchmarked against traditional scheduling methods, including First-Come, First-Served (FCFS), Shortest Processing Time (SPT), and Genetic Algorithms (GA). Empirical results demonstrate that the reinforcement learning algorithm is comparable to traditional scheduling methods in the deterministic case, while outperforms conventional techniques in dynamic environments, exhibiting superior adaptability and efficiency across various scheduling scenarios. These findings underscore the significant potential of AI-driven methodologies to improve operational efficiency in complex scheduling tasks, offering valuable contributions to manufacturing, logistics, and other industries where optimal resource allocation is paramount.