A reinforcement learning enhanced memetic algorithm for multi-objective flexible job shop scheduling toward Industry 5.0

Flexible job shop scheduling problem (FJSP) with worker flexibility has gained significant attention in the upcoming Industry 5.0 era because of its computational complexity and its importance in production processes. It is normally assumed that each machine is typically operated by one worker at any time; therefore, shop-floor managers need to decide on the most efficient assignments for machines and workers. However, the processing time is variable and uncertain due to the fluctuating production environment caused by unsteady operating conditions of machines and learning effect of workers. Meanwhile, they also need to balance the worker workload while meeting production efficiency. Thus a dual resource-constrained FJSP with worker’s learning effect and fuzzy processing time (F-DRCFJSP-WL) is investigated to simultaneously minimise makespan, total machine workloads and maximum worker workload. Subsequently, the reinforcement learning enhanced multi-objective memetic algorithm based on decomposition (RL-MOMA/D) is proposed for solving F-DRCFJSP-WL. For RL-MOMA/D, the Q-learning is incorporated into memetic algorithm to perform variable neighbourhood search and further strengthen the exploitation capability for the algorithm. Finally, comprehensive experiments on extensive test instances and a case study of aircraft overhaul shop-floor are conducted to demonstrate effectiveness and superiority of the proposed method.