Q-learning-driven multi-population cooperative evolutionary algorithm with local search for scheduling of network-shared manufacturing resources

Abstract

The variability in the availability of network-shared manufacturing resources and the release times of orders pose challenges to the operational decision-making of industrial internet platforms. This paper addresses these characteristics by studying the identical parallel machine scheduling problem, aiming to minimize total weighted tardiness under constraints of arbitrary release times and multiple machine unavailability periods. To address this research problem, a decoding mechanism based on machine idle periods is first proposed, effectively solving the impact of machine unavailability periods on the scheduling scheme. Secondly, a multi-population cooperative evolutionary algorithm is designed in which the mechanisms of selection, crossover, mutation, and information exchange between populations are improved. The optimal scheduling properties of two jobs on the same machine and different machines are analyzed, resulting in the design of two local search mechanisms. Additionally, Q-learning is introduced to enhance the adaptability of algorithm parameters by dynamically adjusting them within the multi-population cooperative evolutionary algorithm, resulting in a Q-learning-driven multi-population cooperative evolutionary algorithm with local search (Q-MPCEA-LS). Finally, comparative experiments between the Q-MPCEA-LS algorithm and various metaheuristic algorithms are conducted. The experimental results show that, across all instances, the average relative error in the average value metric of the Q-MPCEA-LS algorithm is 40.0%, 0.1%, 44.2%, and 75.9% lower than that of Q-MPCEA-LS without local search, Q-MPCEA-LS without Q-learning-based dynamic parameter adjustment, the iterative hybrid metaheuristic algorithm, and the hybrid genetic immune algorithm, respectively. These results validate the effectiveness of the individual components and the overall effectiveness of the Q-MPCEA-LS algorithm.