Q-Learning-Driven Accelerated Iterated Greedy Algorithm for Multi-Scenario Group Scheduling in Distributed Blocking Flowshops

Abstract

This paper focuses on the distributed blocking flowshop group scheduling problem under multiple processing time scenarios and due dates (DBFGSP_UPT). Initially, a mathematic model is formulated to achieve a balance between the mean and standard deviation of total tardiness across various scenarios, and its correctness is validated via the Gurobi solver. Next, an accelerated iterated greedy algorithm integrated with Q-learning selection mechanism ($QAIG$) is proposed. The $QAIG$ involves: a rapid evaluation method, tailored for the total tardiness criterion by using a hierarchical approach, which is first proposed to significantly reduce the time complexity of the insertion-based method; a self-calibrating parameter method, which dynamically selects appropriate numbers of groups to be destroyed, is designed to improve the diversity of solutions; and a Q-learning mechanism is integrated into the local search strategy framework to facilitate the selection of high-quality local search schemes. Finally, we conduct a comparative analysis across 810 test instances. Comprehensive numerical experiments and comparative analyses demonstrate that the proposed $QAIG$ surpasses existing state-of-the-art algorithms in terms of the average relative percentage increase.