Parallel-machine scheduling: variants, formulations, and a critical analysis of metaheuristic approaches

• Ano de defesa: 2024
• Autor(a) principal: Rolim, Gustavo Alencar
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: aprendizado por reforço; atrasos e adiantamentos; earliness and tardiness; makespan; meta-heurísticas; metaheuristics; parallel-machine scheduling; programação da produção em máquinas paralelas; reinforcement learning; tempo de fluxo; tempo total de produção; total completion time
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/18/18156/tde-25062025-095024/

This thesis focuses on the study of the unrelated parallel-machine scheduling problem and its variants. In recent years, this problem has gained significant attention due to its wide range of applications, from semiconductor manufacturing to the execution of computer programs. Despite its practical importance, many of its variants are N P-hard, making the solution of large-sized instances challenging for most exact approaches. Therefore, this thesis pursues two major objectives: first, to propose new variants of the unrelated parallel-machine scheduling problem and develop fast and efficient heuristic solutions; and second, to survey and evaluate the current literature to identify the algorithmic components that lead to better metaheuristic results. To achieve these objectives, we structured this work into three stages. In the first stage, we investigated a parallel machine scheduling problem where jobs are subject to a common due window, with the goal of minimizing the total sum of weighted earliness and tardiness penalties. We identified structural properties, presented two mathematical formulations, developed two constructive heuristics, and extended the adaptive large neighborhood search metaheuristic. In the second stage, we conducted a comprehensive survey of the current literature on metaheuristics for scheduling problems with setups and makespan minimization. We analyzed neighborhood structures, solution representations, and demonstrated the advantages of using acceleration techniques. Additionally, we implemented and evaluated a set of 12 heterogeneous metaheuristics under standardized experimental conditions. Finally, in the third stage, we studied a parallel batch scheduling problem with job families and setups, aimed at minimizing the total sum of completion times. We introduced a mathematical formulation, established a dominance relationship, and proposed both a constructive heuristic and a stochastic local search algorithm with multiple neighborhood structures. We also explored the incorporation of Q-learning, a model-free reinforcement learning algorithm, for selecting neighborhood structures. Overall, we not only proposed effective algorithms that outperformed existing approaches across various benchmark instances, but also demonstrated that methods with diverse neighborhood structures and few control parameters tend to deliver superior outcomes.