Algoritmo evolutivo multiobjetivo para o escalonamento de produção com restrição na indústria farmacêutica

• Ano de defesa: 2024
• Autor(a) principal: Kohara, Debora Toshie
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Uberlândia; Brasil; Programa de Pós-graduação em Ciência da Computação
• Palavras-chave em Português: Otimização Multiobjetivo com restrição; Algoritmos Genéticos Multiojetivos; Escalonamento de Bateladas; Manufatura Farmacêutica; Constrained Multiobjective Optimization; Multiobjective Genetic Algorithms; Batch Scheduling; Pharmaceutical Manufacturing; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO; Computação; ODS::ODS 9. Indústria, Inovação e infraestrutura - Construir infraestrutura resiliente, promover a industrialização inclusiva e sustentável, e fomentar a inovação.
• Link de acesso: https://repositorio.ufu.br/handle/123456789/44694; http://doi.org/10.14393/ufu.di.2024.637

Batch scheduling in chemical processes under real-world conditions, such as demand variability and lost sales constraints, has been investigated. These problems are typically non-linear with computationally expensive evaluation functions, making solution convergence challenging. They are characterized by multiple conflicting objectives and constraints, increasing the complexity of exploring viable solutions. This complicates the adoption of certain techniques, but Multiobjective Evolutionary Algorithms (MOEAs) have been explored due to their flexible and efficient operators, applicable to various non-linear multiobjective problems. In this work, new strategies for population initialization, individual selection, and local search-based mutation are incorporated into an NSGA-II-based MOEA, applied to batch scheduling in a pharmaceutical production line as a multiobjective constrained problem. The quality of non-dominated solutions was evaluated considering the number of valid solutions (NS) and multiobjective metrics such as hypervolume (hv), error rate (E), and inverted generational distance (IGD+). Experimental results showed that the evolutionary approach significantly improved the reference model in both static environments (where a single demand set is used for evaluation) and dynamic environments (where demands change over MOEA generations). In static environments, our approach, compared to the reference, improved by 23.3% the average of E, 82.5% in IGD+, 0.3% in hv, and had a 30.2% deterioration in NS. In dynamic environments, it improved by 38.7% in E, 83.7% in IGD+, 86.6% in NS, and 0.8% in hv. We observed that integrating local search into the mutation operator reduced E by 10.5%, IGD+ by 0.1%, hv by 0.1%, and NS by 10.7%, compared to the algorithm using only the proposed initialization and selection methods. Dominance analysis indicated that local search in the number of batches is effective in dynamic scenarios.