Multi-Objective Multi-Exemplar Particle Swarm Optimization Algorithm With Local Awareness

Many machine learning algorithms excel at handling problems with conflicting objectives. Multi-Objective Optimization (MOO) algorithms play a crucial role in this process by enabling them to navigate these trade-offs effectively. This capability is essential for solving complex problems across diverse scientific and engineering domains, where achieving optimal solutions often requires balancing multiple objectives. One of these MOO algorithms Multi-Objective Particle Swarm Optimization (MOPSO) extends it to handle problems with multiple objectives simultaneously, but like many swarm-based algorithms, MOPSO can suffer from premature convergence or local optima solutions. Therefore, this article introduces a novel Multi-Exemplar Particle Swarm Optimization with Local Awareness (MEPSOLA) as a potent solution. The algorithm presents a developed multi-objective-aware criterion for multi-exemplar selection, adeptly balancing exploration and exploitation to avoid local optima and enhance performance across multiple objectives. It also introduces a conditional and periodic Tabu search tailored specifically for exemplar selection enhancement, improving both exploration and exploitation capabilities and avoiding premature convergence. Additionally, our method employs an improved initialization phase using equal sampling for each decision variable to ensure a comprehensive exploration of the entire solution space. A comprehensive assessment utilizing standard mathematical functions such as Fonseca-Fleming (FON), Kursawe (KUR), ZDT1, ZDT2, ZDT3, and ZDT6, and a comparison with state-of-the-art benchmarks in the field such as the Multi-Objective Evolutionary Algorithm (MOEA), Non-Dominated Sorting Genetic Algorithm (NSGA-II), and NSGA-III, validate the efficiency of MEPSOLA. Notably, MEPSOLA’s solutions outperform other benchmarks in key metrics across the majority of mathematical problems, for instance in set coverage, where our method dominates other methods’ solutions by 99.22%, 69%, and 93.58 %, respectively, highlighting its enhanced capability in optimizing capability within complex multi-objective optimization scenarios.