Metaheuristic Training of Multilayer Perceptrons for Medical Classification: A Comparative Study of HO, Puma, SFOA, and WSO

The performance of Multi-Layer Perceptron (MLP) models depends on the effective optimization of weight and bias parameters during the training process. Gradient-based learning methods may not always produce stable and high-performance results due to factors such as initial sensitivity, the need for derivative information, and stagnation at local minima in complex, high-dimensional search spaces. In this study, four current metaheuristic algorithms for optimizing the weight and bias parameters of MLPs were evaluated comparatively: Hippopotamus Optimization Algorithm (HO), Puma Optimizer, Starfish Optimization Algorithm (SFOA), and War Strategy Optimization Algorithm (WSO). The algorithms were tested on the Diabetes, Diagnosis, Mammographic Mass, and Wisconsin Breast Cancer datasets. The same MLP architecture and experimental conditions were used for each dataset; the data were evaluated with an 80% training and 20% testing split. Each method was run in 10 independent runs, with a population size of 30 and a number of iterations of 100, and candidate solutions were limited to the range [-10, 10]. Performance was measured by accuracy, precision, and F1-score; the overall ranking of the methods was examined using the Friedman test. The findings show that metaheuristic performance is sensitive to the dataset. In general, Puma-based training produced more competitive results in accuracy and F1-scores in most datasets, while HO excelled in precision and stability in some scenarios. Although SFOA and WSO produced competitive results in certain datasets, they did not exhibit consistent superiority across all metrics. The results demonstrate that the selection of meta-heuristic methods for MLP training should consider dataset characteristics and target metrics and provide a practical comparison framework for medical decision support applications. In addition, mean and standard deviation values are reported for each dataset, and the strengths of the methods are discussed based on the best run results.