Alawad, N. A., Abed-alguni, B. H., Paul, D., and Al-Betar, M. A. “The effect of different heuristic initialization approaches on the performance of binary population-based optimization algorithms”, Journal of Computational Science, 98, 2026.
Population initialization in binary population-based optimization algorithms (BPOAs) plays a pivotal role in improving the convergence speed towards optimal solutions, particularly in feature selection problems. This study compares and evaluates frequently used heuristic population initialization procedures within a unified and controlled experimental framework. All strategies are embedded within the same Binary Differential Evolution (BDE) algorithm and evaluated using identical parameter settings, fitness functions, and real-world benchmark datasets to ensure fair comparison. Nine initialization methods are assessed in terms of population diversity, classification accuracy, and selected feature subset size. Experimental results demonstrate that, compared to other commonly used initialization methods, the k-means clustering and Sobol quasi-random sequence approaches consistently produce higher population diversity. When integrated into the same BDE framework, these strategies also achieve superior classification accuracy while selecting smaller feature subsets across all datasets. In contrast, Gaussian perturbation-based initialization exhibits low diversity and inferior performance. The findings demonstrate the strong impact of population initialization on optimization effectiveness and provide practical guidance for selecting appropriate initialization strategies in binary population-based optimization for real-world feature selection applications.