<p>Feature selection constitutes a fundamental component of machine learning and Genetic Algorithms (GAs) are extensively employed in feature selection. However, conventional GAs are afflicted by premature convergence and difficulty in preserving population diversity. To mitigate these limitations, this study proposes a real-coded multi-population dynamic competitive genetic algorithm (MPDCGA) for feature selection. In this innovative framework, the population initialization mechanism based on mRMR and cosine similarity furnishes a diverse initial solution, the dynamic competition operator explores the optimal feature subset through coevolutionary processes, and the adaptive similarity crossover operator improves the global search efficiency while augmenting the capability to extract potentially salient features. To comprehensively evaluate the performance of MPDCGA, adequate experiments were conducted on 16 UCI datasets. The experimental results demonstrate that MPDCGA effectively circumvents the limitations of local optimality, achieving superior feature selection accuracy and robustness.</p>