Application of cluster analysis methods for dynamic search space adjustment in genetic algorithms

This study investigates the application of cluster analysis techniques to improve the efficiency of genetic algorithms (GAs) in solving multidimensional global optimization problems, particularly those relevant to the aerospace industry. The research focuses on dynamic search space adjustment in GAs through statistical filtering of individual clusters. The proposed methodology involves: (1) developing a dynamic correction approach for variable domains by partitioning the population into clusters using both fixed-number clustering algorithms (k-means, k-medians, agglomerative, and spectral clustering) and density-based methods (DBSCAN); (2) evaluating cluster quality metrics including population size and average fitness; and (3) eliminating clusters that contribute insignificantly to the evolutionary process. The primary objective is to enhance algorithm convergence speed by 25–30 % (as demonstrated in benchmark testing) while maintaining solution quality in mixed optimization problems through effective search space adaptation at each iteration. The three-stage method comprises: (1) current population clustering, (2) elimination of clusters with below-average population size and fitness, and (3) dynamic boundary adjustment for remaining individuals' domains. Experimental results demonstrate the method's potential for integration into aerospace design systems, significantly reducing computation time while improving parameter optimization accuracy. Furthermore, the approach shows promise for hyperparameter optimization in various machine learning models, particularly in neural network architecture synthesis - including deep neural networks and specialized topologies.