Comparative Evaluation of Random, Sobol, and Dragon-Family-Guided PSO and Quantum Inspired PSO on CEC 2022 Benchmarks and Cleveland Heart Disease Wrapper Feature Selection

Author:
Publisher website
Google Scholar
Abstract
In Particle Swarm Optimization (PSO) and Quantum Inspired PSO (QI-PSO), the initialization source can shape early population diversity, search coverage, convergence behavior, and the later response of candidate generation operators. This effect becomes important in both continuous optimization, where particles move in a real-valued search space, and wrapper feature selection, where candidate solutions are evaluated as binary feature masks through classifier performance. Dragon curves offer another way to define structured search sources because their ordered point sequences can be mapped into a normalized search domain and reused as geometric guides. Based on this idea, this study examines whether random initialization, Sobol initialization, QI candidate generation, and dragon-curve-based geometric guidance change method behavior across continuous optimization and wrapper feature selection. The PSO variant set included Random-PSO, Sobol-PSO, Heighway-PSO, Twindragon-PSO, and Terdragon-PSO, while the QI-PSO variant set included Random-QI-PSO, Sobol-QI-PSO, Heighway-QI-PSO, Twindragon-QI-PSO, and Terdragon-QI-PSO. Heighway, Twindragon, and Terdragon curves were used as ordered geometric sources for particle initialization and stagnation-triggered perturbation. The methods were tested on CEC 2022 benchmark functions in 10-dimensional and 20-dimensional settings and on the UCI Cleveland Heart Disease dataset through wrapper feature selection with KNN-based fitness and additional SVM evaluation. In CEC 2022, the PSO variants produced lower average ranks and lower average mean fitness values than the QI-PSO variants. Random-PSO achieved the lowest average rank and runtime, whereas Terdragon-PSO produced the lowest average mean fitness and the lowest mean best fitness trajectory among the PSO variants. Dragon-guided PSO variants produced more best mean fitness records, but with higher computational cost. In the Cleveland task, Sobol-QI-PSO produced the highest observed mean balanced accuracy among the wrapper FS methods under both KNN and SVM. However, the Holm-corrected comparisons indicated statistically similar balanced accuracy across the compared wrapper methods. Terdragon-PSO produced the highest feature subset stability, while Random-PSO produced the most compact subsets. These results show that initialization, QI candidate generation, and dragon-based guidance vary with problem representation and evaluation dimension. Future studies should evaluate new metaheuristic components through final fitness, runtime, classifier response, selected feature count, selection frequency, and feature subset stability rather than through one performance measure.

This publication has 47 references indexed in Scilit: