EVALUATION OF PARTICLE SWARM ALGORITHM MODIFICATIONS ON SUPPORT VECTOR MACHINE HYPERPARAMETER OPTIMIZATION TUNING FOR RAIN PREDICTION

Abstract

The Particle Swarm Optimization (PSO) algorithm, though simple and effective, faces challenges like premature convergence and local optima entrapment. Modifications in the PSO structure, particularly in acceleration coefficients ( and ), are proposed to address these issues. Techniques like Time Varying Acceleration Coefficients (TVAC), Sine Cosine Acceleration Coefficients (SCAC), and Nonlinear Dynamics Acceleration Coefficients (NDAC) have been implemented to enhance convergence speed and solution quality. This research evaluates various PSO modifications for improving convergence and robustness in rainfall potential prediction using Support Vector Machine (SVM) classification. The UAPSO-SVM algorithm C=0.82568 and γ=0.01960 excels in initial exploration, discovering more optimal global solutions with smaller variability. In contrast, TVACPSO-SVM shows gradual improvement but requires more iterations for stability, while SBPSO-SVM achieves the fastest convergence at iteration 14 but risks overfitting. Robustness analysis reveals all PSO-SVM variants maintain stable performance despite variations in dataset subset sizes, with accuracy stabilizing after a spike at 20%.. Therefore, PSO modifications enhance convergence speed and resilience to data fluctuations, improving their effectiveness for rainfall prediction.

Keywords

• Modified Particle Swarm Optimization
• Support Vector Machine
• , Robustness
• Convergence