CUDA-AMSR-PSO: A NUMBA-CUDA ACCELERATED ADAPTIVE MULTI-SWARM REINFORCED PARTICLE SWARM OPTIMIZATION FRAMEWORK FOR SCALABLE FEATURE SELECTION IN BIG DATA ANALYTICS

Abstract

Feature selection in high-dimensional data involves balancing computational cost with model performance, especially in wrapper-based methods where models are retrained repeatedly. In this work, we present CUDA-AMSR-PSO, a GPU-accelerated approach based on adaptive multi-swarm particle swarm optimization with a reinforcement-inspired method for adjusting inertia weights during the search. Using Numba-CUDA, particle updates are performed in parallel within a single kernel, which helps the method handle large feature spaces more efficiently. Instead of relying on fixed parameters, the algorithm adapts its behaviour based on observed improvements in fitness, while the multi-swarm structure helps reduce premature convergence. The fitness function also considers relationships between features, encouraging the selection of compact but informative subsets. Experiments show that the approach achieves effective dimensionality reduction with competitive classification performance. Profiling results indicate that CPU-based wrapper evaluations remain the main bottleneck, suggesting that further improvements will require optimizing both the search process and the evaluation stage.