Title: Enhancing Multi-Column RBF Neural Networks via Adaptive and Non-Adaptive Particle Swarm Optimization

Abstract:

Radial basis function neural networks (RBFNs) trained through gradient descent offer a robust, fully connected architecture suitable for both shallow and deep learning applications. Among gradient-based techniques, the Error Correction (ErrCor) method stands out as a state-of-the-art approach, optimizing accuracy by selecting the most effective hidden units. Conversely, Particle Swarm Optimization (PSO), a population-based algorithm, leverages collective swarm intelligence to optimize RBFN parameters, ensuring robustness against local minima and facilitating global search capabilities. An enhanced variant, Adaptive PSO (APSO), further refines this process by dynamically tuning swarm parameters to accelerate convergence. While both ErrCor and PSO yield competitive convergence and improved results, they encounter scalability issues with large datasets, including burdensome kernel computations and the need for extensive hidden layer structures.

To address these limitations, the Multi-Column RBFN (MCRN) framework has been introduced, boosting ErrCor performance by deploying multiple small RBFNs within a parallel architecture. Building on the success of MCRN, this study proposes two new methodologies to enhance PSO-based training: Multi-Column RBFN with PSO (MC-PSO) and Multi-Column RBFN with APSO (MC-APSO). These approaches utilize parallel RBFN structures trained via evolutionary swarm techniques. Each individual RBFN is trained independently on a specific spatial subset of the data using either PSO or APSO, resulting in specialists tailored to their respective subsets. During the testing phase, only those RBFNs associated with the spatial neighbors of the test instance contribute to the final multi-column output. This specialization mechanism boosts accuracy, while the parallel processing significantly improves computational speed.

We assessed the proposed methods across several benchmark datasets. The results indicate that both MC-PSO and MC-APSO surpass ErrCor, standard PSO, APSO, and MCRN in terms of accuracy and recall metrics. Furthermore, these proposed models exhibit reduced training and testing durations in the majority of experimental scenarios.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC