Radial basis function network training using a nonsymmetric partition of the input space and particle swarm...

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computational Science

Background:

• Radial Basis Function (RBF) networks are powerful tools for modeling complex data.
• Training RBF networks often involves challenges in determining optimal network structure and parameters.
• Existing methods may lack efficiency or lead to overly complex models.

Purpose of the Study:

• To introduce a novel, integrated algorithm for the complete training of RBF networks.
• To enhance model accuracy and parsimony through an optimized training process.
• To reduce computational time while improving predictive performance.

Main Methods:

• A nonsymmetric fuzzy means (FM) algorithm is employed to determine RBF centers.
• Linear regression is utilized for calculating synaptic weights.
• Particle Swarm Optimization (PSO) is integrated to optimize the fuzzy partition, creating a PSO-based nonsymmetric FM algorithm.

Main Results:

• The proposed algorithm successfully determined all RBF network parameters.
• Evaluated on 12 benchmark datasets, the RBF networks trained by the new method showed superior performance.
• Models generated exhibited higher prediction accuracies and simpler structures compared to other techniques.

Conclusions:

• The PSO-based nonsymmetric FM algorithm offers a robust and efficient approach for RBF network training.
• This method effectively balances accuracy, parsimony, and computational speed.
• The findings suggest a significant advancement in neural network training methodologies.