Automatic determination of radial basis functions: an immunity-based approach

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Radial basis function (RBF) neural network performance relies heavily on basis function parameter selection.
• Fixed RBF parameters allow straightforward determination of output weights via linear least squares, reducing training time.
• A key challenge is efficiently determining the number, position, and dispersion of RBF centers.

Purpose of the Study:

• To develop an unsupervised clustering algorithm for RBF network parameter optimization.
• To propose a novel approach inspired by the vertebrate immune system.
• To automatically determine the number, position, and dispersion of RBF centers.

Main Methods:

• An algorithm inspired by vertebrate immune system models is introduced.
• The method performs unsupervised cluster analysis to define RBF parameters.
• Performance is evaluated against random, k-means, and literature methods.

Main Results:

• The proposed algorithm automatically defines RBF centers, leading to parsimonious solutions.
• Simulations demonstrate effectiveness in regression and classification problems.
• The immune-inspired approach offers an efficient alternative for RBF network training.

Conclusions:

• The immune-inspired algorithm provides an effective method for unsupervised RBF parameter determination.
• This approach simplifies RBF network training and improves efficiency.
• The method shows promise for various machine learning applications.