On the structure and initial parameter identification of Gaussian RBF networks

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Intelligence

Background:

• Radial basis function (RBF) networks are powerful tools for function approximation and classification.
• Efficient initialization of RBF network structure and parameters is crucial for optimal performance.
• Fuzzy decision trees offer a comprehensible model for rule-based systems.

Purpose of the Study:

• To propose an efficient initialization scheme for generalized Gaussian RBF networks.
• To leverage fuzzy decision trees for RBF network structure and parameter initialization.
• To investigate the functional equivalence between fuzzy decision trees and RBF networks.

Main Methods:

• Utilizing fuzzy ID3-like induction algorithms to generate fuzzy decision trees.
• Establishing a functional equivalence property between fuzzy decision trees and generalized Gaussian RBF networks.
• Employing stochastic gradient descent for RBF network learning.

Main Results:

• The proposed scheme enables efficient initialization of RBF network structure and parameters.
• The resulting RBF networks are compact and comprehensible.
• The initialized RBF networks demonstrate acceptable classification accuracy.

Conclusions:

• Fuzzy decision trees provide an effective basis for initializing generalized Gaussian RBF networks.
• The method results in RBF networks that are easy to induce and understand.
• The approach balances network complexity with classification performance.