Multiobjective evolutionary optimization of the size, shape, and position parameters of radial basis function...

Area of Science:

• Computational Intelligence
• Machine Learning
• Neural Networks

Background:

• Optimizing radial basis function neural networks (RBFNNs) is crucial for approximating complex target functions from input-output data.
• Existing evolutionary algorithms may benefit from enhanced operators for more efficient parameter tuning.

Purpose of the Study:

• To develop and evaluate a multiobjective evolutionary algorithm for optimizing RBFNNs.
• To introduce novel genetic operators inspired by singular value decomposition (SVD) and orthogonal least squares (OLS) for improved network performance.

Main Methods:

• A multiobjective evolutionary algorithm was designed to optimize RBFNNs.
• New mutation operators were developed using SVD and OLS matrix transformations.
• These operators introduce local or global modifications to the radial basis functions (RBFs) within the network's population.

Main Results:

• The efficiency of various genetic operators was analyzed.
• Global mutation operators demonstrated superior performance in adjusting RBFNN parameters.
• The proposed evolutionary approach effectively optimizes RBFNNs for function approximation tasks.

Conclusions:

• The integration of SVD and OLS-based mutation operators enhances evolutionary optimization of RBFNNs.
• Global mutation operators are particularly effective for fine-tuning RBFNN parameters.
• This work provides an improved methodology for developing accurate RBFNN models.