Generalized regression neural networks with multiple-bandwidth sharing and hybrid optimization.

Area of Science:

• Machine Learning
• Neural Networks
• Data Science

Background:

• Standard generalized regression neural networks (GRNNs) often use a single bandwidth for all kernels, limiting flexibility.
• Previous methods for multiple bandwidth reduction in GRNNs typically rely on data clustering, which can be computationally intensive.

Purpose of the Study:

• To propose a novel algorithm for function approximation extending GRNNs with a multiple-bandwidth configuration.
• To develop an efficient scheme for dramatic bandwidth reduction while preserving model complexity.

Main Methods:

• The algorithm partitions training patterns into groups, assigning a shared bandwidth to each group.
• Grouping is based on local nearest neighbor distance analysis and principal component analysis with fuzzy clustering.
• A hybrid optimization procedure combines a particle swarm optimizer variant and a quasi-Newton method for bandwidth tuning.

Main Results:

• The proposed method achieves significant bandwidth reduction without compromising model complexity.
• Training utilizes a modified leave-one-out validation error to enhance network generalization.
• The algorithm demonstrates competitive regression performance against other techniques on diverse datasets.

Conclusions:

• The novel GRNN extension offers an effective approach to function approximation with optimized bandwidth selection.
• The method provides a balance between model complexity and generalization capability.
• This technique shows promise for applications requiring accurate regression analysis.