Two fast and accurate heuristic RBF learning rules for data classification.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computational Science

Background:

• Radial Basis Function (RBF) networks are widely used for classification.
• Determining optimal network parameters (spreads, centers, neuron count) is crucial for efficiency.
• Existing RBF learning methods can be complex and may result in large network sizes.

Purpose of the Study:

• To introduce novel RBF learning methods for classification.
• To achieve higher efficiency with a reduced number of neurons.
• To maintain a fast and simple learning algorithm.

Main Methods:

• Heuristics are employed to determine RBF network spreads, centers, and hidden neuron count.
• Neurons are added recursively, maximizing coverage of training data at each step.
• A termination condition based on data coverage or maximum neuron count is used.
• Power exponential distribution function serves as the hidden neuron activation function.

Main Results:

• The proposed methods result in networks with fewer neurons and lower Vapnik-Chervonenkis (VC) dimension.
• Improved generalization properties are demonstrated.
• Data becomes linearly separable in the hidden layer output space, enabling zero training error.
• Simulations show comparable performance against Support Vector Machines (SVM) and other leading RBF methods.

Conclusions:

• The developed RBF learning methods offer an efficient and effective approach to classification.
• The recursive neuron addition strategy optimizes network size and performance.
• The theoretical underpinning ensures linear separability and zero training error potential.