Learning Subspace-Based RBFNN Using Coevolutionary Algorithm for Complex Classification Tasks.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computational Neuroscience

Background:

• Real-world classification tasks often involve complex data distributions.
• Classification accuracy depends on intrinsic sample properties within feature subspaces.

Purpose of the Study:

• To propose a novel algorithm for constructing Radial Basis Function Neural Network (RBFNN) classifiers using subspace learning.
• To improve classification accuracy and network simplicity for complex datasets.

Main Methods:

• Feature subspaces are generated for each RBFNN hidden node during learning.
• Connection weights are adjusted to create subspaces with dominant features.
• A coevolutionary algorithm optimizes network structure and dominant features via two subpopulations.

Main Results:

• The proposed algorithm yields RBFNN models with superior classification accuracy.
• The developed models exhibit simpler network structures compared to existing algorithms.
• The approach effectively utilizes local sample characteristics within subspaces.

Conclusions:

• The novel subspace learning algorithm offers a more flexible and efficient method for complex classification problems.
• This approach enhances RBFNN performance by leveraging feature subspace properties.
• The coevolutionary strategy ensures global optimality for the estimated RBFNN.