A fully complex-valued radial basis function network and its learning algorithm.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Signal Processing

Background:

• Complex-valued neural networks offer advantages in processing complex data.
• Existing complex-valued radial basis function (CRBF) networks have limitations in their activation functions and parameter initialization.

Purpose of the Study:

• To propose a novel fully complex-valued radial basis function (FC-RBF) network with an improved activation function and learning algorithm.
• To enhance the performance of CRBF networks through effective neuron selection and parameter initialization.

Main Methods:

• Development of a fully complex-valued activation function (sech(.)) with Gaussian-like characteristics, mapping C(n) to C.
• Implementation of a K-means clustering-based scheme for neuron selection and center initialization.
• Evaluation using complex XOR, synthetic function approximation, two-spiral classification, non-minimum phase equalization, and adaptive beam-forming problems.

Main Results:

• The proposed FC-RBF network exhibits superior convergence properties compared to existing CRBF networks.
• Demonstrated enhanced approximation and classification abilities in various synthetic and real-world complex-valued tasks.
• Outperformed split-complex CRBF, CMRAN, and CELM in benchmark comparisons.

Conclusions:

• The proposed FC-RBF network with its novel activation function and initialization scheme provides a significant advancement in complex-valued machine learning.
• The FC-RBF network is a promising tool for complex signal processing and data analysis tasks.