Orthogonal least squares based complex-valued functional link network.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Neural Networks

Background:

• Functional link networks (FLNs) are single-layer neural networks that introduce nonlinearity.
• Traditional FLNs can face challenges with parameter learning and local minima in complex-valued domains.

Purpose of the Study:

• To present a fully complex-valued functional link network (CFLN) utilizing multivariate polynomials.
• To address the challenge of selecting an optimal subset of polynomial terms for efficient network design.
• To demonstrate the advantages of a purely complex domain computation for CFLNs.

Main Methods:

• Developed a complex-valued functional link network (CFLN) employing multivariate polynomials as nonlinear expansion terms.
• Utilized the orthogonal least squares (OLS) method constructively for parsimonious monomial subset selection.
• Evaluated the CFLN on function approximation, wind prediction, and nonlinear channel equalization tasks.

Main Results:

• The proposed CFLN, optimized with OLS, demonstrated a simpler network structure.
• Achieved favorable performance across diverse applications, including real-world data prediction.
• The complex-valued approach showed advantages in parameter count and design speed compared to real-valued equivalents.

Conclusions:

• The OLS-based CFLN offers an efficient and effective solution for nonlinear problems.
• Computing in the complex domain provides benefits in terms of network complexity and learning speed.
• The CFLN approach successfully avoids local minima, leading to robust parameter learning.