A Compact Constraint Incremental Method for Random Weight Networks and Its Application.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Neural Networks

Background:

• Incremental random weight networks (IRWNs) often suffer from poor generalization and complex structures.
• This is largely due to random, unguided assignment of learning parameters, leading to redundant hidden nodes and reduced performance.

Purpose of the Study:

• To develop a novel Incremental Random Weight Network with Compact Constraint (CCIRWN) that guides the assignment of random learning parameters.
• To address the limitations of weak generalization and complicated network structures in traditional IRWNs.

Main Methods:

• Introduced a compact constraint using the Greville iteration method for learning parameter configuration.
• Ensured simultaneous quality of hidden nodes and convergence of the CCIRWN.
• Analytically assessed the output weights of the CCIRWN.
• Proposed two learning methods for constructing the CCIRWN.

Main Results:

• The CCIRWN demonstrated favorable generalization ability across various applications.
• Performance was evaluated on 1-D nonlinear function approximation, real-world datasets, and industrial data-driven estimation.
• Numerical and industrial examples confirmed the effectiveness of the compact structure.

Conclusions:

• The proposed CCIRWN effectively overcomes the limitations of traditional IRWNs.
• The compact constraint approach leads to improved generalization and a more streamlined network architecture.
• CCIRWN offers a promising solution for applications requiring efficient and accurate function approximation and estimation.