A rapid learning and dynamic stepwise updating algorithm for flat neural networks and the application to time-series...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Neural Networks

Background:

• Flat neural networks, particularly functional-link networks, are utilized for nonlinear function approximation.
• Traditional training methods can be computationally intensive and complex.

Purpose of the Study:

• To propose a fast learning algorithm for optimizing weights in flat neural networks.
• To enable efficient, on-the-fly weight updates for dynamic systems.

Main Methods:

• Formulating flat neural networks as linear systems solvable by linear least-square methods.
• Developing a dynamic stepwise updating algorithm for real-time weight adjustments.
• Testing the algorithm on diverse time-series data and nonlinear system identification.

Main Results:

• The proposed algorithm achieves optimal weights efficiently.
• Instantaneous weight updates are feasible for new data patterns and network enhancements.
• The model demonstrates strong performance on infrared laser, chaotic, and financial time-series data.

Conclusions:

• The proposed fast learning algorithm offers an attractive solution for real-time processes requiring efficient neural network training.
• This approach simplifies the training of flat neural networks by leveraging linear system formulations.
• The dynamic updating capability makes the model suitable for adaptive and evolving systems.