A new algorithm for online structure and parameter adaptation of RBF networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Neural Networks

Background:

• Online adaptation of radial basis function (RBF) neural networks is crucial for modeling dynamic systems.
• Existing methods often struggle with simultaneous structural and weight adaptation or efficient extrapolation.

Purpose of the Study:

• To present a new adaptive training method for RBF neural networks capable of online structural and weight modification.
• To address the challenges of modeling time-varying dynamical systems, including extrapolation and network size management.

Main Methods:

• The proposed algorithm dynamically modifies the number of hidden layer nodes and output weights.
• Network centers are selected based on a fuzzy partition of the input space, ensuring coverage.
• Inactive hidden nodes are pruned to maintain network efficiency.

Main Results:

• The adaptive method successfully models time-varying dynamical systems, demonstrating effective extrapolation.
• The algorithm manages network size by removing inactive hidden node centers.
• Performance is validated through applications and comparisons with other adaptive training techniques.

Conclusions:

• The novel adaptive training method offers robust online learning for RBF neural networks in dynamic environments.
• It effectively balances adaptation, extrapolation, and network resource management.
• This approach is well-suited for real-time applications involving evolving systems.