A Fast Adaptive Tunable RBF Network For Nonstationary Systems.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Neural Networks

Background:

• Radial basis function (RBF) neural networks are widely used for system identification.
• Traditional RBF networks often struggle with dynamic environments and require significant parameter tuning.
• On-line learning methods are crucial for adapting to changing system dynamics.

Purpose of the Study:

• To develop a novel on-line learning approach for RBF neural networks.
• To enhance the adaptability and performance of RBF networks in dynamic system identification.
• To introduce a flexible and fast method for real-time system modeling.

Main Methods:

• Utilized an RBF network with individually tunable nodes and a fixed, small model size.
• Employed the multi-innovation recursive least squares algorithm for on-line weight vector adaptation.
• Implemented a node replacement strategy for insignificant nodes when residual error increases.
• Developed fast algorithms for optimizing the structural parameters of newly added nodes.

Main Results:

• The proposed on-line learning scheme demonstrated significant improvements in modeling performance.
• The adaptive RBF network effectively handled nonstationary systems.
• The method proved to be flexible and fast for on-line system identification tasks.
• Simulation results indicated superior performance compared to existing approaches.

Conclusions:

• The novel on-line learning RBF neural network offers a robust solution for system identification.
• The adaptive node management and parameter optimization enhance performance in dynamic environments.
• This approach provides a significant advancement for real-time modeling of nonstationary systems.