Initialization by a novel clustering for wavelet neural network as time series predictor.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Neural Networks

Background:

• Wavelet Neural Networks (WNNs) are powerful tools for complex pattern recognition tasks.
• Effective parameter initialization is crucial for WNN performance and training efficiency.
• Existing initialization methods, like Zhang's heuristic approach, have limitations in automatic structure determination.

Purpose of the Study:

• To propose a novel, dynamic clustering-based initialization method for Wavelet Neural Networks (WNNs).
• To enable automatic determination of WNN architecture, including neuron number and initial parameter values.
• To improve the stability and speed of the WNN optimization process.

Main Methods:

• A dynamic clustering procedure is introduced to derive WNN structure and initial parameters.
• The method utilizes input patterns and activating wavelet functions for parameter derivation.
• Performance is evaluated through three simulation examples, comparing against Zhang's heuristic approach.

Main Results:

• The proposed method automatically determines the optimal WNN structure.
• It provides superior initial parameter values compared to Zhang's method.
• The new initialization leads to a more stable and faster optimization process.

Conclusions:

• The novel dynamic clustering approach offers an effective solution for WNN initialization.
• Automatic structure determination and improved parameter initialization enhance WNN training.
• This method presents a significant advancement for WNN applications requiring efficient learning.