Genetic programming of polynomial harmonic networks using the discrete Fourier transform

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Intelligence

Background:

• Multilayer feed-forward neural networks are widely used for time series processing.
• Traditional neural network activation functions can limit performance with complex frequency data.
• Genetic programming offers a powerful framework for evolving network structures and functions.

Purpose of the Study:

• To present a novel hybrid genetic programming system for evolving polynomial harmonic networks.
• To investigate the efficacy of incorporating polynomial activation functions with harmonic inputs.
• To evaluate the performance of the proposed system in time series processing.

Main Methods:

• Development of a genetic programming system to evolve polynomial harmonic networks.
• Utilizing discrete Fourier transform to derive harmonics with non-multiple frequencies.
• Employing tree-structured topology for efficient evolutionary structural search.
• Comparing the proposed system against traditional genetic programming and harmonic GMDH algorithms.

Main Results:

• The evolved polynomial harmonic networks demonstrated superior performance in time series processing.
• The hybrid system outperformed evolutionary polynomial manipulation, Koza-style genetic programming, and harmonic GMDH.
• The novel approach effectively handles harmonics with non-multiple and irregular frequencies.

Conclusions:

• Polynomial harmonic networks represent a promising advancement in neural network architectures for time series analysis.
• The proposed genetic programming system offers an effective method for evolving complex neural network structures.
• This hybrid approach provides a robust solution for processing complex time series data with irregular frequency components.