A Modified General Regression Neural Network (MGRNN) with new, efficient training algorithms as a robust 'black...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Neural Networks

Background:

• General Regression Neural Networks (GRNN) provide a foundation for regression tasks.
• Existing feed-forward networks often face practical challenges with data structures and gradient computation.

Purpose of the Study:

• To introduce a Modified General Regression Neural Network (MGRNN) as an accessible 'black box' tool.
• To enhance the training efficiency and data flexibility of regression neural networks.
• To enable gradient computation for regression surfaces.

Main Methods:

• The MGRNN architecture is based on the established GRNN, with determined network structure and weights.
• Kernel widths are trained using two supervised algorithms, requiring minimal, robust user-defined parameters.
• Generalization capabilities are evaluated using benchmarks like intertwined spirals, Mackey-Glass time series, and PROBEN1.

Main Results:

• The MGRNN demonstrates effective regression surface generation from input data.
• The training algorithms are fast, reliable, and robust to parameter variations.
• The MGRNN successfully handles arbitrary data structures without requiring equal-length vectors and computes gradients.

Conclusions:

• The MGRNN presents a practical and versatile solution for regression problems.
• It overcomes common limitations of traditional feed-forward neural networks.
• The network's ability to handle varied data and compute gradients enhances its applicability.