Tuning of the structure and parameters of a neural network using an improved genetic algorithm

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Neural networks require effective methods for structure and parameter tuning.
• Standard genetic algorithms (GAs) have limitations in optimizing complex network architectures.
• Dynamic adaptation of network structure alongside parameter learning is a significant challenge.

Purpose of the Study:

• To present an improved genetic algorithm (GA) for optimizing neural network structure and parameters.
• To introduce a novel neural network architecture incorporating switches for adaptive learning.
• To demonstrate the efficacy of the improved GA and proposed network in real-world applications.

Main Methods:

• Development of an improved genetic algorithm (GA) with enhanced optimization capabilities.
• Design of a neural network architecture featuring switches in its links.
• Manual selection of hidden nodes based on achieving satisfactory learning performance (fitness value).

Main Results:

• The improved GA demonstrated superior performance over the standard GA on benchmark test functions.
• The proposed neural network successfully learned both input-output relationships and network structure.
• Application examples in sunspot forecasting and associative memory validated the method's effectiveness.

Conclusions:

• The improved GA offers a more effective approach for neural network tuning compared to standard methods.
• The proposed switch-enabled neural network architecture facilitates adaptive learning of both function and structure.
• The presented methodology shows significant potential for complex learning tasks like time-series forecasting and memory modeling.