Statistical analysis of the parameters of a neuro-genetic algorithm

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Hybrid methods combining artificial neural networks (ANNs) and evolutionary algorithms (EAs) are increasingly utilized for their robustness.
• These methods offer advantages in designing ANNs by optimizing initial weights, architecture, and learning rules.
• The G-Prop method is a notable example of such a hybrid approach.

Purpose of the Study:

• To conduct an exhaustive analysis of the G-Prop method.
• To determine optimal parameters for G-Prop, including population size, selection rate, initial weight range, and training epochs.
• To investigate the impact of genetic operators on classification precision and network size.

Main Methods:

• Exhaustive analysis of the G-Prop method and its required parameters.
• Determination of parameter significance and relative importance using Analysis of Variance (ANOVA).
• Experimental comparison of G-Prop with varied parameter settings and other published hybrid methods.

Main Results:

• The study identified significant parameters influencing neural network performance and learning within hybrid methods.
• ANOVA confirmed the significance and relative importance of various G-Prop parameters.
• Experiments demonstrated the impact of genetic operators on classification accuracy and network complexity.

Conclusions:

• Parameter tuning is crucial for optimizing the performance of hybrid ANNs and EAs like G-Prop.
• The findings provide insights into suitable parameter values for enhancing classification tasks.
• The G-Prop method's effectiveness is validated through systematic parameter analysis and comparison.