Robustness, evolvability, and optimality of evolutionary neural networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Optimization problems seek optimal function solutions by adjusting variables.
• Artificial Neural Networks (ANNs) minimize error surfaces during training for better generalization.
• Optimal training does not guarantee optimal generalization due to overfitting/underfitting.

Purpose of the Study:

• To evaluate the evolvability, optimality, and robustness of SEPA with different perturbation functions.
• To investigate the impact of perturbation functions on SEPA's performance.
• To confirm the effectiveness of simultaneous ANN structure and weight adaptation.

Main Methods:

• Developed SEPA (Structure Evolution and Parameter Adaptation) for simultaneous ANN structure and weight evolution.
• Tested SEPA's performance using various perturbation functions.
• Analyzed generalization performance, stability, and robustness.

Main Results:

• SEPA demonstrated stable and robust generalization performance across different perturbation functions.
• The feedback loop between architecture evolution and weight adaptation compensates for individual shortcomings.
• Simultaneous adaptation prevents bias towards either weight or architecture space.

Conclusions:

• Simultaneous adaptation of ANN structure and weights is crucial for effective ANN design.
• SEPA's design ensures robust generalization by balancing architecture and weight optimization.
• This approach mitigates issues of overfitting and underfitting in ANNs.