A Novel Black Widow Optimization Algorithm Based on Lagrange Interpolation Operator for ResNet18.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Hyper-parameters significantly influence neural network training and performance.
• Optimizing the learning rate is crucial but challenging due to task/dataset dependency and trial-and-error methods.
• Evolutionary computation offers automated hyper-parameter tuning for improved efficiency.

Purpose of the Study:

• To propose a novel algorithm, LIBWONN (Lagrange interpolation-based black widow optimization algorithm), for optimizing neural network learning rates.
• To enhance the training efficiency and performance of ResNet18 models.
• To address the complexities and time-consuming nature of manual learning rate tuning.

Main Methods:

• Developed LIBWONN, integrating Lagrange interpolation with the black widow optimization algorithm.
• Evaluated LIBWONN on 24 benchmark functions from CEC2017 and CEC2022.
• Compared LIBWONN against nine advanced metaheuristic algorithms.
• Tested LIBWONN's performance on ResNet18 using six diverse, publicly available datasets.

Main Results:

• LIBWONN demonstrated superior convergence and stability compared to nine other metaheuristic algorithms on benchmark functions.
• LIBWONN achieved significant accuracy improvements on both training (6.99%) and testing (4.48%) sets for ResNet18.
• The proposed algorithm outperformed the standard Black Widow Optimization (BWO) algorithm.

Conclusions:

• LIBWONN effectively optimizes the learning rate for ResNet18, surpassing existing metaheuristic approaches.
• The integration of Lagrange interpolation enhances the performance of the black widow optimization algorithm.
• LIBWONN offers a promising automated solution for improving neural network training and generalization across various applications.