An enhanced blood-sucking leech optimization for training feedforward neural networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Optimization Algorithms

Background:

• Feedforward neural networks (FNNs) offer a scalable and efficient framework for computation.
• The Blood-Sucking Leech Optimization (BSLO) algorithm mimics leech foraging for global and local search.
• Training FNNs requires efficient optimization to determine connection weights and bias thresholds.

Purpose of the Study:

• To enhance the Blood-Sucking Leech Optimization (BSLO) algorithm by integrating the simplex method, creating the SBSLO.
• To train Feedforward Neural Networks (FNNs) using the proposed SBSLO algorithm.
• To evaluate the training efficiency, prediction accuracy, and robustness of SBSLO compared to existing methods.

Main Methods:

• Integration of the simplex method into the BSLO algorithm to form SBSLO.
• Utilizing SBSLO to optimize connection weights and bias thresholds in FNNs.
• Comparative analysis of SBSLO against 12 other optimization algorithms on 17 datasets.

Main Results:

• The SBSLO algorithm effectively trains FNNs, quantifying discrepancies between predicted and actual outputs.
• SBSLO demonstrates improved training efficiency, prediction precision, stability, and robustness.
• The enhanced algorithm leverages BSLO's exploration with the simplex method's exploitation, mitigating local optima and enhancing convergence.

Conclusions:

• The proposed SBSLO algorithm offers a superior approach for training FNNs.
• SBSLO enhances optimization performance by combining global exploration with refined local exploitation.
• The method shows significant improvements in efficiency, accuracy, stability, and convergence speed for FNN training.