Tuning extreme learning machine by an improved electromagnetism-like mechanism algorithm for classification problem.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Neural Networks

Background:

• Extreme Learning Machine (ELM) is a learning algorithm for single hidden-layer feedforward neural networks (SLFNs).
• ELMs offer fast learning, good generalization, and easy implementation compared to gradient-based methods.
• However, random input weights and hidden biases in ELMs necessitate more hidden neurons and lack optimal network structure guarantees.

Purpose of the Study:

• To develop a new learning algorithm that overcomes the limitations of traditional ELMs.
• To enhance the generalization performance and network structure optimization of ELMs.

Main Methods:

• A novel algorithm, DAEM-ELM, is proposed, combining an improved electromagnetism-like mechanism (EM) algorithm (DAEM) with Moore-Penrose (MP) generalized inverse.
• DAEM incorporates three distinct solution updating strategies inspired by the dragonfly algorithm (DA).
• Input weights and hidden biases are tuned via DAEM, while output weights are analytically determined using MP generalized inverse.

Main Results:

• The proposed DAEM-ELM algorithm demonstrates superior generalization performance.
• DAEM-ELM outperforms traditional ELM and other existing evolutionary ELM algorithms in experimental evaluations.

Conclusions:

• DAEM-ELM effectively addresses the disadvantages of traditional ELMs by optimizing network parameters.
• The integration of DAEM and MP generalized inverse offers a promising approach for enhancing ELM performance.