A New Initialization Approach in Particle Swarm Optimization for Global Optimization Problems | JoVE Visualize

Area of Science:

Computational Intelligence
Optimization Algorithms
Machine Learning

Background:

Particle Swarm Optimization (PSO) is a population-based metaheuristic inspired by bee swarming behavior, widely applied to optimization problems.
Effective population initialization is crucial for PSO's diversity and convergence, with quasirandom sequences outperforming random initialization.
Existing initialization methods like Sobol-based PSO (SO-PSO) and Halton-based PSO (H-PSO) show promise but can be further improved.

Purpose of the Study:

To introduce a novel population initialization technique, WELL (using low-discrepancy sequences), to enhance PSO performance.
To develop and evaluate the WE-PSO algorithm for solving large-dimensional optimization problems.
To compare WE-PSO against standard PSO and other quasirandom-based PSO variants.

Main Methods:

The proposed WE-PSO algorithm utilizes the WELL initialization technique based on low-discrepancy sequences.
WE-PSO was validated on fifteen standard unimodal and multimodal benchmark test problems.
The approach was also applied to artificial neural network (ANN) learning and compared with backpropagation and other PSO variants.

Main Results:

WE-PSO demonstrated superior performance compared to standard PSO and other multimodal problem-solving techniques on benchmark functions.
The proposed method achieved higher accuracy scores in artificial neural network learning tasks than standard backpropagation, H-PSO, and SO-PSO.
The study highlights the significant impact of the WE-PSO initialization technique on cost function quality, integration, and diversity.

Conclusions:

The WE-PSO algorithm, with its novel WELL initialization, effectively enhances population diversity and convergence in PSO.
WE-PSO offers a robust and efficient solution for complex optimization problems, including large-dimensional search spaces and ANN training.
The findings underscore the critical role of initialization strategies in optimizing the performance of population-based algorithms.