Abstract
For multi-objective particle swarm optimization, the selection of the global best becomes an interesting topic, because it balances convergence and diversity. But the global best selected by the existing strategies has a high probability of not dominating the particle. The flight towards the global best not dominating the particle is expected to cause some objectives to become worse, thus does not surely promote convergence. As the accumulation of the flights, the algorithm suffers from slow convergence. Therefore we propose the dominating strategy to accelerate the convergence by decreasing that probability. Experimental results show our strategy outperforms other strategies.
Introduction
Multi-objective Particle Swarm Optimization (MOPSO) aims to find a set of optimal trade-offs, known as the Pareto front, by simulating the movement of particles through a search space. A critical component of this process is the selection of a "global best" (gBest) leader for each particle to follow. The choice of gBest must balance two competing goals: convergence (approaching the true Pareto front) and diversity (maintaining a wide distribution of solutions).
Current strategies, such as grid-based or random selection, often choose a gBest that does not strictly dominate the current particle. When a particle moves toward a non-dominating leader, it may improve in some objectives while regressing in others. This stochastic "one step forward, one step back" behavior results in inefficient search paths and slow convergence.
The Dominating Strategy
To address these inefficiencies, we propose the Dominating Strategy. This method prioritizes selecting a leader that strictly dominates the current particle, ensuring that the expected flight path improves the particle across all objectives.
The selection process follows three steps:
- Initialization: A member is randomly selected from the archive of non-dominated solutions.
- Dominance Search: The algorithm traverses the archive to identify a member that dominates the current particle (i.e., better in at least one objective and not worse in any).
- Fallback: If a dominating member is found, it is assigned as the gBest. If no such member exists after a full search, the algorithm reverts to the initial random selection.
As shown in the table above, the Dominating strategy offers "Good" optimality and supports asynchrony, which is essential for modern parallel computing environments, with a time complexity of $O(kAN)$.
Experimental Evaluation
The strategy was evaluated using the Walking Fish Group (WFG1-WFG9) test suite, which includes complex problems featuring non-separability, deception, and degeneration. We compared our proposed method (Strategy A) against four standard approaches: Grid (B), Random (C), Sigma (D), and Nearest (E).
Performance was measured using two metrics:
- Dominating Ratio: The frequency with which the selected gBest dominates the particle.
- Hypervolume (HV): A combined measure of convergence and diversity.
Results and Discussion
The experimental results demonstrate that the Dominating Strategy (A) consistently achieves a significantly higher dominating ratio than all competing methods across all WFG benchmarks. This confirms that the strategy successfully identifies leaders that provide a "sure" direction for convergence.
Furthermore, in terms of Hypervolume, Strategy A outperformed the Grid, Random, Sigma, and Nearest strategies in nearly every test case. While the time complexity is slightly higher than purely random selection, the gains in convergence speed and solution quality justify the computational cost, especially in scenarios with expensive objective evaluations.
Conclusion
The proposed Dominating Strategy effectively accelerates MOPSO convergence by ensuring that particles follow leaders that offer clear Pareto improvements. While the strategy excels at convergence, future work will explore hybridizing it with grid-based methods to further enhance the uniformity and diversity of the final solution set.
Summary updated