Scholarlink

<p>This paper proposes a parallelization method for accelerating MOPSO/D without degrading its solution accuracy. The approach uses particle swarm optimization (PSO) for solving multi-objective optimization problems. To address the computational overhead associated with increasing the swarm size, we introduce two distinct approaches within the decomposition-based MOPSO (MOPSO/D) framework. First, we present a virtual global best (Vgbest) method that enables efficient sharing of the best solutions found during the optimization process across parallel processing units. Second, we propose a method that divides the objective space so that subproblems assigned to each particle can be explored in parallel. Since partitioning the objective space tends to degrade solution accuracy near region boundaries, we propose a method for compensating by sharing historical local best solutions among neighboring cores. The distributed parallel MOPSO/D with the Vgbest method is implemented in Python on a multi-core CPU and evaluated using standard benchmark problems. Experimental results indicate that the proposed methods effectively reduce computation time, achieving up to approximately 2.15 times speedup, while generally maintaining comparable solution accuracy.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Parallelization of MOPSO/D on multi-core CPUs with virtual global best and overlap-aware partitioning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月20日

查看全文
  • Mikiko Sato,
  • Mitsuhiko Uchida,
  • Yuji Sato

摘要

This paper proposes a parallelization method for accelerating MOPSO/D without degrading its solution accuracy. The approach uses particle swarm optimization (PSO) for solving multi-objective optimization problems. To address the computational overhead associated with increasing the swarm size, we introduce two distinct approaches within the decomposition-based MOPSO (MOPSO/D) framework. First, we present a virtual global best (Vgbest) method that enables efficient sharing of the best solutions found during the optimization process across parallel processing units. Second, we propose a method that divides the objective space so that subproblems assigned to each particle can be explored in parallel. Since partitioning the objective space tends to degrade solution accuracy near region boundaries, we propose a method for compensating by sharing historical local best solutions among neighboring cores. The distributed parallel MOPSO/D with the Vgbest method is implemented in Python on a multi-core CPU and evaluated using standard benchmark problems. Experimental results indicate that the proposed methods effectively reduce computation time, achieving up to approximately 2.15 times speedup, while generally maintaining comparable solution accuracy.