计算机科学与探索 ›› 2020, Vol. 14 ›› Issue (8): 1409-1426.DOI: 10.3778/j.issn.1673-9418.1908037

• 人工智能 • 上一篇    下一篇

具有动态子空间的随机单维变异粒子群算法

邓志诚,孙辉,赵嘉,王晖   

  1. 1. 南昌工程学院 信息工程学院,南昌 330099
    2. 江西省水信息协同感知与智能处理重点实验室,南昌 330099
    3. 鄱阳湖流域水工程安全与资源高效利用国家地方联合工程实验室,南昌 330099
  • 出版日期:2020-08-01 发布日期:2020-08-07

Stochastic Single-Dimensional Mutated Particle Swarm Optimization with Dynamic Subspace

DENG Zhicheng, SUN Hui, ZHAO Jia, WANG Hui   

  1. 1. School of Information Engineering, Nanchang Institute of Technology, Nanchang 330099, China
    2. Jiangxi Province Key Laboratory of Water Information Cooperative Sensing and Intelligent Processing, Nanchang 330099, China
    3. National-Local Joint Engineering Laboratory of Water Engineering Safety and Effective Utilization of Resources in Poyang Lake Area, Nanchang 330099, China
  • Online:2020-08-01 Published:2020-08-07

摘要:

传统粒子群算法采用整体维度更新策略,常因某一维或某几维未达到最优解,导致粒子适应值变差。针对此问题,提出具有动态子空间的随机单维变异粒子群优化算法,从优质粒子全维空间中,构造动态子空间,并随机选择异于子空间的一维进行变异。子空间大小动态变化:前期选取多数维度组成子空间,增大变异维度的多样性;后期选取少数维度组成子空间,增强粒子精细搜索的能力。同时,根据Pareto定律,使种群在前期20%迭代次数内,探索新解空间区域,后期80%迭代次数内,进行有效的平衡搜索,加快种群收敛速度。使用多类型基准测试函数,在30、50和100维下进行仿真实验,结果表明,该算法在收敛速度和精度上,不仅优于新改进的粒子群算法,而且优于新改进的人工蜂群算法和萤火虫算法。

关键词: 粒子群优化算法(PSO), 单维变异, 动态子空间, Pareto定律

Abstract:

The traditional particle swarm optimization algorithm adopts the overall dimension updating strategy, and the particle??s fitness value deteriorates frequently because the optimal solution is not reached in a certain dimension or a few dimensions. Aiming at this problem, a stochastic single-dimensional mutated particle swarm optimization algorithm with dynamic subspace is proposed. The dynamic subspace is constructed from the high-quality particle's dimension, and one-dimension different from the subspace is randomly selected to mutate. The subspace size changes dynamically. In the early stage, most of the dimensions are used to form the subspace, which increases the diversity of the mutated dimension. Later, a few dimensions are selected to form the subspace, which enhances the ability of the particle to search fine. At the same time, according to Pareto's law, the population explores the new solution space region within the first 20% iterations, and performs an effective balanced search within 80% of the iterations in the later period to accelerate the population convergence speed. Simulation experiments are carried out in 30, 50 and 100 dimensions using multi-type benchmark functions. The results show that the proposed algorithm is not only superior to state of the art particle swarm optimization algorithms, but also the artificial bee colony and firefly algorithm in convergence speed and precision.

Key words: particle swarm optimization (PSO), single-dimensional mutation, dynamic subspace, Pareto's law