基于多尺度选择性学习和探测-收缩机制的PSO算法

A PSO Algorithm Based on Multiscale-Selective-Learning and Detecting-Shrinking Strategies

针对粒子群算法逃离局部最优能力差、易早熟收敛、求解精度低等缺点,提出了一种具有多尺度选择性学习和探测-收缩机制的PSO算法.在多尺度选择性学习机制中,粒子根据其自身进化状态在拓扑结构、邻居个体、目标变量维等多个尺度上进行选择性学习,提升粒子个体的学习效率;在探测-收缩机制中,算法利用历史信息指导种群最优解进行探测,提高其逃离局部最优的能力,当判断种群历史最优解处于全局最优解附近时,执行空间收缩策略,将种群的搜索空间限定在较小的一个区域,增强算法的开采能力,提高算法的求解精度.通过和其它PSO算法在22个典型测试函数的实验对比表明,本算法能有效克服早熟收敛、加快收敛速度、提高求解精度.

To overcome the shortcomings the traditional particle swarm optimization algorithm （PSO）, such as poor ability to escape a local optimal ,premature convergence and low precision,we proposed a new PSO based on multiscale-se- lective-learning and detecting-shrinking strategies,which called MDPSO in short. In the multiscale-selective-leaming strate- gy, a particle executes a multiscale learning process to improve its studying efficiency by adopting its topology, selecting neighbors, and choosing target variable dimensions. In the detecting-shrinking strategy ,particles＇ historical best solutions are periodic sampling and some useful information, which extracting from the sampling results, is used to direct the best solutions to carry out a detecting operation. The aims of the strategy are to improve PSO＇s global searching ability and to help the popu- lation escape a local optimal solution. While the best solution situating around a global optimal solution, the algorithm imple- ments the shrinking strategy to confine the search space to a small one the aims of which are to improve the PSO＇s exploitation ability and to increase the accuracy of the solutions. The proposed method was applied to twenty-two typical benchmark functions, and the comparisons of the performance between MDPSO and other eight PSO algorithms were experimented. The results suggest that the proposed strategies can effectively overcome the premature convergence, speed up the convergence and improve solutions accuracy.