基于群密度的改进果蝇优化算法及在异常检测中的应用

针对传统果蝇算法面临的收敛稳定性差、难以协调全局搜索及局部搜索能力等缺点，提出一种基于群密度的改进果蝇优化算法。首先，借鉴现有算法的优势，将果蝇种群分为搜索果蝇和跟随果蝇，并分别使用两类果蝇进行全局化搜索与局部精细化搜索。然后，为提高算法全局搜索的稳定性，在每次迭代过程中使用基于最优区间回避的分区采样策略更新搜索果蝇的位置；该策略在每次迭代过程中获得表现最优的若干只果蝇以构造最优果蝇组，根据最优果蝇组中果蝇个体在每个维度上的取值范围确定最优区间，并通过对最优区间外的其他区间分区采样以确定搜索果蝇的新位置。最后，为协调算法的全局搜索能力与局部搜索能力，引入群密度的概念，通过计算果蝇群密度并结合相关阈值实现不同种群规模的动态调整。针对典型测试函数的实验结果表明，基于最优区间回避的分区采样策略相对于传统随机函数具有更强的全局优化性能。与传统优化算法相比，本文算法在保证收敛速度的同时获得了较高的寻优精度及稳定性，在综合性能上得到明显提升。在KDDcup99数据集上的异常检测仿真实验结果表明，本文基于分区采样及群密度的果蝇优化算法能有效避免局部最优，在获取异常检测分类器的重要参数最佳取值方面起到一定作用。

Improved Fruit Fly Optimization Algorithm Based on Population Density and Its Application in Anomaly Detection

To address the problems that the traditional fruit fly algorithms cannot steadily bad converge,and effectively balance the global and local searching ability,a novel fruit fly optimization algorithm based on population density was proposed.Firstly,by utilizing the advantages of existing methods,the fruit flies were divided into the searching fruit flies and the following fruit flies,which were then used for global searching and local searching,respectively.Secondly,in order to improve the stability of global searching process,the partition sampling strategy based on optimal interval avoidance was used to update the positions of searching fruit flies in each iteration process.The strategy selected the fruit flies with the best performances in each iteration to construct the optimal fruit fly group,and determined the optimal interval according to the ranges of the fruit flies in each dimension.Then,the new positions of the searching fruit flies were determined by sampling the interval except for the optimal interval.Finally,in order to balance the global and local searching ability,the conception of population density was introduced,and the dynamic population size adjustment of different types of fruit flies was achieved by thresholding the population density.The experimental results of typical test functions showed that the partition sampling strategy based on optimal interval avoidance achieved higher global optimization ability compared to traditional rand functions.Compared to traditional optimization algorithms,the proposed algorithm obtained high optimization accuracy and stability while guaranteeing the convergence speed,achieving obvious improvements on comprehensive performances.The simulation results of the anomaly detection showed that,the fruit fly algorithm based on the partition sampling and population density can avoid local optimum effectively,and is effective in obtaining the optimal values of important parameters of the anomaly detection classifier.