多策略改进哈里斯鹰优化算法及其在高维特征选择中的应用

Multi-Strategy Improved Harris Hawks Optimization Algorithm and Its Application in High-Dimensional Feature Selection

摘要: 为解决原始哈里斯鹰优化算法(HHO)在迭代过程中易陷入局部最优、种群多样性下降、后期收敛精度不足等问题，提出一种融合多策略的改进哈里斯鹰优化算法(IHHO)。该算法首先引入经验互益策略，整合历史种群的有效搜索信息，增强种群个体间的信息交互能力，提升算法跳出局部最优的潜能；其次设计能量自适应扰动策略，依据猎物逃逸能量动态调整扰动比例，实现全局广域探索与局部精准开发的动态平衡；最后提出精英方差引导开发策略，通过计算精英个体的逐维方差确定局部搜索步长，解决算法后期搜索的局部振荡问题，提升局部优化精度。通过时间复杂度分析明确了IHHO的计算成本，同时给出了算法的完整实现步骤。为验证算法的性能，选取CEC2020标准测试函数集，将其与原始HHO、白鲸优化算法(BWO)、算术优化算法(AOA)等5种经典群智能优化算法进行对比实验，设置相同的实验参数与迭代条件，从最优值、平均值、标准差等指标进行分析，并通过Friedman统计检验与收敛曲线验证算法的综合性能。实验结果表明，所提出的算法在CEC2020测试函数集上的Friedman平均排名为1.43，显著优于对比算法，在复杂高维非线性优化问题中展现出更优的全局寻优能力、收敛速度与鲁棒性。将IHHO算法应用于高维特征选择问题，以k-最近邻分类错误率为适应度函数，选取8个高维数据集开展实验，从适应度值、分类准确率与特征压缩效果等维度进行验证。结果显示，IHHO在Breast、CNS、Leukemia等多数高维数据集中取得了更低的平均适应度值与更高的分类准确率，部分数据集分类准确率达到100%，所选特征子集兼具良好的判别性与泛化性，有效实现了高维特征的精准筛选与冗余特征消除。研究表明，所提多策略改进方法能显著提升哈里斯鹰优化算法的综合性能，并在连续函数优化与高维特征选择领域均具有良好的应用效果与推广价值。

Abstract: To address the problems that the original Harris Hawks Optimization (HHO) algorithm is prone to falling into a local optimum, suffers from the decline of population diversity, and has insufficient convergence accuracy in the later iteration stage, an improved Harris Hawks Optimization algorithm fused with multiple strategies (IHHO) is proposed. Firstly, an experience mutual benefit strategy is introduced into the algorithm to integrate the effective search information of historical populations, enhance the information interaction ability among population individuals, and improve the algorithm’s potential to jump out of the local optimum. Secondly, an energy-adaptive perturbation strategy is designed to dynamically adjust the perturbation ratio based on the prey’s escape energy, thereby achieving a dynamic balance between global wide-area exploration and local precise development. Finally, an elite variance-guided exploitation strategy is proposed, which determines the local search step size by calculating the dimension-wise variance of elite individuals, thus solving the problem of local oscillation in the later search stage of the algorithm and improving the local optimization accuracy. The computational cost of IHHO is clarified through time complexity analysis, and the complete implementation steps of the algorithm are given simultaneously. To verify the optimization performance of IHHO, the CEC2020 standard test function set is selected for comparative experiments with five classic swarm intelligence optimization algorithms including the original HHO, Beluga Whale Optimization (BWO) and Arithmetic Optimization Algorithm (AOA). The same experimental parameters and iteration conditions are set, and the performance is analyzed from the indicators such as the optimal value, mean value and standard deviation. In addition, the comprehensive performance of the algorithm is verified by the Friedman statistical test and convergence curves. The experimental results show that the Friedman average ranking of IHHO on the CEC2020 test function set is 1.43, which is significantly better than that of the comparison algorithms. IHHO exhibits superior global optimization ability, convergence speed and robustness in solving complex high-dimensional nonlinear optimization problems. The IHHO algorithm is applied to the high-dimensional feature selection problem, with the k-nearest neighbor classification error rate as the fitness function. Eight high-dimensional datasets are selected for experiments, and the algorithm is verified from the dimensions of fitness value, classification accuracy and feature compression effect. The results show that IHHO achieves lower average fitness values and higher classification accuracy in most high-dimensional datasets such as Breast, CNS and Leukemia, with the classification accuracy reaching 100% in some datasets. The selected feature subsets have both good discriminability and generalization ability, which effectively realize the accurate screening of high-dimensional features and the elimination of redundant features. The research shows that the proposed multi-strategy improvement method can significantly improve the comprehensive performance of the Harris Hawks Optimization algorithm, and IHHO has good application effects and popularization value in the fields of continuous function optimization and high-dimensional feature selection.