用于水声目标识别的自适应遗传样本选择算法

针对训练样本集中含有噪声样本、冗余样本以及无关样本，导致分类系统分类性能下降、不稳定的水声目标识别问题，提出了一种新的自适应遗传样本选择算法(Adaptive Genetic Instance Selection Algorithm, AGISA)。算法先随机生成初始种群，接着利用设计的遗传算子(跨代选择、自适应交叉和简化最近邻变异)指导种群进化，每代中对分类贡献大且选择样本数目少的个体适应度值高。提取了实测3类水声目标的多域特征，进行样本选择和分类识别仿真实验，结果表明：AGISA可以选出有效样本子集，在样本维数下降约73%的情况下，支持向量机分类器的正确分类率能提高约2.5%；并且AGISA具有较好的收敛性、稳定性，所得优化样本子集具有较好泛化能力且能明显减少分类的时间。

An adaptive genetic instance selection algorithm for underwater acoustic target classification

In this paper, a new adaptive genetic instance selection algorithm (AGISA) is proposed for underwater acoustic target classification. The AGISA is proposed to address the problem that the classification performance in classifying underwater acoustic targets declines and becomes unstable as the training instance set contains noise samples, redundant samples and irrelevant samples. The AICISA generates an initial population randomly, and then generates new generations through designed genetic operators (cross-generational selection, adaptive crossover and reduced nearest neighbor mutation). In each generation, antibodies with less number of features and with high classification accuracy are given higher fitness values. The multi-field features are extracted from 3 classes of underwater targets, and used in instance selection and classification experiments. Experimental results show that AGISA can select the subset of efficient instances, and there is about 2% increase in the accuracy of SVM classifier when the number of features decreases about 70%. AGISA has good convergence and stability, and the instance subset obtained by AGISA achieves good generalization ability, which can reduce the classification time obviously.