基于粗糙集和蚁群算法的决策库用于毒性分类  

Rough set theory and ant colony algorithm based decision rule base approach to toxic classification

作　　者：贺益君[1] 陈德钊[1] 

机构地区：[1]浙江大学化学工程与生物工程学系,浙江杭州310027

出　　处：《浙江大学学报（工学版）》2009年第3期481-485,共5页Journal of Zhejiang University：Engineering Science

基　　金：国家自然科学基金资助项目(20276063)

摘　　要：粗糙集理论适于处理离散属性,对于连续属性,需经离散化,其本质是搜索最小断点集以及最小属性约简,两者均为NP难问题,为此提出了样本可分辨矩阵和覆盖策略,并引入权重,将其归结为约束最小化问题,采用蚁群算法求解.引入了启发式信息的动态计算方法,并结合后验的信息素,计算选择概率,逐步引导蚁群可行解构造,2类信息的结合有助于提高寻优性能.将该方法用于2类同系化合物的毒性作用机制分类研究,可有效地实现断点集最小化和属性最小约简,由此便于建立分类规则库.相比判别分析、径向基网络和支持向量机3种方法,该规则库具有更加良好的预测性能,且易于专业分析和理解.Dealing with continuous attributes, discretization is needed as a preprocess step in the rough set theory, which in nature is to seek the minimal cutting points set. Solving minimal attribute reduction is another core task. Both have been shown as NP（non-deterministic polynomial time）-hard problems. The problems of the minimal cutting points set and minimal attribute reduction was syncretized as a unified constraint minimal problem based on the notion of the discernibility matrix and weight factor. The ant colony algorithm was proposed to solve this unified problem. The dynamic computation method of heuristic information was introduced, combined with the posterior pheromone, and the selection probability of each column can be computed, which can guide the feasible solution construction step by step. A combination of these two types of information helps to improve the optimization efficiency. The proposed method was applied to classification of organic compounds toxic action mechanisms and phenols toxic action mechanisms.Comparison of results showed that the proposed method had better prediction accuracies than the discriminant analysis, radial basis function neural network and support vector machine, and the classification model with rules was more understandable and easier to be explained by professional knowledge.

关 键 词：毒性分类 粗糙集理论 可分辨矩阵 蚁群算法 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]