摘要
文中对如何从含有离散变量和连续变量的混合数据中学习 Bayesian网进行了研究 ,提出了一种基于遗传算法的连续变量离散化算法 .在该算法中给出了兼顾离散模型准确度和复杂度的适应度函数 ;并基于对离散化的实质性分析 ,定义了离散策略等价的概念 ,由此制定了离散策略的编码方案 ;进一步设计了变换离散策略的遗传算子 .该算法不存在局部极值问题 ,且不需要事先给定变量序关系 .模拟实验结果表明 ,该算法能有效地对连续变量离散化 ,从而使得从混合数据中学到的 Bayesian网具有较好性能 .
Based on genetic algorithm, this paper presents a discretization algorithm called DGA. Compared with univariate discretization that deals with the single variables individually, the result from DGA is more exact because it is a multivariate discretization method, whereby each variable is discretized taking into account its interaction with the other variables. Besides, it searches for the best discretization strategy by genetic operators, avoiding finding local maxima and specifying an ordering between the variables in advance which are inevitable for deterministic search adopted by previous discretization algorithms. In this paper, (1) fitness function is given that pays attention to not only accuracy and concision of discretization model, but also accuracy and concision of learned structure; (2) encoding is described by giving definition of discretization policy equivalence based on essential of discretization strategy; (3) genetic operators are designed that switch individuals to evolve good discretization policy. Experimental results show that this algorithm can effectively discretize continuous variables so that the learned structure takes on good performance.
出处
《计算机学报》
EI
CSCD
北大核心
2002年第8期794-800,共7页
Chinese Journal of Computers
基金
国家"八六三"高技术研究发展计划 ( 86 3-30 6 -ZD0 5 -0 1-2 )
国家自然科学基金 ( 6 98830 0 3)
教育部高校博士点专项科研基金项目
教育部符号计算与知识工程重点实验室资助
