贝叶斯网参数学习中连续变量离散化方法研究

A Discretization Method of Continuous Variable in Bayesian Network Parameter Learning

连续变量离散化是贝叶斯网络参数学习中面临的一个重要问题,它的好坏将直接影响到贝叶斯网络的推理效果。目前缺少一种有效的手段用于评价连续变量离散化的好坏,通过研究,提出了推理信息量的概念,并采用作为衡量连续变量离散化好坏的标准。在连续变量离散化的过程中,采用遗传算法通过迭代的方式寻求最优解,其中,推理信息量作为衡量个体适应度的标准。实例分析证明,推理信息量大的推理效果好要优于推理信息量小的推理效果。

Discretization of continuous variable is a very important problem in Bayesian network parameter learning,and it affects the effectiveness of Bayesian network reasoning directly. There is a lack of standard used to evaluate discretization of continuous variable at present, therefore a notion of reasoning information is presented here by study ,which is used as the measure standard of diseretization . According to the simple Bayesian network, a diseretization model is built, and the optimal solution is found by GA , and in this process , reasoning information is used as the function of individual fitness degree . Example analysis proved that using greater amout of reasoning information can obtain better reasoning effect than using less reasoning information.