遗传算法在决策表连续属性离散化中的应用研究

连续属性的离散化是压缩数据和简化分析的重要手段 ,也是模式识别、机器学习和粗集分析等领域研究的难点 .目前已出现多种离散方法 ,存在的主要问题是对离散效果影响较大的侯选分割点集选择带有较强的主观性 .最优离散化是 NP-困难问题 ,大多数离散化算法采用的启发式也难以得到较满意的离散效果 .基于粗集理论 ,探讨了上述问题 ,把分割点的优选问题转化为 0 -1整数规划 ,并提出一种用实数编码的遗传算法来计算最优分割点集.

The Discretization of Continuous Attributes Using Genetic Algorithms

The discretization of continuous attributes is an important method for compressing data and simplifying analysis, which is of the focuses in the domains of pattern recognition, machine learning and rough sets. Some discretization algorithms have been used such as MD, discretization based on entropy but there exist disadvantages in them. For example, the choice of initial set of cut dots is hard to be determined. The optimal discretization has been proved to be NP\|hard. Heuristics used by most algorithms usually give local minima though results sometimes are satisfactory. Based on the rough set theory, the problems mentioned above are firstly discussed in this paper. Then we transform the discretization of continuous attributes into 0\|1\|integer programming, which can be solved successfully by existent software such as lindo. Furthermore, a genetic algorithm using decimal encoding is proposed to compute the optimal discretization.