基于互信息下粒子群优化的属性约简算法

摘要

最小属性约简是粗糙集理论中属性约简的优化问题.在寻找最小属性约简的问题上,基于粒子群优化的属性约简算法(ARPSO算法)优于传统的属性约简算法.在现有的ARPSO算法中,正域部分通常被作为启发式信息,但是它并不能够很好地衡量不确定性,而互信息是粗糙集理论中一种更有效的度量不确定信息的重要工具.为此,提出基于互信息下的粒子群优化的属性约简算法(MIPSO算法),该算法把互信息作为适应度函数,通过增强粒子能迅速靠近吸引子的这一特性,改进了内嵌区域震荡搜索的粒子群优化算法(简记为RSPSO算法),防止算法较早的陷入局部最优,使得粒子群中的粒子更快的找到最优值,因此使得算法尽可能实现全局收敛.实验结果表明,该算法不仅提高了寻优的能力,加快了算法的速度,提升了算法的精度,而且也能够使得约简后剩余属性的互信息值与约简前所有属性的互信息值近似相等.%Minimum attribute reduction is the optimum problem in the attribute reduction of the rough sets theory.To seek the minimum attribute reduction,the attribute reduction algorithm based on the particle swarm optimization (ARPSO algorithm) beats the traditional attribute reduction algorithm.In existed ARPSO algorithms,the positive region is usually taken as the heuristic information,however,it is not precision enough to measure the uncertainty.The mutual information is a more efficient tool to measure the uncertainty in the rough sets theory.To handle this problem,an attribute reduction algorithm based on the particle swarm optimization takes the mutual information(MIPSO algorithm)as a term in the fitness function,The proposed MIPSO algorithm improves the regional shock search embedded particle swarm optimization algorithm (RSPSO) by enhancing the speed which the particle is close to the attractor,preventing from being local optimum early and finding the optimum as soon as possible.Consequently,the global convergence of the MIPSO algorithm is guaranteed as soon as possible.The experimental results show that the proposed MIPSO algorithm not only improves the optimization ability,accelerates the speed and improves the accuracy,but also can keep the mutual information value of all attributes before reducing approximately equal to the value of remaining attributes after reducing.