Genetic clustering algorithm

Abstract

The genetic algorithm of clustering of analysis objects in different data domains has been offered within the hybrid concept of intelligent information technologies development aimed to support decision-making. The algorithm makes it possible to account for different preferences of the analyst in clustering reflected in a calculation formula of fitness function. The place of this algorithm among those used for cluster analysis has been shown. The algorithm is simple in its program implementation, which increases its usage reliability. The used technology of evolutionary modeling is rather expanded in the mentioned algorithm. Firstly, the decimal chromosomes coding is used instead of the traditional binary coding. This has resulted from the fact that the chromosome genes condition is multiple and not binary. Moreover, this is due to the absence of the genetic operator of inversion in this algorithm. Secondly, a new genetic operator used for filtering has been implemented. This operator eliminates chromosomes that do not meet the required clusters quantity condition in a task. Such chromosomes can appear in the stochastic process of their evolution. The presented algorithm has been studied in a series of simulation experiments. As a result, it has been found that stabilization of splitting into clusters is reached when the number of completed generations of evolution is 200 and more, and the population size is rather small: from 150 chromosomes (in this case no considerable amount of random-access store is required). The calculations carried out on real data showed for this algorithm the high quality of clustering and the acceptable computing speed of the same order with the computing speed of SOM and “k-means” algorithms.