Feature Selection Study Based On Gene-Lock And Link Agent Genetic Algorithm

Feature selection is the important data preprocessing method of pattern recognition and data mining. With the research objects more complex, the feature dimensions of the objects are increasingly high. Because a large number of high-dimensional data objects contain many redundant features and even noise characteristics, the use of feature selection methods to reduce the redundant or noise characteristics and search time as well as improve the results'quality has become a hotspot. Usually reducing search time and improving search results'quality is contradictory. According to the different requirements of the real-time and quality in the work, the genetic algorithm will be used for the search of feature selection. Paper studies the relevant algorithm. The main findings are as follows:First, when the good quality feather subset is selected from high-dimensional feature sets of the strong demand for real-time, the implementation time of the conventional feather selection approaches is long and the efficient of that is very low. So this paper proposed a feather selection algorithm based on gene-lock genetic algorithm (GLGA).The gene-lock operator of GLGA is to avoid duplication search of the gene; the adaptive stopping criteria based on the feather symbol register is to accelerate the convergence speed of the algorithm.Second, because the search results'quality of network environment (lattice) genetic algorithm doesn't meet the requirement of project well, this paper proposed a feather selection algorithm based on link agent genetic algorithm (LAGA). This algorithm includes link agent structure, neighborhood competition, adaptive crossover, adaptive mutation, replacement strategy, and adaptive stopping criteria. It can keep the diversity of the agents well, and effectively inherit the good genes of good individuals as well as search new space.The paper selects two experimental data sets from the international Machine Learning data sets UCI. The GLGA and LAGA are compared with the other three genetic algorithms respectively, and the paper uses three different evaluation criteria for each of the four feature selection algorithms to feather selection and BP neural network classification experiments. Study results show that the implementation time of GLGA algorithm is very short and the efficiency is high, so it can be used for real-time feather selection. In addition, the classification rate of feather subset selected by GLGA corresponds to the others elected by the other three feature selection algorithms and sometimes even higher more. For high-dimensional feature sets, GLGA algorithm also has significant drop-dimensional effect and lower dimensional complexity of network classifier. For LAGA, Study results show that the quality of feather subset selected by LAGA is better than the others selected by the other three feature selection algorithms. The feature subset selected by LAGA has better stability, higher classification rate and lower dimensional complexity of network classifier.