Bat Algorithm With Principle Component Analysis

Large-scale problem is one hot topic in evolutionary computation. Due to the amount local optima, the population diversity is one important issue when solving such problems. In this article, two new variants with principal component analysis(PCA_BA and PCA_LBA) are designed to solve it. The main contributions are listed as follows:In the standard version of bat algorithm, there are many correlate individuals with overlapped information. To obtain the large population diversity, principal component analysis method(PCA) is employed to produce some incorrelate individuals. Furthermore, some original individuals are randomly selected into this new population to preserve the fixed population size. To test the performance, CEC'2008 large-scale benchmark functions are used and compared with three variants of bat algorithm, simulation results show the improved algorithm is effective;In PCA_BA and PCA_LBA, all individuals are analyzed by PCA, however,there are little influences for those individuals with small correlation coefficients.With this manner, a correlation threshold is employed to make a classification for population: high correlate individuals and low correlate individuals, and only high correlate individuals are considered by PCA. To provide a proper threshold,golden section method is employed to test. Simulation results indicate the validity of this modification;For evolutionary algorithms, population diversity is the main problem in the first period aiming to escaping from local optima, while the local search is the main consideration in the latter period to provide a precise local optimum.However, PCA_BA and PCA_LBA always focus the population diversity during all generations, to overcome this shortcoming, a generation threshold is introduced to provide a classification for the first and later period, and this new variant is significantly improved the performance.