Abstract: In this paper, a general definition of meta-learning is proposed. The selection of learning algorithms in meta-learning is investigated from the point of bias/variance decomposition as well as the effect of correlation on its accuracy. In order to obtain classifiers with variable correlation, artificial datasets are generated based on the simulating algorithm presented in the paper. Experiments are performed on UCI datasets and simulated datasets and show that meta-learning outperforms several combining methods averagely; and that negative correlation measured by Q statistic benefits meta-learning approach.