Abstract: After analyzing the relationship among the generalization error of the neural networks ensemble, the generalization error, and the diversity of the individual neural network, a individual neural networks active learning algorithm ALA is proposed, which encourages individual neural network to learn from the expected goal and others, so all individual neural networks are trained simultaneously and interactively. In the stage of combining individual neural networks, a selective approach is proposed, which adds a bias to every individual network and selects part of individual networks to constitute an ensemble. Experiment results show that an efficient neural networks ensemble system can be constructed by using the individual networks learning algorithm ALA and selective ensemble approach.