This paper describes a method for improving the generalization performance of bagging ensemble by means of using Bayesian approach. We examine the Bayesian prediction using bagging leaning machines for regression problems, and show a method to reduce the generalization loss defined by the square error of the prediction for test data. We examine and validate the effectiveness via numerical experiments using the CAN2s as learning machines, where the CAN2 is a neural net for learning efficient piecewise linear approximation of nonlinear functions.