Selective baggiing based incremental learning

In this article, we introduce selective bagging based incremental learning, an algorithm for incremental learning using selective ensemble. Selective bagging gains new information from the incremental data by selecting proper components. In the first situation of the incremental learning process, we train component predictors by bootstrap sampling on the original data set, and then constitute the ensemble predictor by selecting proper component predictors based on a genetic algorithm. In the next situation, we re-select proper component predictors from the original component predictors on the incremental data set; or more new component predictors are trained on the incremental data set, and a new ensemble predictor is constituted by selecting some proper predictors from all component predictors on all validation data. The proposed algorithm enables the resulting ensemble predictor to learn new information from new data set without forgetting previously acquired knowledge. Experiments on handwritten digit recognition indicate that selective bagging based incremental learning is a promising learning algorithm.