The move ensemble method

According to the analysis of the position relationship between expectation value (EV) and prediction values (PV) of component models in an ensemble model, a Move Ensemble method (MEM) based on a new sampling strategy is proposed in this paper. The MEM is to create up-move training data set by increasing EVs and to create down-move training data set by decreasing EVs. Then the up-move component model and the down-move component model are built on the up-move training data set and the down-move training data set, respectively. In terms of the peculiarity of this pair component model, a nonlinear combining method, which with neural network classification principle to choose the fittest weight vector for the two component models, is used. Two simulated experiments proved that the proposed move ensemble model outperforms the single model.