Abstract: Dynamic adaptive selection ensemble based on local classification accuracy estimation is introduced to improve the performance of neural network ensemble.Based on Bayesian theory,it can be proved that the perfor-mance of the dynamic adaptive selection ensemble can approximate the optimal Bayesian classifier if certain hypotheses are met.According to this conclusion,member network selection methods based on hard decision and soft decision are introduced.Experiment is made on five data sets selected from the UCI machine learning database.The experimental results show that the dynamic adaptive selection ensemble is better than conventional voting and averaging methods,and the performance is not sensitive to the size of neighborhood.Furthermore,the soft decision method is of better performance than the hard decision method.