In the educational measurement, the item response theory (IRT) has overcome many shortcomings the ways in which educational and psychological tests are usually constructed, evaluated and used. Many research results on item selection are based on the item response models, and obtained a good quality of testing. However, the traditional item selection methods are not flexible enough to satisfy the target information of the designed test. In the paper, we propose an AI technique-neural network to select test items such that the difference of information between the constructed test and the desired test is greatly reduced. The simulation results are similar to the related works proposed in the past ten years. In addition the time complexity of the proposed neural network method is the same as the traditional methods. The proposed method significantly reduces the errors of the information between the constructed test and the desired test while maintaining the efficiency of the item selection process, and then gives a new research direction on the test development.