The Descended Dimension Method of Parameter-estimation of BP Neural Network Based on Item Response Theory

摘要

Parameters and individual ability in Dichotomously Scored of Response Theory model are estimated with Back-Propagation Neural Network. The dimension of Scoring matrixes X is descended by using scoring rate or passing rate or coefficient of correlation or guess rate when estimating those item parameters. The method is simulated in computer, and the results show that the item parameters estimation is more precise than the current international popular software, such as BILOG,PARSCALE etc. The well-trained Neural Network can output the estimate value in field test and need fewer examinees and items. The difference between estimate values and true values is very small.