Research On IRT Parameter Estimation Of ANN Method

Computerized Adaptive Testing(CAT)is a kind of testing form which combines modern measurement theory with computer technology,and its basic idea is "measured by people" and "tailored".CAT has the advantages of efficient saving,flexible measurement and high security.Based on the Item Response Theory(IRT),CAT adaptively selects questions suitable for the ability level of the participant(examinee)in the examination to be assigned to the person to answer,and when the difficulty of the questions matches the level of the person's ability,the efficiency of the testing will be greatly improved.How to estimate the item parameters and the ability parameters efficiently,quickly and accurately is the difficulty and key in the research of CAT.In estimating the item parameters and capability parameters,the traditional mathematical statistics method often needs a large sample quantity to achieve a certain degree of accuracy,and the calculation time is relatively long.Aiming at the item parameters of the questions and the ability parameter estimation of the persons,this paper studies and proposes a parameter estimation method based on Artificial Neural Network(ANN),in order to estimate the parameters of each parameter under unknown conditions.In this paper,the Three-parameter Logistic Model with binary scoring is used as the research object,using General Regression Neural Network(GRNN)is carried out as Artificial Neural Network model.The actual testing data and simulated data based on Monte Carlo method are used to verify,and the accuracy of capability parameters and item parameters estimation of neural networks under different sample quantities is studied by comparing mathematical statistics methods.The research work mainly includes the following contents:(1)The mathematical statistics method under the large sample estimates the item parameters and capability parameters of the testing response matrix at the same time,and the estimation results are labeled.(2)Combining label data,GRNN is trained to determine the structure of neural network for estimating item parameters and the structure of neural network for estimating capability parameters.(3)The item parameters and the ability parameter estimation results of the GRNN method and the mathematical statistics method are compared under different sample sizes.Finally,the GRNN is trained by simulating the test data,the network structure used to estimate the parameters,and it is applied to the actual testing through experiments.The results of comparative experiments show that,the accuracy of estimating ability parameters and item discrimination,difficulty and guessing parameters of GRNN method is better than that of mathematical statistics method in the case of small sample size.