Research On Applying ANN To Estimate IRT Parameters With Small Sample

Item response theory (IRT) is the basis of constructing item bank. Before doing so, item parameters must be precisely estimated. On one hand, traditional parameter estimation method is based on mathematical statistics (MS), which required a large number of examinees. On the other hand, the number of examinees participating into the pretest must be limited for the purpose of item security. The two make the research of item parameter estimation a very important task of IRT, especially in the pretest with small number of examinees.In this thesis, a new item parameter estimation method based on artificial neural networks (ANN) was proposed. In order to verify its performance, the three-parameter Logistic model scoring dichotomously and general regression neural networks were selected as the researched IRT and ANN model, respectively. And then, simulation experiment was designed and conducted according to Monte Carlo method. Conclusions are drawn as following:1. An ANN modeling method, which takes the item response pattern and its IRT parameter as ANN's input and output respectively, was proposed here. Theoretical analysis indicates that this method is somewhat advantageous.2. According to three kinds of performance criterions, the ANN method and MS method were compared under several experimental conditions. The results show that: the ANN method is somewhat more advantageous than the MS method on most criterions under all conditions. Particularly, if the MS method cancels the item prior constrains, its performance deteriorates seriously, which makes the ANN method perform quite well plausibly.3. Analysis of variance on the ANN method's performance indicates that: the ANN method can't perform well on item's all parameters at the same time; researchers should make a tradeoff among the parameters, such as caring more on difficulty than the others, and then, special pretest measures should be taken.