ABSTRACTSeveral studies that attempted to predict item difficulty using linear weighted composites have been done at ETS and elsewhere. The variables in the composites were coded item characteristics. The validity of these predictions was evaluated using multiple correlations between predicted and actual item difficulties. Results of these studies indicate that the validities of the linear predictions were not zero, but neither were they high. Also, success varied with the particular prediction task attempted, with validities varying from .45 to .77. The figures were promising, but more precision would be helpful for item design and test specification. The present study sought to improve on the validity of the linear predictions by using a “neural net,” a technique that is widely used for nonlinear prediction and is borrowed from artificial intelligence applications. Predictions made using the neural net might be more accurate than those made using linear prediction for two reasons: (a) the latter is a special case of neural net prediction, one in which the variables do not interact, and (b) it is thought that item characteristics do interact to produce item difficulty. Unfortunately, it was found that using the neural net did not lead to substantial improvement in prediction when using the same characteristics as arguments that had proven superior for linear prediction. Perhaps the performance of the neural net did not compare more favorably with that of linear prediction because the variables used had already been chosen to be optimal for linear prediction. For this reason, in an extension of the original project, variables that were optimal for the neural net were sought using a search procedure known as “genetic algorithm.” As used in this study, the genetic algorithm is a constrained search procedure intended to find the best of a fixed number of predictors, that is, the same number used in linear prediction. Applying the genetic algorithm did, indeed, lead to improved prediction. And as expected, most of the variables that were effective in linear prediction did not survive the search. However, it proved to be useful to use the variables that were effective in linear prediction as the starting set for the genetic algorithm, rather than using a randomly selected set. Only two variables were found that were useful for prediction in both the linear system and a neural net; several other variables were found that might be useful in a neural net application but were not effective in a linear application. Substantial capitalization on chance was noted in this study, which can weaken the substantive inferences. Finally, results of this study indicate that the performance of the genetic algorithm could be improved by including chromosomes that contain variables found to be effective in a linear prediction system.