THE RELATION BETWEEN FIT AND PREDICTION FOR ALTERNATIVE FORMS OF LEARNING-CURVES AND RELEARNING CURVES

Learning-curve models fitted to initial data are used to predict subse quent performance; however, the model that fits the initial data best may not predict best in future periods - a paradox documented in appli cations of other prediction models. Little evidence exists about the m agnitude of the problem in the domain of learning curves and relearnin g curves. Using laboratory data, the authors examine the predictive ab ility of alternative models, examine the strength of the relation betw een goodness-of-fit and predictive ability, and test whether this rela tion is the same for both learning curves and relearning curves. Altho ugh the correlations between measures of goodness-of-fit and predictiv e ability are not high, one curve (a log-log-linear model recently int roduced to the literature) tended to dominate the rankings on the basi s of predictive ability for both learning curves and relearning curves . This curve also tended to provide the best fit in the estimation per iod as a relearning curve, and the second-best fit as a learning curve .