A symbolic genetic programming approach for identifying models of learning-by-doing

In this study, we apply a symbolic regression approach to generate and investigate new potential univariate learning curve functional forms to forecast human learning responses efficiently and stably. Past studies have compared learning models in the literature to one another. Yet, continued interest in model development and comparison suggests that the question remains open as to whether there are other useful and yet-undiscovered models. We address the question of whether the existing literature contains the best model choices, or if additional forms have merit. We employ a multigenic genetic programming algorithm to secondary field data from a range of manual sewing tasks. We identified an array of potentially useful empirical forms and examined whether these forms match or improve upon extant forms. Among two-parameter functional forms, the log linear form performed well in efficiency and stability for both models of cumulative experience, and cumulative working time. A three-parameter hyperbolic model was found and top-ranked as a model of cumulative work and a model of cumulative time in the three-parameter learning curve functional forms. We also found that 4-parameter models show characteristics of over-fitting and have small marginal differences in efficiency and stability for models of cumulative working time, which suggests that a three-parameter model may be a good choice, in general.