Adjustment for cognitive interference enhances the predictability of the power learning curve

Abstract

Learning curves, which express performance as a function of the cumulative number of repetitions when performing a given task, have a long tradition of supporting managerial decisions in production and operations management. Performance generally improves as the number of repetitions of a given task increases, with the latter being a primary proxy to reflect experience. A learning curve is usually a maximum-likelihood trend-line that best fits raw data points by splitting them to above and below it. However, its curvature does not always accurately capture the scatter around it, which reduces its accuracy. This paper advocates for an improved learning curve, one that accounts for the variable degree of cognitive interference that occurs while learning when moving from one repetition to the next. To capture this phenomenon, this paper accounts for memory traces of repetitions to measure the residual (interference-adjusted), not the nominal (maximum), cumulative experience. Two alternative learning curve models were developed. The first model aggregates the residual cumulative experience for each repetition while fitting the data. The second model is an approximate expression and, as a continuous model, much easier to implement. The models were tested against data from different learning environments (such as production and assembly), alongside a more traditional power (log-linear) form of the learning curve and its plateau version. The results show that the interference-adjusted models fit the data very well, such that they can serve as valuable tools in production and operations management.