Data Representation for Predicting Performance with Learning Curves

Abstract

Mathematical learning curve models can be used to predict the time or cost required to perform future cycles in a repetitive construction activity. The analyst has a choice of several methods of representing the data, usually trading off between response and stability of forecasting information. Traditionally, learning curve data has been evaluated using either unit data or cumulative-average data. This paper evaluates those two methods and two other techniques: the moving average and the exponentially weighted average. For the 54 construction activities evaluated, unit data gives the most accurate prediction of the time or cost to complete the remaining cycles of the activity. Cumulative-average data gives the least accurate prediction. Compared to unit data, the exponentially weighted average can predict future performance with only a slight loss of accuracy early in the activity, but equal accuracy later in the activity. The exponentially weighted average may offer an improved combination of stability and response, depending on the smoothing parameter chosen.