One of the most utilised pavement deterioration models is of a form called Family Group (also referred to as Family Curves, Survivor Curves, Family Types or Regression Trees in Data Mining). This modelling method has been used quite successfully in deterioration modelling and divides the network into separate groups believed to deteriorate in a similar manner. Each group is then represented by some deterioration function. This form of modelling has the advantage of providing an overall solution that is easily interpreted by the separate series of simple regression curves. The difficulty with modelling of this type is in defining an acceptable level of complexity ensuring the model does not suffer overfitting. The question of an acceptable level of complexity requires answering for both the division of the network into families (groups) and the form of the deterioration curve representing each family. This paper introduces Minimum Message Length Regression Trees (MMLRT) which can search and select from all possible combinations an appropriate number of families and corresponding curves. This ensures that the final models are of an appropriate level of complexity as supported by the data and will more accurately infer future deterioration levels from previously unseen data. (a) For the covering entry of this conference, please see ITRD abstract no. 0612AR242E.
Abstract