Predicting the effectiveness of rolling dynamic compaction using genetic programming

Rolling dynamic compaction (RDC) is a soil compaction method that involves impacting the ground with a non-circular roller. This technique is currently in widespread use internationally and has proven to be suitable for many compaction applications, with improved capabilities over traditional compaction equipment. However, there is still a lack of knowledge about a priori estimation of the effectiveness of RDC on different soil profiles. To this end, the aim of this paper is to develop a reliable predictive tool based on a machine-learning approach: linear genetic programming (LGP). The models are developed from a database of cone penetration test (CPT)-based case histories. It is shown that the developed LGP-based correlations yield accurate predictions for unseen data and, in addition, that the results of a parametric study demonstrate its generalisation capabilities. Furthermore, the selected optimal LGP-based model is found to yield superior performance when compared with an artificial neural network model recently developed by the authors. It is concluded that the LGP-based model developed in this study is capable of providing reliable predictions of the effectiveness of RDC under various ground conditions.