Genetic programming for predictions of effectiveness of rolling dynamic compaction with dynamic cone penetrometer test results

Abstract

Rolling dynamic compaction (RDC), which employs non-circular module towed behind a tractor, is an innovative soil compaction method that has proven to be successful in many ground improvement applications. RDC involves repeatedly delivering high-energy impact blows onto the ground surface, which improves soil density and thus soil strength and stiffness. However, there exists a lack of methods to predict the effectiveness of RDC in different ground conditions, which has become a major obstacle to its adoption. For this, in this context, a prediction model is developed based on linear genetic programming (LGP), which is one of the common approaches in application of artificial intelligence for nonlinear forecasting. The model is based on in situ density-related data in terms of dynamic cone penetrometer (DCP) results obtained from several projects that have employed the 4-sided, 8-t impact roller (BH-1300). It is shown that the model is accurate and reliable over a range of soil types. Furthermore, a series of parametric studies confirms its robustness in generalizing data. In addition, the results of the comparative study indicate that the optimal LGP model has a better predictive performance than the existing artificial neural network (ANN) model developed earlier by the authors.