Bearing Capacity Modeling of Composite Pile Foundation Using Parameter-Optimized RBF Neural Networks

Abstract

Radial basis function-artificial neural networks (RBF-ANNs) are used for bearing capacity modeling of composite foundation reinforced with deep mixing piles. Although RBF-ANNs possess significant advantages in terms of strong generalization, flexible adaptability to multi-independent variables and sufficient avoidance of local minima, their performance may be directly affected by two uncertain parameters, the width of radial basis kernel function (spread) and the goal error of training (err_goal). Up to now still no mature methods to determine the optimal parameter values. As an exploration, a novel method is proposed to determine the optimal parameter values by thoroughly searching over the possible interval of uncertain parameters. Moreover, a technique of reconstructing more samples from few original samples is put forward to improve the prediction precision of the RBF-ANNs. The proposed techniques are applied to the bearing capacity modeling of composite foundation reinforced with deep mixing piles. The results demonstrate that the uncertain parameter optimization and sample reconstruction techniques are capable of significantly improving the performance of RBF-ANNs.