Particle swarm hybridize with Gaussian Process Regression for displacement prediction

Abstract

Gaussian Process Regression (GPR) as a new kernel machine learning technique holds many advantages such as programming easily, self-adaptive acquisition of hyper-parameters and prediction with probability interpretation. Presently, the hyper-parameters of GPR are got by maximizing likelihood function of training samples based on conjugate gradient algorithm. However, the algorithm has the shortcomings of too strong dependence on initial value in optimization effect, difficultly in determination of iteration steps and easily falling into local optimum. The author proposes particle swarm optimization (PSO) and genetic algorithm (GA) is respectively used to search the optimal hyper-parameters during the training process automatically then formed the PSO/GA-GPR algorithm. Finally, the two different hybrids algorithm are adopted to predict the displacement through the typical landslide cases analysis in order to verify the extrapolation ability of both approaches. From the deformation prediction results of landslide displacement, it can be concluded that the PSO-GPR coupling model obviously improved the prediction precision than that of GA-GPR, so it can be utilized in displacement prediction of geotechnical engineering and meanwhile be served as a reference for similar projects.