Ensemble learning approach based on stacking for unmanned surface vehicle's dynamics

Abstract

The 3 degrees of freedom (DOF) model, including surge motion, sway motion and yaw motion, with differential thruster is proposed to describe the unmanned surface vehicle (USV)'s dynamics. The experiment is carried out in the Qing-huai river and the data obtained from different zigzag trajectories is filtered by gaussian filtering method. The base learners, Backpropagation (BPNN), Support vector machine (SVM) with RBF kernel and SVM with Linear kernel, are selected to identify the dynamic model of USV. To guarantee that base learners have a strong robustness and generalization ability, the initial weights in BPNN and hyper-parameters in SVM are optimized by cross validation (CV), genetic algorithm (GA), particle swarm optimization (PSO) and cuckoo search algorithm (CS) method. The results show that the CS method has a better optimal capacity in predicting the USV's dynamics. These methods on prediction USV's dynamics have their own advantages and disadvantages. Furthermore, The Ensemble learning (EL) method, called stacking, integrating these base learners is firstly proposed to identify the USV's dynamics. The results demonstrate that the EL method has more accuracy in identifying the dynamic models than base-learners.