Variance-tuned ensemble Gaussian process regression for learning and prediction of ship maneuvering motion

Abstract

With the development of maritime transportation, ship maneuverability is attracting increasing attention, the learning and modeling of ship maneuvering motion is the basis of studying the ship maneuverability. Machine learning methods such as Gaussian process regression (GPR) are widely employed in the learning and prediction of ship dynamics. However, the predictions by the learning model always exhibit distinct deviations when the ship motion to be predicted are quite different from those of the training dataset. To deal with this problem, a robust method based on variance-tuned ensemble GPR learning method is proposed. Multiple base predictors are trained based on GPR each from a different bootstrap dataset. The weight values of each base predictor are adaptive with variance-tuned strategy during the whole prediction process. Taking the ONR Tumblehome ship and KRISO Very Large Crude Oil Carrier 2 tanker as study objects, the results indicate that the proposed method exhibits strong generalization capabilities and achieves high accuracy in predictions.