Importance Weighted Gaussian Process Regression for Transferable Driver Behaviour Learning in the Lane Change Scenario

Abstract

Due to advantages of handling problems with nonlinearity and uncertainty, Gaussian process regression (GPR) has been widely used in the area of driver behaviour modelling. However, traditional GPR lacks the ability of transferring knowledge from one driver to another, which limits the generalisation ability of GPR, especially when sufficient data for driver behaviour modelling are not available. To solve this limitation, in this paper, a novel GPR model, Importance Weighted Gaussian Process Regression (IWGPR) is proposed. The importance weight (IW) represents the probabilistic density ratio between two drivers and the unconstrained least-squares importance fitting (ULSIF) is applied to calculate IW. Meanwhile, an IW-based model selection (IWMS) method is proposed to help the model select optimal parameters. Using IWGPR, sufficient historical data collected from one driver can be used to model another driver with insufficient data, and thus improve the generalisation ability of GPR. To verify the proposed algorithm, a toy regression problem is used to illustrate the working mechanism of IWGPR. With simulated and naturalistic driving data, three experiments for driver behaviour modelling in the lane change scenario, are designed and carried out. Experimental results indicate that IWGPR performs better than GPR when sufficient data are not provided by the new driver, which proves the generalisation ability of IWGPR. Meanwhile, the comparative study between different transferable driver behaviour learning methods is detailed and analysed.