Model-based Individualization of Human-like Steering Controllers

Abstract

Current autonomous steering controllers are designed based on either optimal control strategies or on average human driver steering behavior. Individual drivers may wish to steer differently than those controllers do. Such conflicts can be mitigated by customizing individual steering behavior. This thesis aims to develop a method to individualize human-like steering controllers, using a parameter estimation technique for a human-like steering model based on a global optimization algorithm. With this technique, the accuracy of estimated parameters is investigated in order to help to understand the nature of driver steering behavior. A linear human-like cybernetic steering model inspired by human physical steering control actions is used in this thesis. In this steering model, seven parameters relate vehicles states to steering activity, of which the parameter values can be interpreted in a physically meaningful way. Preliminary research has demonstrated that this model structure enables parameter estimation of model parameters. But the current implemented parameter estimation technique of the steering model still needs requirements on initial guesses of the parameters and the accuracy of estimated parameters is under-investigated. In this study, parameter estimation of the steering model is investigated by formulating an optimization problem to minimize the prediction error. To deal with the nonlinearity of the optimization problem, a genetic algorithm is complemented with the Levenberg-Marquardt algorithm. This thesis presents the parameter estimation process in two steps. In the first step, the steering model generates a data set by simulating it with known parameters. This data set is used to estimate the steering model parameters and the results are compared with the known parameter set. Accurate parameter estimation results could not be reached in this first step, which is shown to be caused by over-parameterization of the steering model. Two parameters in the neuromuscular system of the steering model are then simplified to mitigate the over-parameterization problem and the other five parameters are remained. These five parameters can be estimated accurately by examining the metrics of VAF and Euclidean distance after the model simplification. In the second step, a simulation data set from a different validated human-vehicle model is used for parameter estimation to demonstrate the algorithm. It is shown that with the simplified steering model, the proposed parameter estimation technique can estimate consistent parameter values with good VAFs. A cross-validation test is also implemented to give the corroborative evidence to indicate over-parameterization of the steering model.