| This thesis considers the problem of adaptive estimation and adaptive control through the use of multiple models. The concept differs from conventional parameter adaption in that multiple models can instantaneously change the behavior of the estimator or controller. However, this dissertation is not concerned with instantaneous switches among models, but instead with a probability-based weighting of the outputs of different models. This type of soft-switching still allows for rapid variation in the emphasis among the models, but instead blends the outputs of the various models, rather than isolating just one output.; The architecture developed in this thesis applies multiple models to both the problem of multiple target tracking and the problem of adaptive control of a single plant, where the plant itself does not change over time, but the control objectives for the plant do vary. Firstly, existing theory on multiple model state estimation is reviewed and then extended to include nonlinear models and nonlinear observers. Secondly, linear control theory involving multiple models is examined with existing and new stability results being presented. This stability theory is adapted to a class of nonlinear systems, which is affine in the control input. Finally, a method for combining the quantities of a multiple model target tracker and a multiple model controller is introduced.; The proposed application for the theory presented here is the Adaptive Cruise Control (ACC) problem in the automotive industry. Through longitudinal control over a vehicle, an ACC system can achieve two principal objectives. The first is to improve traffic flow capacity. The second is to reduce driver workload while operating the vehicle. It is this second objective that is emphasized in this thesis, with multiple controllers stressing different levels of road safety and ride-comfort. The tools of this dissertation are applied to the ACC problem both in simulation and experimentally. By employing different models in the multiple target tracker, various maneuvers of preceding vehicles can be detected and identified. Subsequently, as maneuvers are detected, smooth switching among the controllers allows the control system to tradeoff safety and comfort in a manner reminiscent of realistic human driver behavior. |
