Steven Joseph Szwabowski

Stochastic dynamic programming control policies for fuel efficient in-traffic driving

This paper demonstrates a methodology, based on stochastic dynamic programming, for developing a control policy that prescribes vehicle speed to minimize on average a weighted sum of fuel consumption and travel time, while travelling along the same route or a set of routes in a given geographic area. Given the current road grade, traffic speed and vehicle speed, the control policy prescribes an offset in vehicle speed relative to current traffic speed, which when added to the predicted value of traffic speed, gives a vehicle speed set point for an adaptive cruise control system. It is shown that transition probability matrices necessary to generate the control policy can be constructed from gathered data. A virtual testing environment based on CarSim is used for simulations that can effectively handle vehicle following and adaptive cruise control scenarios. Comparative fuel savings are shown to depend on time of travel (off-peak hours or rush hour) and traffic assumptions.

Modeling of vehicle driving conditions using transition probability models

This paper considers modeling of vehicle driving conditions using transition probability models (TPMs) for applications of dynamic optimization. The properties of transition probabilities for vehicle speed, vehicle acceleration, and road grade are discussed based on the analysis and experimental vehicle data. The KL-divergence is shown to provide an effective metric that can differentiate similar driving conditions from dissimilar ones.

Evolving Markov chain models of driving conditions using onboard learning

This paper describes simple and suitable for real-time implementation algorithms for on-board learning of Markov Chain models of driving conditions (e.g., driver wheel torque request, vehicle speed, surrounding traffic speed, road grade, road curvature etc.). The use of Kullback-Liebler (KL) divergence is proposed as a stopping and re-initialization criterion for learning, permitting an evolving set of Markov Chain models to be generated for different route segments. Examples based on learning models of road grade and vehicle speed are reported. Assuming that a set of learned Markov Chain models and of associated control policies is available onboard of the vehicle, the use of KL divergence is also advocated for selecting the control policy that matches the current driving conditions. Potential applications of this approach include optimal energy management in Hybrid Electric Vehicles (HEV) and fuel efficient Adaptive Cruise Control.