Modelling complex variation in driver behavior data.

Automobile engineering has traditionally focused on understanding and optimizing the performance of the various systems under the hood.  Attention is now turning to the system behind the wheel, the driver.  Car companies are in need of accurate models for how drivers "work", i.e. how they manipulate the brake, accelerator, transmission and steering in response to the driving environment.  These models are particularly important in the development of electric vehicles, where particular driver behavior could strain the batteries and shorten the (already limited) range of the vehicle.  The goal of this project is to use detailed time series data on driver controls and engine performance and produce a state space model that captures the unique features of each driver's behavior.

Data for this project comprise highly multivariate time series on 18 drivers, with measurements 10 times a second on: location, speed, gas and brake pedal position, engine load, etc.  The project will require the use of unsuperivised learning on this high-dimensional, temporally correlated data to identify possible subgroups in driver behaivor. Parameters that are important for determining individual driver behavior will be identified.  The ideal student for this project will have experience with (one or more of) time series, state space, & phase plane modelling. The end result of this project will advance modelling of driver behavior which will help produce better estimates of electric car drive times.

Supervisors: Tom Nichols, Paul Jennings (WMG)