System-Level Optimization of Longitudinal Acceleration of Autonomous Vehicles in Mixed Traffic

Mixed traffic scenarios present challenges to autonomous vehicles due to the high degree of randomness introduced by human drivers combined with their larger reaction times and perception errors. In this paper we address those challenges on a longitudinal control level by designing optimal car-following models which aim to maximise simultaneously the vehicle population's speed, efficiency, comfort, and safety. We use the agent-based simulation mixed traffic tool BEHAVE to design a scenario covering all driving phases and formalize the four different objective functions to be optimized. We take on a multi-objective optimization approach in order to analyse the trade-offs that occur between the chosen traffic metrics. Furthermore, we design a methodology to scalarize the multi-objective problem and find a single optimal well-balanced parameter set maximizing the formulated objective functions. The optimized model is able to gain significant performance increase in terms of efficiency, comfort and safety, while giving away a significantly smaller percentage of average speed.