dc.description.abstract | The purposes of this research are 1) studying the driving behavior of human drivers
and Connected and Automated Vehicles (CAVs); 2) developing appropriate
car-following models for CAVs to form a cooperative strategy in order to enhance
traffic stability, reduce traffic oscillation and improve safety; 3) generating CAV’s
driving model through a learning based method; and 4) using learning based method
to develop a cooperative driving strategy in signalized intersections. Therefore, in
the first part of this research, we show two demonstrations about the model
development through an approach that modifying existing car-following models.
The proposed methods are applied at highway sections with on-ramp and priority
junction. By comparing with human drivers, the result shows that with a proper
controlling mechanism, an increasing percentage of autonomous vehicles will
reduce the total travel time and smooth traffic oscillations.
Developing driving models for Connected and Automated Vehicles through
modifying a classical car-following model seems acceptable. However, those
models are affected and constrained by empirical equations. The classical models,
used to be applied for simulating human driving behaviors, may not be an ideal
model for the Connected and Automated Vehicles due to the difference between
machine and human.
Fortunately, technology innovations, most notably, machine learning techniques
offer another modeling approach. In the second part of this research, we develop
car-following controllers for Connected and Automated Vehicles based on
reinforcement learning to dampen or eliminate traffic oscillations (or stop and go
driving behaviors) caused by human drivers. By taking advantage of reinforcement
learning, the controller has the capability of self-learning and self-correction.
Compared to traditional modeling approaches, it significantly reduces the modeling
constraints. Two case studies are established to evaluate the model's performance.
Our results demonstrate that the generated model from reinforcement learning is
able to improve travel efficiency as well as reduce the negative impact of traffic oscillations. | |