Simulated CAVs Driving and Characteristics of the Mixed Traffic Using Reinforcement Learning Method

Using cooperative gaming method can be a promising approach to mimic various driving tasks in the field of automated driving. This paper presents a Deep Reinforcement Learning approach for modelling Connected and Automated Vehicles (CAVs) in heterogeneous traffic. First, the Gipps models were integrated into regular vehicle agent. Second, an enhanced Q-learning was employed as the modelling platform for CAVs, to strengthen the capability of the simulation system in realistically reproducing CAV lane-changing and car-following behaviour. Third, extensive simulation studies based on a two-lane highway stretch show that the inclusion of CAVs considerably improves traffic flow, mean speed, and traffic capacity. We also simulated managed lane policies to determine how CAVs should be distributed across lanes in various conditions. Such understanding is essential for research concerning CAV, as well as, the CAV implication for future traffic management.