Modeling Lane-Changing Behavior Based on a Joint Neural Network

Abstract

This paper proposes a joint neural network model to imitate lane-changing behaviors. Specifically, lane-changing decision-making process is captured by probabilistic neural network (PNN) and lane-changing decision-making process is learned by back-propagation neural network (BPNN). The link between the two neural networks is the target gap for lane-changing. After testing and calibrating the joint neural network model, simulation experiments are designed to study heterogeneous traffic flow at an off-ramp bottleneck. Numerical simulations are conducted in various traffic scenarios with different market penetration rates (MPRs) of intelligent vehicles (IVs) and proportions of exit vehicles. Finally, the performance of heterogeneous flows is evaluated from the perspectives of average speed, road capacity, and safety. The results show that joint neural network can accurately predict the gap types chosen for lane changes and vehicle trajectory during lane-changing. For the traffic system, road capacity obtains the least value when the MPR of IVs is 50%. Moreover, frequent lane-changing movements upstream the off-ramp bottleneck determine the areas at greatest risk. However, when MPR of IVs is over 80% or proportion of exit vehicles is below 15%, both traffic efficiency and safety can be significantly improved. This work provides some insights into the application of machine learning algorithms to traffic flow modeling, and conducts quantitative analysis on the impact of key parameters on traffic systems. Findings of this work can support management and operation of automated highway systems in the future.