Date of Award

August 2023

Degree Type

Thesis

Degree Name

Master of Science

Department

Computer Science

First Advisor

Tian Zhao

Committee Members

Christine Calynn T Cheng, Xiao Qin

Keywords

Car Following, Reinforcement Learning, Simulation of Urban MObility (SUMO), Soft Actor Critic (SAC), Traffic Simulation, Velocity Control

Abstract

Agents in multi-agent traffic simulation tend to be more dependent on the rules and existing instructions to move mechanically and unnaturally imitating human behaviors. The agents will not accelerate or decelerate as humans do. Humans have an irregular pattern of acceleration and deceleration when it comes to real-time driving. This includes hitting breaks when not necessary and sometimes even driving above the speed limit to catch up. In prior works, other factors such as drag and simulation-specific parameters were not considered in the models. Additionally, the models were not tested on the traffic simulation frameworks like SUMO. Instead, they utilized simple numerical models to simulate the environment and evaluate the performance of the models. Therefore, there is a need to further investigate and incorporate these additional factors, as well as validate the models on the SUMO platform, to enhance the realism and applicability of the research. It is also difficult to calibrate SUMO to a given traffic scenario as traffic engineers might need to specify manually the vehicle specifications while designing the experiments. It would be easier for engineers to populate the road network with pre-trained agents that require minimal tuning which includes specifying maximum acceleration, deceleration, and minimum and maximum speed of the vehicles to be simulated. We propose a unified system for agents to decide when to accelerate and decelerate with the help of deep reinforcement learning aided by a combination of factors such as instantaneous speed, time, and other important metrics. The proposed system will aid the agents to behave more like humans by acting based on the surrounding agents in complex situations. This in turn can help create a diverse traffic flow that can mimic real-life traffic scenarios.

Download

Share

COinS