07/07/2021
By Liming Jiang
Ph.D. Candidate: Liming Jiang
Defense Date: Friday, July 23, 2021
Time: 9 - 11:30 a.m.
Location: Via Zoom. Please contact Liming_Jiang@student.uml.edu for the link.
Thesis/Dissertation Title:
Modelling The Cooperative and Ethical Behavior of Autonomous Vehicles in A Mixed Autonomy Environment by Artificial Intelligence
Committee Advisor:
Yuanchang Xie, Civil & Environmental Engineering, University of Massachusetts Lowell
Committee Members*:
Nathan Gartner, Civil & Environmental Engineering, University of Massachusetts Lowell
Danjue Chen, Civil & Environmental Engineering, University of Massachusetts Lowell
Polichronis Stamatiadis, Civil & Environmental Engineering, University of Massachusetts Lowell
Nicholas G. Evans, Philosophy, University of Massachusetts Lowell
Brief Abstract:
This dissertation addresses two critical aspects of future transportation systems consisting of both human-driven vehicles and vehicles of different levels of automation: cooperation and ethics. Cooperative behavior is often seen among human drivers. Similar to human drivers, autonomous vehicles (AV) can be programmed to be cooperative and achieve specific goals. Depending on the control algorithm design, such cooperative behavior may or may not contribute to better performance of traffic systems and lead to ethical consequences. Many challenges arise when developing cooperative AV, including mixed autonomy, scalability, stochasticity, nonlinearity, delayed feedback, and ethics (e.g., being responsible and fair). Such issues are explored from four perspectives in this dissertation by utilizing artificial intelligence and microscopic traffic simulation. These four relatively independent but related studies cover human driver behavior modeling, cooperative merge control, ethical decision making, and dampening stop-and-go traffic.
The first study aims to clone human driver behavior using artificial intelligence (AI), where the cooperative behavior of human drivers is implicitly learned. It adopts an innovative Mixture of Experts (MoE) method to simultaneously model vehicles’ longitudinal and lateral driving behaviors based on the naturalistic driving HighD dataset, and to address the potential heterogeneity in driver behavior. In addition, this study applies the SHapley Additive exPlanations (SHAP), a game theoretic approach, to interpret the results of the MoE models and to understand the individual impacts of variables as well as their interaction effects on drivers’ longitudinal and lateral responses. The proposed MoE is able to accurately model human driver behavior and has good potential to be integrated into microscopic traffic simulation tools. In the future, the MoE and SHAP approach can be used to model human driver behavior in a mixed autonomy environment when such data is available.
The second study focuses on how to improve the merge control of AV prior to lane reduction points due to either accidents or constructions. A Cooperative Car-following and Merging (CCM) control strategy is proposed to model cooperation among vehicles. The CCM control extends the traditional adaptive cruise control concept and considers vehicles in adjacent lanes for longitudinal control. It explicitly takes courtesy of AV into consideration, and quantifies the impacts of courtesy on traffic operations given the coexistence of AV and Human-Driven Vehicles (HDV). The results suggest that being cooperative and showing courtesy is important for improving safety and traffic operational efficiency at lane reduction points.
In the third study, Deep Reinforcement Learning (DRL) is adopted to control the longitudinal behavior of AV. The purpose is to see how cooperative AV behaving altruistically can help to dampen the stop-and-go traffic shockwaves and stabilize traffic flow. The Soft Actor Critic (SAC) reinforcement learning (RL) algorithm, which is well suited for continuous control problems, is chosen for the DRL agents. The reward function is carefully designed to include altruism and to achieve three main objectives: safety, efficiency, and oscillation dampening. SUMO simulation results show that with a properly designed reward function DRL agents can effectively mitigate the stop-and-go traffic shockwaves and improve safety.
Previous research on cooperative and autonomous vehicles (CAV) suggests they can substantially improve traffic system operations in terms of mobility and safety. However, these studies do not explicitly take each vehicle’s potential gain/loss into consideration and ignore their levels of willingness to cooperate. They do not account for ethical issues such as fairness either. In the fourth study, several cooperation/courtesy strategies are proposed to address the above issues. These strategies are grouped into two categories based on non-instrumental and instrumental principles. Non-instrumental strategies make courtesy/cooperation decisions based on some courtesy proxies and a user-specified courtesy level, while instrumental strategies are based only on courtesy proxies related to local traffic performance. Also, a new CAV behavior modeling framework is proposed based on our previous work on cooperative car-following and merging (CCM) control to evaluate the proposed instrumental and non-instrumental strategies.
Overall, this research shows that without cooperation the potential of vehicle connectivity and automation cannot be fully exploited. Smart AV acting selfishly will not generate the most desirable traffic control outcomes. It is promising to adopt AI-based approaches to design and model cooperative AV behavior. Also, adding cooperative behavior to AV needs to further consider ethical issues such as fairness and flexibility (i.e., individual willingness to cooperate), instead of system mobility performance only. It is anticipated that vehicle cooperation will become an increasingly important research topic in the future for AV, and will play a critical role in improving traffic operations and safety.