Deep Reinforcement Learning and Game Theoretic Monte Carlo Decision Process for Safe and Efficient Lane Change Maneuver and Speed Management

Author

Shahab KarimiFollow

Date of Award

5-2023

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Mechanical Engineering

Committee Chair/Advisor

Prof. Ardalan Vahidi

Committee Member

Prof. John Wagner

Committee Member

Prof. Javad Velni

Committee Member

Dr. Phanindra Tallapragada

Abstract

Predicting the states of the surrounding traffic is one of the major problems in automated driving. Maneuvers such as lane change, merge, and exit management could pose challenges in the absence of intervehicular communication and can benefit from driver behavior prediction. Predicting the motion of surrounding vehicles and trajectory planning need to be computationally efficient for real-time implementation. This dissertation presents a decision process model for real-time automated lane change and speed management in highway and urban traffic. In lane change and merge maneuvers, it is important to know how neighboring vehicles will act in the imminent future. Human driver models, probabilistic approaches, rule-base techniques, and machine learning approach have addressed this problem only partially as they do not focus on the behavioral features of the vehicles. The main goal of this research is to develop a fast algorithm that predicts the future states of the neighboring vehicles, runs a fast decision process, and learns the regretfulness and rewardfulness of the executed decisions. The presented algorithm is developed based on level-K game theory to model and predict the interaction between the vehicles. Using deep reinforcement learning, this algorithm encodes and memorizes the past experiences that are recurrently used to reduce the computations and speed up motion planning. Also, we use Monte Carlo Tree Search (MCTS) as an effective tool that is employed nowadays for fast planning in complex and dynamic game environments. This development leverages the computation power efficiently and showcases promising outcomes for maneuver planning and predicting the environment’s dynamics. In the absence of traffic connectivity that may be due to either passenger’s choice of privacy or the vehicle’s lack of technology, this development can be extended and employed in automated vehicles for real-world and practical applications.

Recommended Citation

Karimi, Shahab, "Deep Reinforcement Learning and Game Theoretic Monte Carlo Decision Process for Safe and Efficient Lane Change Maneuver and Speed Management" (2023). All Dissertations. 3310.
https://tigerprints.clemson.edu/all_dissertations/3310

Author ORCID Identifier

0000-0003-4369-6151