Vindula Jayawardana is a Ph.D. student in Electrical Engineering and Computer Science at MIT, working with Prof. Cathy Wu. He received his B.S. in Computer Science and Engineering from the University of Moratuwa in Sri Lanka in 2018 and an M.S. in Electrical Engineering and Computer Science from MIT in 2022. His research interests are in reinforcement learning and its applications in roadway automations. In particular, he focuses on learning generalizable control methods for energy-efficient autonomous driving and its prospective impact on climate change mitigation goals. Outside of work, he likes reading, traveling, and listening to music.
With the development of connected vehicle technologies, there are newfound opportunities for impactful roadway automations, such as efficient autonomous vehicle control for traffic smoothing and emission reduction. Given such problems often manifest as multi-agent dynamical systems with nonconvex objectives and nonlinear dynamics, learning-based methods are often leveraged to solve them. Yet, the implicit necessity for learning-based methods to effectively generalize across traffic scenarios remains a critical challenge. This work first aims to identify the pitfalls of not considering problem variations in learning controllers and show that it can lead to evaluation overfitting. Second, synergizing conventional control and learning-based control is proposed to learn controllers that generalize. Last, learned robust controllers are used to mitigate metropolitan roadway carbon emissions with autonomous vehicles as a means of system control, showing the prospective impact of roadway automations in achieving climate change mitigation goals. Overall, this research aims to learn controllers that generalize across traffic scenarios, with the goal of realizing large-scale roadway automations.