A major roadblock in today’s approach to machine learning is sample complexity, which is the question of how much data is needed for learning algorithms to achieve the right level of performance. The more data, the more energy required, and the greater the impact on the environment. To address this, Nandan Tumu, an Electrical and Systems Engineering (ESE) doctoral student in the School of Engineering and Applied Science, explores more efficient methods and discovered that physics-informed and constrained learning could significantly reduce the need for extensive sampling.
By integrating this approach with conformal prediction, a method for distribution-free uncertainty quantification, Tumu has found a way to control complex systems efficiently and reliably. An innovative pairing of physics-informed and constrained learning and conformal prediction has become the driving force of his research, promising to unlock the potential of larger multi-agent systems, such as fleets of drones or driverless cars, or infrastructure like power grids and wind farms.
Optimizing transportation systems has been a motivating application for Tumu’s research. In 2023, he joined a team at Pacific Northwest National Labs (PNNL) as a summer intern to develop machine learning methods for traffic system control. In their recent paper, “Differentiable Predictive Control for Large-Scale Urban Road Networks,” Tumu and his collaborators address one of the most pressing issues of our time: traffic congestion and its contribution to CO2 emissions. Since transportation is a major driver of global emissions, optimizing traffic networks is essential for reducing energy consumption and mitigating climate change.
Tumu’s novel approach leverages Differentiable Predictive Control (DPC), a physics-informed machine learning methodology developed at PNNL, to advance traffic management.
The practical implications of Tumu’s research will be evaluated through PNNL’s collaboration with the city of Coral Gables, Florida, as a part of the AutonomIA project funded by ARPA-E. The goal is to implement these advanced traffic control algorithms—strategies for managing traffic lights and signals—in a real-world setting, in order to significantly reduce travel time and energy consumption.
Tumu’s research extends beyond urban traffic. In collaboration with PNNL, he is applying and advancing DPC methodologies to enhance the efficiency of existing wind farms. “This extension aligns with my overarching research vision of developing control algorithms for networked cyber-physical systems to enhance efficiency and performance,” Tumu says. “By incorporating physics-based information and uncertainty quantification, I aim to create improved control algorithms that leverage real-world data.”
This story is by Liz Wai-Ping Ng. Read more at Penn Engineering.
