As artificial intelligence becomes more integrated into our daily lives, it’s essential that these systems are able to accurately make decisions in the real world, and react appropriately in complex, ever-changing environments.
That’s the goal of research by Kaustubh Sridhar, a doctoral student in the Penn Research in Embedded Computing and Integrated Systems Engineering (PRECISE) Center in the School of Engineering and Applied Science. Sridhar’s research on AI decision-making aims to make current models of the world more accurate, to create AI systems that can better learn from expert data, and to create systems that can adapt to new situations quickly.
“I would like to continue doing AI and decision-making research, possibly in industry, after I graduate so that I can help improve autonomous agents in the real world,” says Sridhar. “And of course, my hope is that such agents will help people in a wide range of tasks ranging from household chores to autonomous driving.”
“In undergrad, I spent my efforts inventing control algorithms for quadrotors, robots with four rotors, which I really liked,” says Sridhar. “I later expanded my boundaries to discover that generally intelligent decision making is the most challenging problem for obtaining robots that can help humans in the real world.”
Throughout his doctoral studies, Sridhar has been working to build general algorithms that are both intelligent and safe for any robot and any decision-making agent. He is a two-time recipient of Outstanding Reviewer awards in top machine learning conferences. His work was nominated for the Best Paper award at the International Conference on Cyber-Physical Systems (ICCPS) in 2023. His most recent research has focused on semi-parametric methods for machine learning and decision-making applications that can be used for robot learning and learning-enabled digital or physical systems, such as autonomous vehicles, electric grids, and cloud computing.
“The key takeaways include that simple ways to combine the benefits of non-parametric components with neural networks leads to better generalization in learning both dynamics models and policies,” he says. “Some conclusions that surprised me were that semiparametric methods can provide rigorous guarantees for real-world performance in diverse domains that otherwise cannot be found.”
Read more at Penn Engineering Today.
