Developed ZARTS, a novel zero-order optimization-based neural architecture search (NAS) approach, addressing issues with the first/second-order approximations used in DARTS with rigorous theoretical proofs.
Conducted a survey on ML methods for the SAT problem(accepeted by Mach. Intell. Res.)
Participated in SAT Competition 2022, developing two Kissat-based solvers with multi-armed bandit and local search techniques.
Georgia Institute of Technology, Machine Learning Group
July 2020 - Dec. 2020
Developed and implemented an automated program synthesis pipeline using CFG-GP to design and evaluate CDCL SAT solver heuristics, achieving results on par with manually designed state-of-the-art heuristics.