LANS Seminar

Description: Trust region methods have shown much stability for noisy non-convex problems. A subclass of these methods, called ASTRO, uses adaptive sampling to build quadratic local approximations and move to new iterates. ASTRO attains competitive worst-case complexity results via exploiting dependence of noisy observations and underlying problem’s sample-path structure. When the objective function is the output of a stochastic simulation whose inputs follow a probability model, we show the effect of common random numbers in the reduction of overall sample complexity. We also investigate the analytical and practical pros and cons of quadratic regularization for this algorithm.

Bio: Sara Shashaani is an Associate Professor in Edward P. Fitts Department of Industrial and Systems Engineering and Affiliated Faculty with the Operations Research Graduate Program at North Carolina State University. Her research interests lie in the intersection of stochastic optimization and Monte Carlo simulation with a focus on continuous zeroth-order nonconvex problems and on digital twins. She is a 2024 Goodnight Early Career Innovator and a 2025 MGB-SIAM Early Career Fellow, and has received  multiple research grants from the National Science Foundation and the Office of Naval Research. She has been an active member and elected board member of the INFORMS Simulation Society. She has also co-created SimOpt: an open-source library and benchmarking platform for simulation optimization. Her research has contributed to renewable energy, climate adaptation, advanced manufacturing, and public health application areas.

Please note that the meeting URL for this event can be seen on the cels-seminars website which requires an Argonne login.