초청세미나

Black-Box Optimization for Optical Systems

■ Abstract

A black-box optimization problem frequently appears in various real-world applications such as experimental design, hyper-parameter optimization, material discovery, and optics. The goal of black-box optimization is to maximize the objective function by sequentially querying inputs and receiving noisy evaluations, i.e., bandit feedback. Then, an algorithm for this problem generally aims to find an optimal input with the minimum number of queries. The main benefit of black-box optimization is that it is possible to optimize objective functions that are non-differentiable, unknown, or stochastic, which can be easily encountered in real-world problems. However, to optimize the unknown stochastic objective function, a main challenge arises in dealing with the trade-off between estimating the unknown objective function from noisy evaluations (so-called exploration) and finding the optimal input based on the estimated function (so-called exploitation). This talk introduces some well-known optimization techniques that can overcome such trade-off and also addresses its applications to optical systems.

■ Biography

He is currently an assistant professor at the Chung-Ang University. He received his B.S. and Ph.D. degrees in electrical and computer engineering from Seoul National University in 2015 and 2021, respectively. His research interests include black-box optimization, bandit theory, reinforcement learning, stochastic optimal control, and their applications.

■ Contact : 정윤찬 교수(yoonchan@snu.ac.kr)