Title: Improving the Practical Scalability and Robustness of Zeroth-Order Optimization Solvers

Abstract: 

Zeroth-order optimization solvers are often deployed in settings where little information regarding a problem's conditioning or noise level is known. Effective solvers must handle complex applications -- from automated materials discovery to quantum circuit compilation -- each presenting unique challenges. These problem features often limit the scale of the problems on which zeroth-order algorithms can be effectively deployed. We overcome this limitation through novel algorithms based on randomized subspace techniques. We also report on our experience developing adaptive algorithms, which leverage information learned online to adapt critical algorithmic features. We illustrate our approaches in trust-region-based reduced-space methods and show how trained policies can even be deployed effectively in nonstationary cases, where the noise seen changes over the decision space.


Bio: 

Stefan M. Wild is a Senior Scientist and Director of the Applied Mathematics and Computational Research Division at Lawrence Berkeley National Laboratory, a research lab primarily funded by the U.S. Department of Energy’s Office of Science. Wild is also adjunct faculty in Industrial Engineering and Management Sciences at Northwestern University. Wild received his Ph.D. in Operations Research and Information Engineering from Cornell University. Wild is a SIAM Fellow and his research has been recognized by the INFORMS Optimization Society's  Egon Balas Prize and the U.S. Presidential Early Career Award for Scientists and Engineers. Wild is Section Editor for SIAM Review and Associate Editor for Data Science in Science, INFORMS Journal on Computing, Journal of Optimization Theory and Applications, and Mathematical Programming Computation. Wild’s primary research focuses on developing model-based algorithms and software for challenging numerical optimization problems and automated learning under uncertainty, with the goal of accelerating and advancing scientific discoveries.