Hello. I am the lead AI research engineer/scientist at SigOpt (acquired by Intel in 2020). Currently, I work on productionizing Bayesian optimization, and more broadly, sequential decision making problems. I am especially interested in applying sequential optimization techniques in scientific and engineering domains such as materials simulation and design. Prior to SigOpt, I obtained my Ph.D. in electrical engineering from Princeton University, where I was advised by Prof. Warren B. Powell. My doctoral studies focused on approximate dynamic programming, stochastic optimization, and optimal learning, with an application in managing grid-level battery storage.
I spend most of my time at SigOpt developing and productionizing Baysian optimization software.
It takes a great deal of engineering effort and care to ensure that
the platorm is industry-grade, extensive, and robust under vastly different user demands.
I have lead the development of many features such as multiobjective optimization,
constrained optimization, model-aware optimization, as well as various backend computational improvements.
The bulk of my work is now available as open source software
sigopt-server
and
libsigopt
.
These projects are collaboration work with materials scientists from the University of Pittsburgh. At the high level, we frame materials design and discovery in the context of sequential decision making problems. In the first project, we develop a constrained Bayesian optimization method to accelerate the fabrication process of an optical device. In the second project, we use multiobjective Bayesian optimization to discover and study the Pareto optimal anti-reflective nanostructures through numerical simulations.
This project aims to co-optimize battery storage for multiple revenue streams. In particular, we are interested in the energy arbitrage and frequency regulation as the two main modes of operation. For the first time, we are able the model the problem down to the two-second resolution, which replicates the dynamics of the regulation signal. We also introduce the idea of low-rank value function approximation for backward dynamic programming.