Research Portfolio

Can a vehicle safely teach itself to avoid obstacles from scratch? Researchers approach the idea of flipping a switch and getting guaranteed safe control with a healthy dollop of skepticism - it's essentially impossible in the real world. Yet exploring this question reveals insights that motivate the development of new control algorithms. 

By investigating the minimal assumptions required for safe learning-based control, I have developed a suite of theory and practical algorithms that lend probabilistic robustness in a learning-based control setting.  This includes algorithms for nonlinear systems - even systems that necessitate learning a latent state space (e.g. visuomotor control synthesis and latent state-spaces). 

I have leveraged my work in this space to reduce charge times for lithium-ion batteries by over 30% compared to industry standards, reduce the energy consumption of YouTube's video compression infrastructure by over 4%, and create progress towards safe autonomous systems.

Modern trends show scalability is the difference maker in tackling complex tasks with machine learning. Despite rapid advancements in scalable ML, the cost in dollars and carbon to train large models is still prohibitive for most applications.

My research in this space explores information-geometric methods for improving the efficiency of training algorithms by exploiting sparsity. My vision is to create greater progress towards democratizing access to powerful tools, while reducing the energy expenditure of the broader industry. 

How do we explore degradation mechanisms in batteries when degradation happens across months or years? Where do we build a system of telescopes to best observe cosmic rays? Not all data are equal, and some data samples may cost thousands of dollars to observe. How do we determine the best manner to design an experiment to maximize its information gain? My work in this space explores analytic and practical methods to design optimal experiments for parameter inference. These insights are useful from designing battery management systems to improving system-wide fuel economy and safety of connected and automated vehicles. 

Stochastic optimization deals with uncertainty and randomness in the underlying system. If the true distributions of relevant random variables are unknown, how can we guarantee constraints will be satisfied in our optimization model? Provided samples from the true underlying distribution, distributionally robust optimization uses clever mathematical tools to guaranteeing feasibility subject to distributional uncertainty. 

My work in this space extends theory of distributionally robust optimization to a host of nonlinear and nonconvex spaces. I have created progress towards leveraging distributionally robust optimization to quantify uncertainty of machine learning forecasters, and solve optimal control problems for unstructured and high-dimensional problems.