Control-Based Design of Online Optimization Algorithms

In many applications – ranging from control, to image processing, to machine learning – we are interested in solving online optimization problems, that is, optimization problems whose data change over time. Designing online optimization algorithms presents a challenging set of constraints, foremost of which is the need for real-time operations. In this talk, I will discuss the fruitful application of model-based techniques to design efficient online algorithms. After a background on online optimization, I will discuss how we can leverage techniques from (robust) control theory to design online algorithms that outperform the state of the art. I will show theoretical and numerical results for both unconstrained and constrained optimization problems, and discuss some interesting future directions.