In the past and now still, optimization has been the key tool that underlies many problems in both machine learning and control. In machine learning, the technology behind the training of most modern classifiers relies in a fundamental way on optimization. In control, recent advances in nonconvex optimization are leading to a paradigm shift from classical linear control to a principled framework for design of nonlinear controllers that are provably safer, more agile, and more robust.

Nowadays we also observe an inverse phenomenon where ideas from control and learning lead to fascinating new research within the optimization community. For example, the focus of the learning community on scalable and computationally tractable methods has forced optimizers to develop new variants of classical algorithms that are able to keep pace with the vast quantities of data at hand. Optimizers are also thinking harder about notions of optimality that better incorporate robustness as, in many problems in learning, over-optimizing a model on training data leads to poor generalization on test data. In addition, a bevy of interesting theoretical questions in optimization have come out of the success of popular heuristics used in the learning community (e.g., for training of neural networks) and are yet to be answered. In a similar way, fundamental ideas in control theory have led to new algorithmic developments within optimization. For example, a new research direction has emerged that interprets a multitude of optimization algorithms as dynamical systems and uses tools from robust control to improve their convergence rate analyses. Finally, there is exciting new activity at the juncture between machine learning and control. In particular, the problem of learning the parameters of a dynamical system on the fly, while simultaneously controlling it, has received much renewed attention. Advances in this direction are important for various applications in robotics and cyber-physical systems, among other areas of high impact.

The goal of the first edition of the Princeton Day of Optimization is to inspire new research directions at the interface of optimization, control, and machine learning, such as the ones mentioned above. To this effect, we have reached out to prominent researchers whose work straddles at least two of the three areas. We are excited to hear what these wonderful speakers have to say, and we very much hope to see you all there!

—Amir Ali Ahmadi

We gratefully acknowledge the support of our sponsors without whom this event would not have been possible: the department of Operations Research and Financial Engineering (ORFE), the Center for Statistics and Machine Learning (CSML), the School of Engineering and Applied Science (SEAS), the department of Mechanical and Aerospace Engineering (MAE), all at Princeton University, the IBM Thomas J. Watson Research Center, as well as the National Science Foundation.