Data Analytics and Network Optimization
The availability of ever-increasing amounts of data from multitudes of sources is rapidly transforming the way we approach control and optimization problems. While model-driven methods remain at the heart of how we approach most problems, there is now an increasingly important data-driven component to be incorporated with existing methods, while new ones are continuously under development.
In the case of Cyber-Physical Systems and Multi-Agent Systems, the models involved are almost always network systems. In such networked settings, distributed optimization plays a crucial role not only to achieve scalability but also avoid the well-known pitfalls of centralized control and optimization schemes.
A large part of ongoing research in the CODES Lab has been focusing on data-driven optimization methods based on the theory of Perturbation Analysis (PA) which provides gradient estimates from directly observed data without requiring stochastic models that are difficult to build. One of the recent breakthroughs in the development of the PA theory is showing that it can be used with virtually arbitrary stochastic hybrid systems and it is characterized by intrinsic robustness properties with respect to unknown characteristics of the stochastic processes involved. This has led to a large class of Infinitesimal Perturbation Analysis (IPA) algorithms and an IPA Calculus of wide applicability.
A related part of our research aims at developing asynchronous event-driven distributed algorithms for network optimization with applications to a variety of Multi-Agent Systems. Many interactive applets and videos from networked robotic team experiments done in our laboratory can be found at CODES Lab.