Scalable Control using Learning and Adaptation

t the United Nations Summit 2015, our world leaders adopted 17 Sustainable Development Goals. A necessary condition for the completion of these goals is efficient, reliable, and safe infrastructure. For example, Goal 7: Affordable and Clean Energy requires infrastructure robust to loss of the inertia prevalent in conventional power plants, such as coal, gas, and nuclear power. As the nature of consumption and production changes, the networks’ structures and underlying control mechanisms must keep up. Unfortunately, many of the anticipated changes increase the load and introduce additional complexity. Examples are micro-producers of electricity, autonomous vehicles in transportation networks, and increased nodes in communication networks. As complexity can increase by orders of magnitude, controlling these networks requires models at an entirely new scale. Manually sustaining accurate models of individual components becomes infeasible. A solution is to use adaptation and learning to automatically learn and sustain models, taking care to do so in a reliable and scalable way.

In my doctoral studies, I address the fundamentals of scalable modeling's technical challenges using adaptation and learning. I study minimax control and graph realizability of controllers, meaning controllers that respect information exchange constraints in networks. The aim is to synthesize algorithms for scalable, robust adaptive control that automatically sustains accurate models of highly complex networks. Such algorithms can facilitate the complex technologies and infrastructures needed to reach the Sustainable Development Goals.

In order to ensure good future decision-making, you will need to explore your surroundings and learn from past data. This thesis explores a robust way of doing that in the control engineering setting.