Argonne National Laboratory

Upcoming Events

Optimal Control Under Information Constraints for Scheduling and Neuroscience

MCS Seminar
Yorie Nakahira, California Institute of Technology
March 13, 2018 10:00AM to 11:00AM
Building 240, Room 1404

Abstract: This talk focuses on control theory under information constraints, with applications to scheduling, sensorimotor, and biomolecular systems. Information constraints can arise from many practical scenarios (e.g., unknown disturbances, future uncertainties, distributed computation, or noisy communication). Therefore, achieving robustness and efficiency under these information constraints is a universal challenge underlying the design of complex systems.

In the first part of my talk, I will describe a few scalable (distributed) algorithms that are optimal for different settings of deadline scheduling. For maximizing the predictability and stability of the service capacity, we introduce exact scheduling, which works by finishing jobs exactly at their deadlines by using constant speeds and is easy to implement in large-scale systems. We show that under stationary Poisson arrivals, exact scheduling is optimal among all distributed algorithms; under nonstationary job arrivals, a variation of exact scheduling is Pareto-optimal for minimizing the variability and stability of the service capacity among distributed algorithms. These algorithms are evaluated in the Caltech Electrical Vehicle (EV) Testbed as well as by using the Google EV dataset. Compared with the optimal performance of the best offline algorithm or the best centralized algorithm, the performance of the proposed algorithms is also competitive.

In the second part of my talk, I will summarize the control and information theories that I developed to understand the design principles of the sensorimotor motor system and biomolecular systems. Control theory has been used to study human sensorimotor control (at the system level), while separately, information theory has been used to study the information transmission of individual neurons (at the component level). However, because of the lack of an integrated theory, little work has characterized how the constraints at the component level impact the performance at the system level. We are making efforts to provide a holistic perspective of the system level and the component level by developing an integrated theory. This theory will also help characterize the fundamental trade-offs of biomolecular systems and potentially other engineered systems.