LANS Informal Seminar: Ankush Chakrabarty, Ph. D.

Seminar Title: Approximate Computing and Machine Learning for Nonlinear Model Predictive Control with Severe Resource Constraints
Speaker: Ankush Chakrabarty, Ph. D., Postdoctoral Fellow, John A. Paulson School of Engineering and Applied Sciences, Harvard University

Date/Time: 2017-06-20 15:00
Location: Bldg. 240, room 1405

Description:
Nonlinear model predictive control (NMPC) has demonstrated excellent regulatory and tracking performance in a wide range of complex systems since its inception. However, a critical challenge in real-time implementation of this control algorithm is the computational expenditure incurred by repeatedly solving non-convex finite-horizon optimal control problems. With the advent of resource-limited ubiquitous com-puting platforms such as edge devices in the internet-of-things or wearable/implantable technology, it is therefore crucial to investigate variants of the NMPC algorithm capable of deployment under severe re-source constraints.

In this talk, we describe how to formulate approximate NMPCs via supervised learning, sparse sampling, and polynomial regression. We also leverage tools from approximate computing to systematically trade-off controller performance in order to attain ultra-fast speeds of operation, extremely low power consumption, and highly miniaturized implementations of NMPC on embedded platforms. The potential of the proposed algorithm is demonstrated via hardware-in-the-loop experiments: we observe controller updates in nanosec-onds, power consumption <0.5 mW, and an area footprint of <1K gates on a two-dimensional nonlinear system. Finally, we provide sufficient conditions that guarantee feasibility and stability of the approximate NMPC.