2018 AFOSR MURIVerifiable, Control-Oriented Learning On The FlyPO: Dr. Frederick Leve, Dynamics and Control: dycontrol@us.af.milPI: Dr. Ufuk Topcu, The University of Texas at AustinMURI Website
The proposed effort will develop a theoretical and algorithmic foundation for run-time learning and control for physical, autonomous systems. The resulting algorithms will (i) adapt to unforeseen, possibly abrupt changes in the system and its environment; (ii) establish verifiable guarantees---in a sense to be precisely defined---with respect to high-level safety and performance specifications; and (iii) obey and leverage the laws of physics and contextual knowledge. They will also provide quantitative trade-offs between the strengths of their guarantees, the amount of run-time data and a priori side information necessary to establish such guarantees, and the computational requirements.
The proposed approach treats the (limited) data that the system generates and the existing side information as the first-class objects of control-oriented learning. Such side information includes the physical laws that the system obeys, the context in which it serves, and the structure in its mathematical representation. The research plan builds on a coherent composition of learning, verification and synthesis:
Thrust I, on learning, will merge the side information with run-time data to derive bounds on model uncertainty with both finite-dimensional, parametric and infinite-dimensional, functional components.
Thrust II, on verification, will determine---using the learned models---whether a given control strategy retains the possibility of viable operation (e.g., making progress toward the mission objectives without jeopardizing future safety) and assess the likelihood of failure over practically relevant time horizons.
Thrust III, on joint learning and control, will develop strategies that effectively prioritize the dual tasks of learning the dynamics and minimizing the likelihood of mission failure.
The back-end computation engine will provide cross-cutting, efficient optimization algorithms addressing the needs of all thrusts. All three thrusts build on a common working principle supported by the front-end constraint engine: Data and side information must be jointly utilized for control-oriented learning.
The approach also embraces the fact that developing truly autonomous systems---and, in particular, control-oriented learning on the fly---is beyond the reach of any single discipline. It distills ideas---and in many cases discovers novel extensions---from a number of conventionally disparate disciplines into a unified foundation. These disciplines include analysis, control theory, dynamical systems, learning theory, optimization and formal methods. Many of these interdisciplinary connections are pursued for the first time in the proposed effort.
The fundamental research outcomes of the proposed work have the potential to pave the road for the DoD to develop truly autonomous systems aware of their high-level tasks, low-level, physical capabilities, and computational resources; operate in contested environments; and survive disruptions or recognize its impossibility.
The background and tight integration of the team will contribute toward creating a new field of on-the-fly learning receptive to the needs and opportunities in control for autonomous systems. The foundation to be established by the proposed effort will set the stage for more research by a broader community in the years to come.