|Prof. Evangelos Theodorou|
|Prof. Mike DeWeese|
A proposal put forth by School of Aerospace Engineering professor Evangelos Theodorou to explore the fundamental questions influencing learning control for autonomous systems has been approved by the Army Research Office, an element of the U.S. Army Combat Capabilities Development Command’s Army Research Laboratory.
The five-year, $1.6 million project, "From Information Theoretic Control and Learning to Non-Equilibrium Stochastic Thermodynamics: Connections, Interdependencies and Scalable Algorithms" will support Theodorou and his co-PI, University of California-Berkeley physics professor Mike DeWeese in their investigation of the fundamental questions that influence learning control in autonomous systems.
"This is a project for basic research which allow us to think big and think holistically across disciplines,” said Theodorou, who heads up Georgia Tech's Autonomous Control and Decision Systems laboratory.
“Ultimately we want to push the frontier on stochastic optimal control theory and non-equilibrium stochastic thermodynamics to explore new connections and directions that can lead to novel algorithms for decision making under uncertainty. I think this research will have huge impact not only in Autonomy but is also in applied and statistical physics"
The duo's research could bear much fruit for ARO and for basic research, as it will take into account the dynamic and uncertain environments in which autonomous systems in the military routinely operate. Their work will focus in on decision-making, learning, and adaptation -- qualities whose optimization and improvement are heavily influenced by computational, thermodynamic and energetic constraints. Their collaboration will draw on Theodorou and DeWeese's expertise in stochastic optimal control and information theory, statistical physics, and machine learning.
"We will take a holistic view of the learning and control technology embedded in autonomous systems to address foundational questions such as: what is thermodynamics efficiency of decision-making algorithms and is there a unifying theory for the thermodynamics of computation in decision making? What is the role that nonlinearity, noise and morphology play in control of complex systems? How does low-level organization and architecture relate to computation and performance? Can existing connections between information theory and stochastic control generalize to systems operating at multiple temporal and spatiotemporal scales? What are the underlying trade-offs and computational mechanism for switching between model-free and model-based decision making?