From Information Theoretic Control and Learning to Non-Equilibrium Stochastic Thermodynamics: Connections, Interdependencies and Scalable Algorithms
Abstract
This research addresses fundamental questions on embodied learning control for autonomous systems in engineering and applied physics. These systems have to operate in dynamic and uncertain environments and for that they have to be equipped with computational processes to support decision making, learning and adaptation. In addition, they have to be able to explore the environment and exploit their complex dynamics to achieve agility and dexterity. Optimization and improvement of the aforementioned capabilities takes place under the existence of computational, thermodynamic and energetic constraints imposed by the underlying organization and structure of the system in consideration. Motivated by these challenges, in this research we take a holistic view of embodied learning and control to address foundational questions such as: what is thermodynamics ef?ciency 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 relates 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? To address the aforementioned questions we bring expertise from the disciplines of stochastic optimal control and information theory, statistical physics, and machine learning. Our technical approach will rely on i) optimality principles in stochastic control theory, ii) ?uctuation theorem in non-equilibrium stochastic thermodynamics, iii) information theoretic variational inequalities and iv) machine learning algorithms for uncertainty representation and uncertainty quanti?cation. The research will result in novel algorithms for real time planning, perception and control for general classes of systems and develop theories on fundamental scienti?c questions at the core of autonomy and applied physics. These intellectual products will expand the operational capabilities of existing autonomous systems and advance autonomy.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jul 09, 2020
- Source ID
- W911NF2010151
Entities
People
- Evangelos A. Theodorou
Organizations
- Army Contracting Command
- Georgia Tech Research Corporation
- United States Army