Statistical Mechanics for Learning Algorithmic- Based Controllers: The Role or Physics in New Computational Models for Real-Time Control
Abstract
Major Goals: Control of physical systems problems is challenging because of: (a) problem complexity stemming from the nonlinearity, non-holonomic dynamics, and non-convex objective functions; (b) scalability stemming from the high-dimensional state space; and (c) uncertainty owing to poorly characterized dynamics, the effect of the unknown environment or even an adversary. Current control synthesis methods attempting to address these problems are hindered by the difficulty in finding analytic, closed-form solutions. Most physical systems do not admit such nice solutions. Furthermore, owing to problem complexity, controller synthesis is done off-line; not adaptable to changes in operational constraints and mission requirements. Naive computational approaches (e.g., grid-based discretizations) are plagued by the curse of dimensionality. The key question of how to control safely and reliably high-dimensional complex, uncertain systems still remains extremely challenging.
Document Details
- Document Type
- Technical Report
- Publication Date
- Aug 31, 2018
- Accession Number
- AD1091296
Entities
People
- Evangelos A. Theodorou
- Panagiotis Tsiotras
Organizations
- Georgia Tech Research Corporation