DID YOU KNOW? DTIC has over 3.5 million final reports on DoD funded research, development, test, and evaluation activities available to our registered users. Click
HERE to register or log in.
Accession Number:
AD1064069
Title:
The Statistical Physics of Stochastic Optimal Control and Learning
Descriptive Note:
Technical Report,01 Sep 2015,31 Aug 2016
Corporate Author:
Georgia Tech Research Corporation Atlanta United States
Report Date:
2016-11-17
Pagination or Media Count:
13.0
Abstract:
This workshop aims to identify interconnections between the areas of statistical physics, stochastic control theory, and learningadaptation and investigate open questions and future research directions at the intersection of these fields. The investigation will have a theoretical and an application component. Starting with connections between statistical physics and optimal control, concepts from statistical physics have recently been used towards the development of scalable algorithms for stochastic control. Grounded on the fundamental connections between partial differential equations and stochastic differential equations, the aforementioned tools provide probabilistic representations of solutions of partial differential equations and generated new algorithms for stochastic control using forward sampling of stochastic differential equations. On the other side, optimality principles in control theory namely the Pontryagin Maximum principle and Dynamic Programming have been used to generalize fluctuation theorems in stochastic thermodynamics. The use of aforementioned optimality principles has contributed towards a better understanding of the role that feedforward and feedback control laws play in stochastic thermodynamic systems. In machine learning, there have been new probabilistic methods for inference, regression and uncertainty representation. Algorithms for statistical inference have interpretations drawn from statistical physics and stochastic thermodynamics. These interpretations have the potential to result in new algorithms for statistical inference and create a direct connection between statistical physics and machine learningadaptation.
Distribution Statement:
APPROVED FOR PUBLIC RELEASE
