Safe Distributed Reinforcement Learning for Large-Scale Systems
Abstract
Systematic ways to integrate diverse, heterogeneous, and possibly time-varying and uncertain components into complex autonomous systems while guaranteeing system level properties define a holy grail in the science of assured long-term autonomy. With much work being done already on topics such as safe machine learning or reinforcement learning to obtain guarantees on performance and safety of learning enabled autonomous systems, we focus on the next challenging step: how to provide such guarantees in a multi-agent system where multiple learning components are interacting. Our approach combines reinforcement learning with techniques and concepts from optimization and control theory to obtain scalable and structured policies that ensure safety. These techniques include (primal-)dual gradient updates, resilient distributed optimization, and imposition of control-relevant properties such as dissipativity through control barrier functions. The output of this project will be scalable reinforcement learning algorithms for autonomous multi-agent systems with analytical guarantees on safety. Relevance to ARO will be especially maintained through an ongoing collaboration with ARL researchers.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jul 28, 2023
- Source ID
- W911NF2310266
Entities
People
- Hao Zhu
Organizations
- Army Contracting Command
- United States Army
- University of Texas at Austin