Dynamic Decision-Problem Decomposition for Autonomous Systems in Complex Domains
Abstract
Constructing agents that can be autonomous for long periods of time in complex environments requires developing new methods for learning and planning. Existing methods do not scale to realistic applications such as an assistant in a hospital, home or supply depot. The key idea in this proposed work is to decompose the large problems faced by an autonomous agent into a sequence of smaller problems. These smaller problems can be obtained by ignoring some state variables, by imposing constraints that limit the size of the problem, by temporal hierarchical decomposition, by decomposing goals, and other methods of decomposition and abstraction. The goal in this project is to develop and integrated system that exploits all of these approaches so as to tackle problems that are beyond the reach of existing autonomous decision-making systems. Specific problems to be addressed are: * Year 1 Combine existing approaches to state-variable selection with the HPN hierarchical planning and execution framework. Extend state-variable decomposition for interchangeable objects and places. Demonstrate the approach in a simple non-robotic simulator. * Year 2 Integrate approaches to state-space decomposition, temporal decomposition and goal-based decomposition and demonstrate in a realistic simulator involving exogenous dynamics. * Year 3: Explore search and optimization methods for handling multiple goals. Demonstrate an integrated planning system in a realistic robotic simulation operating in a complex environment.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 26, 2023
- Source ID
- W911NF2310034
Entities
People
- Leslie P. Kaelbling
Organizations
- Army Contracting Command
- Massachusetts Institute of Technology
- United States Army