Decision-Theoretic Sequential Decision Making for Observer-Aware, Goal-Directed Behavior in Swarms - ONR White Paper Tracking Number 20-000001141
Abstract
This proposal aims to study the applicability of decision-theoretic multiagent models in combination with new approaches to hierarchical abstraction to generate control policies resulting in goal-directed, observer-aware behavior in robot swarms. It will develop new methods that apply deep learning to partially observable tasks, extend these methods to include the discovery of abstraction hierarchies to scale to larger domains, and finally apply these approaches to the multi-agent case. Additionally, it will develop new methods incorporating auxiliary objective functions that enable the swarm to generate behavior based on measures of an observer s awareness of its objective (e.g., to produce easily interpretable behavior). The scope of the work includes establishing a mathematical basis of the proposed behaviors, model development, computer simulation, and algorithm analysis to understand the requirements for incorporating both purposeful and random behavior into the design of a large multirobot system.The proposal will result in new algorithms for solving partially-observable single- and multiagent problems, using abstraction hierarchies. These algorithms will be widely applicable.Additionally, it will result in new methods for incorporating both purposeful and random behavior into the design of a large multirobot system.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 05, 2021
- Source ID
- N000142112200
Entities
People
- George Konidaris
Organizations
- Brown University
- Office of Naval Research
- United States Navy