Constructing Abstraction Hierarchies for Robust, Real-Time Control
Abstract
This project will develop algorithms that enable artificial agents to make rapid decisions in uncertain environments. Previous research has focused on two distinct approaches to improving decision making: one in which the agent constructs skills that encapsulate frequently executed sequences of actions, and another in which the agent constructs abstract representations that discard unnecessary detail to focus only on the information relevant to decision-making. This project will combine these two approaches to build hierarchies where each layer is composed of a collection of skills and the abstract representation required for deciding which skill should be executed next. It will focus on the basic theory of how such hierarchies work: how they can be constructed, and how plans can be generated from them in particular situations. Finally, it will address the problem of finding the optimal hierarchy, both in the case where we wish to minimize average planning time over a distribution of tasks, and where we wish to ensure rapid decision-making in critical or dangerous scenarios.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 02, 2017
- Source ID
- FA95501710124
Entities
People
- George Konidaris
Organizations
- Air Force Office of Scientific Research
- Brown University
- United States Air Force