Learning to Plan in Hybrid Spaces
Abstract
Our goal is to make intelligent autonomous systems that can operate in complex, uncertain, dynamic domains over long time-horizons. Such systems typically involve domains that are best represented as hybrids of discrete and continuous aspects, including: (a) control of one or more autonomous vehicles for surveillance or other operations; (b) management of a depot or maintenance operation; and (c) control of a home robot companion for the elderly. The most effective way to construct such systems is to use general models of the domain (i.e., what the effects of taking various actions will be) combined with a general-purpose planning or reasoning mechanism. This strategy allows a system to synthesize new plans in response to new contingencies. General-purpose planning methods are highly effective but as the size of the space grows or the horizon becomes long, the search very quickly becomes intractable. So, planning grants generality, but at the cost of what seems like unmanageable computational complexity. Researchers have found several general strategies for taming this complexity: (1) Heuristic search methods that reduce the effective branching factor by focusing the search in “directions” that are likely to be useful; (2) Adaptation methods that generalize previous plans to more quickly deal with related problem instances; (3) Determinization methods that reduce the branching factor induced by uncertainty about action outcomes, by focusing on one or a small number of likely outcomes and then replanning if something unanticipated occurs; and (4) Abstraction methods that break a large planning problem into smaller ones. The focus of these has recently turned to general, domain-independent methods for improving planning performance. We propose to apply machine-learning techniques to learn from previous planning and execution examples to make general-purpose planning methods efficient for use in long-term autonomous operations in high-dimensional hybrid domains.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 02, 2017
- Source ID
- FA95501710165
Entities
People
- Leslie P. Kaelbling
Organizations
- Air Force Office of Scientific Research
- Massachusetts Institute of Technology
- United States Air Force