Dynamic Backtracking

In solving a search problem, it is important to remember which possible solutions you've considered and which you haven't. By doing so, you drastically reduce the number of possibilities you need to consider in the future. This general technique is known as "dependency maintenance" and, until recently, was of limited practical value because there were simply too many possible dependencies to maintain and use at each point in the search.

This technique has been made viable in practice by the introduction of "dynamic backtracking," or "relevance-bounded learning." The basic idea is to only store information about possible solutions that seem likely to be considered later during the overall search; if a piece of information seems unlikely to be useful later, it gets dropped. Relevance-bounded learning is the technique of choice on a wide range of problems of practical interest.

CIRL

Dynamic backtracking was invented at CIRL. It was renamed relevance-bounded learning by Bayardo and Schrag (U Texas at Austin).

Pointers

Dynamic Backtracking.
This article introduces dynamic backtracking, an intelligent backtracking method, developed by Matt Ginsberg. Appeared in JAIR 1993. Compressed postscript document.

Parent areas:
Systematic Methods

Implementation:
PLRS

Back to the CIRL Overview Page.