Volume 39, Number 1, January-March 2005
|Page(s)||55 - 74|
|Published online||15 July 2005|
About the choice of the variable to unassign in a decision repair algorithm
LAAS-CNRS, Toulouse, France; email@example.com; firstname.lastname@example.org
Accepted: 8 February 2005
The decision repair algorithm (Jussien and Lhomme, Artificial Intelligence 139 (2002) 21–45), which has been designed to solve constraint satisfaction problems (CSP), can be seen, either (i) as an extension of the classical depth first tree search algorithm with the introduction of a free choice of the variable to which to backtrack in case of inconsistency, or (ii) as a local search algorithm in the space of the partial consistent variable assignments. or (iii) as a hybridisation between local search and constraint propagation. Experiments reported in Pralet and Verfailllie (2004) show that some heuristics for the choice of the variable to which to backtrack behave well on consistent instances and that other heuristics behave well on inconsistent ones. They show also that, despite its a priori incompleteness, decision repair, equipped with some specific heuristics, can solve within a limited time almost all the consistent and inconsistent randomly generated instances over the whole constrainedness spectrum. In this paper, we discuss the heuristics that could be used by decision repair to solve consistent instances, on the one hand, and inconsistent ones, on the other hand. Moreover, we establish that some specific heuristics make decision repair complete.
Key words: Constraint satisfaction problem / depth first tree search / local search / constraint propagation / backtrack / heuristics / completeness.
© EDP Sciences, 2005
Current usage metrics show cumulative count of Article Views (full-text article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.
Data correspond to usage on the plateform after 2015. The current usage metrics is available 48-96 hours after online publication and is updated daily on week days.
Initial download of the metrics may take a while.