Deepak Mehta

IJCAI Conference 2009 Conference Paper

• David Lesaint
• Deepak Mehta
• Barry O’Sullivan
• Luis Quesada
• Nic Wilson

Hard and soft precedence constraints play a key role in many application domains. In telecommunications, one application is the configuration of callcontrol feature subscriptions where the task is to sequence a set of user-selected features subject to a set of hard (catalogue) precedence constraints and a set of soft (user-selected) precedence constraints. When no such consistent sequence exists, the task is to find an optimal relaxation by discarding some features or user precedences. For this purpose, we present the global constraint SOFTPREC. Enforcing Generalized Arc Consistency (GAC) on SOFT- PREC is NP-complete. Therefore, we approximate GAC based on domain pruning rules that follow from the semantics of SOFTPREC; this pruning is polynomial. Empirical results demonstrate that the search effort required by SOFTPREC is up to one order of magnitude less than the previously known best CP approach for the feature subscription problem. SOFTPREC is also applicable to other problem domains including minimum cutset problems for which initial experiments confirm the interest.

IJCAI Conference 2007 Conference Paper

• Deepak Mehta
• M. R. C. van Dongen

Constraint Satisfaction Problems (CSPs) are ubiquitous in Artificial Intelligence. The backtrack algorithms that maintain some local consistency during search have become the de facto standard to solve CSPs. Maintaining higher levels of consistency, generally, reduces the search effort. However, due to ineffective constraint propagation, it often penalises the search algorithm in terms of time. If we can reduce ineffective constraint propagation, then the effectiveness of a search algorithm can be enhanced significantly. In order to do so, we use a probabilistic approach to resolve when to propagate and when not to. The idea is to perform only the useful consistency checking by not seeking a support when there is a high probability that a support exists. The idea of probabilistic support inference is general and can be applied to any kind of local consistency algorithm. However, we shall study its impact with respect to arc consistency and singleton arc consistency (SAC). Experimental results demonstrate that enforcing probabilistic SAC almost always enforces SAC, but it requires significantly less time than SAC. Likewise, maintaining probabilistic arc consistency and maintaining probabilistic SAC require significantly less time than maintaining arc consistency and maintaining SAC.