Competitive native solvers for Answer Set Programming (ASP) perform a backtracking search by assuming the truth of literals. The choice of literals (the heuristic) is fundamental for the performance of these systems. Most of the efficient ASP systems employ a heuristic based on look-ahead, that is, a literal is tentatively assumed and its heuristic value is based on its deterministic consequences. However, looking ahead is a costly operation, and indeed look-ahead often accounts for the majority of time taken by ASP solvers. For Satisfiability (SAT), a radically different approach, called look-back heuristic, proved to be quite successful: Instead of looking ahead, one uses information gathered during the computation performed so far, thus looking back. In this approach, atoms which have been frequently involved in inconsistencies are preferred. In this paper, we carry over this approach to the framework of disjunctive ASP. We design a number of look-back heuristics exploiting peculiarities of ASP and implement them in the ASP system DLV. We compare their performance on a collection of hard ASP programs both structured and randomly generated. These experiments indicate that a very basic approach works well, outperforming all of the prominent disjunctive ASP systems - DLV (with its traditional heuristic), GnT, and CModels3 - on many of the instances considered.
Experimenting with Look-Back Heuristics on Hard ASP Programs
LEONE, Nicola;MARATEA M;RICCA, Francesco
2007-01-01
Abstract
Competitive native solvers for Answer Set Programming (ASP) perform a backtracking search by assuming the truth of literals. The choice of literals (the heuristic) is fundamental for the performance of these systems. Most of the efficient ASP systems employ a heuristic based on look-ahead, that is, a literal is tentatively assumed and its heuristic value is based on its deterministic consequences. However, looking ahead is a costly operation, and indeed look-ahead often accounts for the majority of time taken by ASP solvers. For Satisfiability (SAT), a radically different approach, called look-back heuristic, proved to be quite successful: Instead of looking ahead, one uses information gathered during the computation performed so far, thus looking back. In this approach, atoms which have been frequently involved in inconsistencies are preferred. In this paper, we carry over this approach to the framework of disjunctive ASP. We design a number of look-back heuristics exploiting peculiarities of ASP and implement them in the ASP system DLV. We compare their performance on a collection of hard ASP programs both structured and randomly generated. These experiments indicate that a very basic approach works well, outperforming all of the prominent disjunctive ASP systems - DLV (with its traditional heuristic), GnT, and CModels3 - on many of the instances considered.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.