Authors:
Ole-Christoffer Granmo
1
and
Noureddine Bouhmala
2
Affiliations:
1
University of Agder, Norway
;
2
Vestfold University College, Norway
Keyword(s):
Satisfiability problem, GSAT, Learning automata, Combinatorial optimization, Stochastic learning.
Related
Ontology
Subjects/Areas/Topics:
Artificial Intelligence
;
Computational Intelligence
;
Constraint Satisfaction
;
Evolutionary Computing
;
Knowledge Discovery and Information Retrieval
;
Knowledge-Based Systems
;
Machine Learning
;
Soft Computing
;
State Space Search
;
Symbolic Systems
Abstract:
The Satisfiability (SAT) problem is a widely studied combinatorial optimization problem with numerous applications, including time tabling, frequency assignment, and register allocation. Among the simplest and most effective algorithms for solving SAT problems are stochastic local-search based algorithms that mix greedy hill-climbing (exploitation) with random non-greedy steps (exploration). This paper demonstrates how the greedy and random components of the well-known GSAT Random Walk (GSATRW) algorithm can be enhanced with Learning Automata (LA) based stochastic learning. The LA enhancements are designed so that the actions that the LA chose initially mimic the behavior of GSATRW. However, as the LA explicitly interact with the SAT problem at hand, they learn the effect of the actions that are chosen, which allows the LA to gradually and dynamically shift from random exploration to goal-directed exploitation. Randomized and structured problems from various domains, including SAT-e
ncoded Logistics Problems, and Block World Planning Problems, demonstrate that our LA enhancements significantly improve the performance of GSATRW, thus laying the foundation for novel LA-based SAT solvers.
(More)