TY - GEN
T1 - Effect of Initial Assignment on Local Search Performance for Max Sat
AU - Berend, Daniel
AU - Twitto, Yochai
N1 - Publisher Copyright:
© 2020 Schloss Dagstuhl- Leibniz-Zentrum fur Informatik GmbH, Dagstuhl Publishing. All rights reserved.
PY - 2020/6/1
Y1 - 2020/6/1
N2 - In this paper, we explore the correlation between the quality of initial assignments provided to local search heuristics and that of the corresponding final assignments. We restrict our attention to the Max r-Sat problem and to one of the leading local search heuristics-Configuration Checking Local Search (CCLS). We use a tailored version of the Method of Conditional Expectations (MOCE) to generate initial assignments of diverse quality. We show that the correlation in question is significant and long-lasting. Namely, even when we delve deeper into the local search, we are still in the shadow of the initial assignment. Thus, under practical time constraints, the quality of the initial assignment is crucial to the performance of local search heuristics. To demonstrate our point, we improve CCLS by combining it with MOCE. Instead of starting CCLS from random initial assignments, we start it from excellent initial assignments, provided by MOCE. Indeed, it turns out that this kind of initialization provides a significant improvement of this state-of-the-art solver. This improvement becomes more and more significant as the instance grows. 2012 ACM Subject Classification Theory of computation ! Theory of randomized search heuristics; Theory of computation ! Stochastic approximation.
AB - In this paper, we explore the correlation between the quality of initial assignments provided to local search heuristics and that of the corresponding final assignments. We restrict our attention to the Max r-Sat problem and to one of the leading local search heuristics-Configuration Checking Local Search (CCLS). We use a tailored version of the Method of Conditional Expectations (MOCE) to generate initial assignments of diverse quality. We show that the correlation in question is significant and long-lasting. Namely, even when we delve deeper into the local search, we are still in the shadow of the initial assignment. Thus, under practical time constraints, the quality of the initial assignment is crucial to the performance of local search heuristics. To demonstrate our point, we improve CCLS by combining it with MOCE. Instead of starting CCLS from random initial assignments, we start it from excellent initial assignments, provided by MOCE. Indeed, it turns out that this kind of initialization provides a significant improvement of this state-of-the-art solver. This improvement becomes more and more significant as the instance grows. 2012 ACM Subject Classification Theory of computation ! Theory of randomized search heuristics; Theory of computation ! Stochastic approximation.
KW - Combinatorial optimization
KW - Local search
KW - Maximum satisfiability
KW - Probabilistic algorithms
UR - http://www.scopus.com/inward/record.url?scp=85088164990&partnerID=8YFLogxK
U2 - 10.4230/LIPIcs.SEA.2020.8
DO - 10.4230/LIPIcs.SEA.2020.8
M3 - Conference contribution
AN - SCOPUS:85088164990
T3 - Leibniz International Proceedings in Informatics, LIPIcs
BT - 18th International Symposium on Experimental Algorithms, SEA 2020
A2 - Faro, Simone
A2 - Cantone, Domenico
PB - Schloss Dagstuhl- Leibniz-Zentrum fur Informatik GmbH, Dagstuhl Publishing
T2 - 18th International Symposium on Experimental Algorithms, SEA 2020
Y2 - 16 June 2020 through 18 June 2020
ER -