Optimizing SMT Solving Strategies by Learning with an Evolutionary Process

TitreOptimizing SMT Solving Strategies by Learning with an Evolutionary Process
Type de publicationCommunication
TypeCommunication avec actes dans un congrès
Date du colloque16-20/07/2018
Titre du colloqueInternational Conference on High Performance Computing & Simulation : Pacos 2018
Titre des actes ou de la revue2018 International Conference on High Performance Computing & Simulation (HPCS)
AuteurGálvez Ramírez, Nicolás
1, 2
, Monfroy, Eric , Saubion, Frédéric , Castro, Carlos
ISBN 978-1-5386-7879-4
Résumé en anglais

This paper deals with program optimization, i.e., learning of more efficient programs. The programs we want to improve are Z3 solving strategies. Z3 is a SMT (SAT Modulo Theory) solver which is currently developed by Microsoft Research. We define strategy generators based on evolutionary processes. SMT solving strategies include various aspects that can affect the performance of a SMT solver dramatically. Each of these elements includes a huge amount of options which cannot be exploited without expert knowledge. We define a generic evolutionary algorithm based on genetic programming concepts. This strategy generation process aims at learning better strategies by successive improvements, using rules that can be combined in order to handle both structures and parameters of the strategies. We experiment 7 different strategies generators on 2 SMT logics (QF_LRA,QF_LIA). The results show that the learned strategies improve default strategies available in the solver.

URL de la noticehttp://okina.univ-angers.fr/publications/ua17116
Lien vers le document en ligne