This article was accepted into the corpus but its outbound wikilinks were never NER-processed — typical at the deepest BFS hop or when the run's entity cap was reached. No expansion funnel to show.
| SATzilla | |
|---|---|
| Name | SATzilla |
| Developer | Portfolio of academic groups |
| First release | 2003 |
| Latest release | 2012 (major public releases) |
| Genre | Algorithm portfolio, automated algorithm selection, solver scheduling |
| License | Mixed academic licenses |
SATzilla
SATzilla is an algorithm portfolio system for selecting and scheduling propositional satisfiability (SAT) solvers on hard instances. It uses empirical performance models and machine learning to predict which among a set of competitive solvers will solve a given instance fastest, combining techniques from Carnegie Mellon University, University of British Columbia, Dresden University of Technology, Max-Planck-Institute for Informatics research groups and contributors from solver competitions such as the SAT Competition. SATzilla influenced research on automated algorithm selection, parameter tuning, and per-instance configuration.
SATzilla frames solver selection as a supervised learning problem: given a set of syntactic and semantic features extracted from a CNF formula, predict runtime or success for candidate solvers and construct a schedule or a single best choice. Its design connects work from Richard M. Karp-inspired algorithm analysis, the Leuven Algorithmic Game Theory community, and machine learning efforts at Massachusetts Institute of Technology and University of California, Berkeley. Typical feature sets include graph-based measures related to the Erdős–Rényi model and clause-variable incidence graphs studied by researchers at Technische Universität München. The system motivated cross-pollination between the International Joint Conference on Artificial Intelligence and the Conference on Neural Information Processing Systems.
The earliest incarnations emerged in the early 2000s amid progress by teams associated with the DIMACS challenges and members of the Zuse Institute Berlin and University of British Columbia who pooled solver portfolios that included versions of MINISAT, CryptoMiniSat, and PicoSAT. Prominent releases used ridge regression, Gaussian processes, and random forests inspired by methodologies from Stanford University and the Weizmann Institute of Science. Major public evaluation rounds were tied to the annual SAT Challenge and the SAT Competition where new solver entries from groups at Princeton University, University of Waterloo, University of Toronto, and University of Sydney were evaluated. Funding and collaboration came from agencies including the National Science Foundation and the Deutsche Forschungsgemeinschaft.
SATzilla’s pipeline combines feature computation, empirical performance modeling, and scheduling. Feature extraction components borrow graph and combinatorial metrics studied at Cornell University and University of Oxford, such as clause-to-variable ratios and community structure akin to analyses from École Polytechnique Fédérale de Lausanne. Performance models employ regression and classification approaches drawn from work at University College London and University of Pennsylvania, often using algorithm configuration tools related to ParamILS developed by researchers at the University of British Columbia. Scheduling layers adapt techniques from the Round-Robin and portfolio literature, integrating cutoff strategies comparable to those discussed at International Symposium on Experimental Algorithms meetings. The modular architecture allowed inclusion of heterogeneous solvers like conflict-driven clause learning engines and lookahead procedures from teams at University of California, Irvine and University of Helsinki.
Evaluations centered on benchmark suites assembled by the SAT Competition and the SAT Challenge, including industrial instances derived from groups in the Electronic Design Automation industry and crafted mathematical problems related to instances studied at Los Alamos National Laboratory. SATzilla demonstrated substantial improvements over single-best solvers on heterogeneous instance sets, often judged by PAR10 and solved-instance counts used by researchers at University of Pennsylvania and University of British Columbia. Meta-analyses published in venues such as the Journal of Artificial Intelligence Research compared SATzilla variants against algorithm selection baselines from ETH Zurich and University of Toronto, showing that accurate feature computation and robust training sets were critical to generalization.
SATzilla and its derivatives regularly influenced rankings at the SAT Competition where entrants included solvers from Microsoft Research and academic teams from University of Waterloo. The approach catalyzed the wider algorithm selection community that later formed benchmarks in the International Conference on Machine Learning and the Principles and Practice of Constraint Programming community. SATzilla’s success spurred commercial interest from firms engaged in verification and synthesis, notably companies collaborating with researchers at Carnegie Mellon University and Stanford University, and inspired algorithm-selection tracks within subsequent solver competitions.
Research spawned numerous variants: cost-sensitive and censored-regression adaptations proposed by groups at Dresden University of Technology and the University of British Columbia, online and instance-adaptive portfolios from teams at Microsoft Research and Google Research, and parallel portfolio schedulers developed at ETH Zurich and École Normale Supérieure. Successor frameworks incorporated algorithm configuration systems like SMAC and combined per-instance selection with automated parameter tuning pioneered in work from Humboldt University of Berlin. Broader algorithm-selection toolkits for combinatorial problems drew on SATzilla’s paradigms and were applied to domains championed by Massachusetts Institute of Technology and Princeton University.
Practitioners applied SATzilla-style selection in industrial verification workflows at organizations such as Intel Corporation and IBM Research for hardware model checking instances and in formal methods toolchains used by teams at Microsoft Research and Nokia Research Center. Academic users incorporated portfolio-based selection into experimental pipelines for constraint satisfaction and automated reasoning, collaborating with labs at University of Oxford and Weizmann Institute of Science. Implementations require benchmark corpora and solver binaries from contributors like MINISAT and CryptoMiniSat teams; users typically retrain models when target instance distributions change, following reproducibility practices advocated at the International Conference on Learning Representations and by the Association for Computing Machinery.
Category:Automated theorem proving Category:Algorithm selection Category:Satisfiability