View: session overviewtalk overviewside by side with other conferences
09:00 | Fast Sampling of Perfectly Uniform Satisfying Assignments SPEAKER: Zayd Hammoudeh ABSTRACT. Validation of complex digital circuitry primarily relies upon constrained-random simulation. However, it is intractable to simulate all satisfying assignments (models). Rather, sufficient verification coverage is achieved by selecting test stimuli uniformly at random from the set of valid models. The current state-of-the-art witness generators use hashing to perform this sampling near-uniformly. We present a new, perfectly-uniform witness generation algorithm based on the exact model counter sharpSAT and reservoir sampling. In experiments across hundreds of SAT benchmarks, our algorithm is faster than the state of the art by ten to over a hundred thousand times. |
09:30 | SPEAKER: Dimitris Achlioptas ABSTRACT. We present a probabilistic model counter that can trade off running time with approximation accuracy. As in several previous works, the number of models of a formula is estimated by adding random parity constraints (equations). Specifically, the binary logarithm of the number of models is approximated by the number of random constrains needed to eliminate all models. One key difference with prior works is that the systems of parity equations used by our algorithm are the parity check matrices of Low Density Parity Check (LDPC) error-correcting codes. As a result, the equations tend to be much shorter, often containing fewer than 10 variables each, making the search for models far more tractable. The price paid for computational tractability is that their corresponding statistical properties are not as good as when (much longer) constraints are used. We show how one can deal with this issue and derive rigorous approximation guarantees by performing more solver invocations. |
10:00 | SPEAKER: Markus Hecher ABSTRACT. In this paper, we introduce a novel algorithm to solve projected model counting (PMC). PMC asks to count solutions of a Boolean formula with respect to a given set of projected variables, where multiple solutions that are identical when restricted to the projected variables count as only one solution. Our algorithm exploits bounded primal or incidence treewidth of the input instance. It runs in time $O(2^{2^{k+4}} n^2)$ where k is the treewidth and n is the input size of the instance. In other words, we obtain that the problem PMC is fixed-parameter tractable when parameterized by treewidth. Further, we take the exponential time hypothesis (ETH) into consideration and establish lower bounds of bounded treewidth algorithms for PMC, which yields that runtime bounds of our algorithm are tight. |
14:00 | ABSTRACT. This paper describes the algorithm implemented in the QBF solver CQESTO, which has placed second in the non-CNF track of the last year's QBF competition. The algorithm is inspired by the CNF-based solver QESTO. Just as QESTO, CQESTO invokes a SAT solver in a black-box fashion. However, it directly operates on the circuit representation of the formula. The paper analyzes the individual operations that the solver performs on the circuit. |
14:30 | Symmetries for QBF SPEAKER: Manuel Kauers ABSTRACT. While symmetries are well understood for Boolean formulas and successfully exploited in practical SAT solving, little is known about symmetries in quantified Boolean formulas (QBF). There are some works introducing adoptions of propositional symmetry breaking techniques for QBF solving, with a theory covering only very specific parts of QBF symmetries. We present a general framework based on symmetric groups that give a concise characterization of symmetries in QBF. Our framework handles universal and existential symmetries in a dual manner and gives rise to a natural generalization of a known symmetry breaking technique from SAT. |
15:00 | SPEAKER: Martin Suda ABSTRACT. We develop new semantics for resolution-based calculi for Quantified Boolean Formulas, covering both the CDCL-derived calculi and the expansion-derived ones. The semantics is centred around the notion of a partial strategy for the universal player and allows us to show in a local, inference-by-inference manner that these calculi are sound. It also helps us understand some less intuitive concepts, such as the role of tautologies in long-distance resolution or the meaning of the ``star'' in the annotations of IRM. Furthermore, we show that a clause of any of these calculi can be, in the spirit of Curry-Howard correspondence, interpreted as a specification of the corresponding partial strategy. The strategy is total, i.e. winning, when specified by the empty clause. |
16:00 | QBF as an Alternative to Courcelle's Theorem SPEAKER: Valia Mitsou ABSTRACT. We propose reductions to quantified Boolean formulas (QBF) as a new approach to showing fixed-parameter linear algorithms for problems parameterized by treewidth. We demonstrate the feasibility of this approach by giving new algorithms for several well-known problems from artificial intelligence that are in general complete in the second level of the polynomial hierarchy. By reduction from QBF we show that all resulting algorithms are essentially optimal in their dependence on the treewidth. The problems that we consider were already known to be fixed-parameter linear by using Courcelle's theorem or dynamic programming, but we argue that our approach has clear advantages over these techniques: on the one hand, in contrast to Courcelle's theorem, we get concrete and tight guarantees for the runtime dependence on the treewidth; on the other hand, we avoid tedious dynamic programming and, after showing some normalization results for CNF-formulas, our upper bounds often boil down to a few lines. |
16:30 | Polynomial-Time Validation of QCDCL Certificates SPEAKER: Tomáš Peitl ABSTRACT. Quantified Boolean Formulas (QBFs) offer compact encodings of problems arising in areas such as verification and synthesis. These applications typically require that QBF solvers not only decide whether an input formula is true or false but also output a witnessing certificate. State-of-the-art QBF solvers based on Quantified Conflict-Driven Constraint Learning (QCDCL) can emit Q-resolution proofs, from which in turn certificates can be extracted. The correctness of a certificate generated in this way is validated by substituting it into the matrix of the input QBF and using a SAT solver to check that the resulting propositional formula (the validation formula) is unsatisfiable. This final check is often the most time-consuming part of the entire certification workflow. We propose a new validation method that does not require a SAT call and provably runs in polynomial time. It uses the Q-resolution proof from which the given certificate was extracted to directly generate a (propositional) proof of the validation formula in the RUP format, which can be verified by a proof checker such as DRAT-trim. Experiments with a prototype implementation show a robust, albeit modest, increase in the number of successfully validated certificates compared to validation with a SAT solver. |
Doors open at 4:30 pm; please be seated by 4:50 pm (attendance is free of charge and all are welcome; please register).