SAT Modulo Monotonic TheoriesbySam BaylessB. Sc., The University of British Columbia, 2010A THESIS SUBMITTED IN PARTIAL FULFILLMENTOF THE REQUIREMENTS FOR THE DEGREE OFDoctor of PhilosophyinTHE FACULTY OF GRADUATE AND POSTDOCTORAL STUDIES(Computer Science)The University of British Columbia(Vancouver)March 2017c© Sam Bayless, 2017AbstractSatisfiability Modulo Theories (SMT) solvers are a class of efficient constraintsolvers which form integral parts of many algorithms. Over the years, dozens ofdifferent Satisfiability Modulo Theories solvers have been developed, supportingdozens of different logics. However, there are still many important applications forwhich specialized SMT solvers have not yet been developed.We develop a framework for easily building efficient SMT solvers for previ-ously unsupported logics. Our techniques apply to a wide class of logics which wecall monotonic theories, which include many important elements of graph theoryand automata theory.Using this SAT Modulo Monotonic Theories framework, we created a new SMTsolver, MONOSAT. We demonstrate that MONOSAT improves the state of the artacross a wide body of applications, ranging from circuit layout and data centermanagement to protocol synthesis — and even to video game design.iiPrefaceThe work in this thesis was developed in a collaboration between the IntegratedSystems Design Laboratory (ISD) and the Bioinformatics and Empirical & Theo-retical Algorithmics Laboratory (BETA), at the University of British Columbia.This thesis includes elements of several published and unpublished works. Inaddition to my own contributions, each of these works includes significant contri-butions from my co-authors, as detailed below.Significant parts of the material found in chapters 3, 4, 5, 7, and 6.1 appearedpreviously at the conference of the Association for the Advancement of ArtificialIntelligence, 2015, with co-authors N. Bayless, H. H. Hoos, and A. J. Hu. I was thelead investigator in this work, responsible for the initial concept and its execution,for conducting all experiments, and for writing the manuscript.Significant parts of the material found in Chapter 8 appeared previously at theInternational Conference on Computer Aided Verification, 2016, with co-authorsT. Klenze and A. J. Hu. In this work, Tobias Klenze was the lead investigator; Ialso made significant contributions throughout, including playing a supporting rolein the implementation. I was also responsible for conducting all experiments.Significant parts of Section 6.2 were previously published at the InternationalConference on ComputerAided Design, 2016, with co-authors H. H. Hoos, andA. J. Hu. In this work, I was the lead investigator, responsible for conducting allexperiments, and for writing the manuscript. This work also included assistanceand insights from Jacob Bayless.Finally, Section 6.3 is based on unpublished work in collaboration with co-authors N. Kodirov, I. Beschastnikh, H. H. Hoos, and A. J. Hu, in which I was thelead investigator, along with Nodir Kodirov. In this work, I was responsible for theimplementation, as well as conducting all experiments and writing the manuscript.iiiTable of ContentsAbstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiPreface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiiTable of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ivList of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viiiList of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ixAcknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32.1 Boolean Satisfiability . . . . . . . . . . . . . . . . . . . . . . . . 32.2 Boolean Satisfiability Solvers . . . . . . . . . . . . . . . . . . . . 62.2.1 DPLL SAT Solvers . . . . . . . . . . . . . . . . . . . . . 72.2.2 CDCL SAT Solvers . . . . . . . . . . . . . . . . . . . . . 92.3 Many-Sorted First Order Logic & Satisfiability Modulo Theories . 132.4 Satisfiability Modulo Theories Solvers . . . . . . . . . . . . . . 172.5 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223 SAT Modulo Monotonic Theories . . . . . . . . . . . . . . . . . . . 243.1 Monotonic Predicates . . . . . . . . . . . . . . . . . . . . . . . 243.2 Monotonic Theories . . . . . . . . . . . . . . . . . . . . . . . . . 28iv4 A Framework for SAT Modulo Monotonic Theories . . . . . . . . . 304.1 Theory Propagation for Finite Monotonic Theories . . . . . . . . 324.2 Conflict Analysis for Finite Monotonic Theories . . . . . . . . . . 404.3 Compositions of Monotonic Functions . . . . . . . . . . . . . . . 434.4 Monotonic Theories of Partially Ordered Sorts . . . . . . . . . . . 484.5 Theory Combination and Infinite Domains . . . . . . . . . . . . . 525 Monotonic Theory of Graphs . . . . . . . . . . . . . . . . . . . . . . 555.1 A Monotonic Graph Solver . . . . . . . . . . . . . . . . . . . . . 565.2 Monotonic Graph Predicates . . . . . . . . . . . . . . . . . . . . 595.2.1 Graph Reachability . . . . . . . . . . . . . . . . . . . . . 605.2.2 Acyclicity . . . . . . . . . . . . . . . . . . . . . . . . . . 655.3 Weighted Graphs & Bitvectors . . . . . . . . . . . . . . . . . . . 685.3.1 Maximum Flow . . . . . . . . . . . . . . . . . . . . . . . 685.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 746 Graph Theory Applications . . . . . . . . . . . . . . . . . . . . . . . 766.1 Procedural Content Generation . . . . . . . . . . . . . . . . . . . 766.2 Escape Routing . . . . . . . . . . . . . . . . . . . . . . . . . . . 856.2.1 Multi-Layer Escape Routing in MONOSAT . . . . . . . . 876.2.2 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . 906.3 Virtual Data Center Allocation . . . . . . . . . . . . . . . . . . . 946.3.1 Problem Formulation . . . . . . . . . . . . . . . . . . . . 946.3.2 Multi-Commodity Flow in MONOSAT . . . . . . . . . . 966.3.3 Encoding Multi-Path VDC Allocation . . . . . . . . . . . 986.3.4 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . 996.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1077 Monotonic Theory of Geometry . . . . . . . . . . . . . . . . . . . . 1087.1 Geometric Predicates of Convex Hulls . . . . . . . . . . . . . . . 1117.1.1 Areas of Convex Hulls . . . . . . . . . . . . . . . . . . . 1117.1.2 Point Containment for Convex Hulls . . . . . . . . . . . . 1127.1.3 Line-Segment Intersection for Convex Hulls . . . . . . . . 1157.1.4 Intersection of Convex Hulls . . . . . . . . . . . . . . . . 116v7.2 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1188 Monotonic Theory of CTL . . . . . . . . . . . . . . . . . . . . . . . 1228.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1248.2 CTL Operators as Monotonic Functions . . . . . . . . . . . . . . 1278.3 Theory Propagation for CTL Model Checking . . . . . . . . . . . 1308.4 Conflict Analysis for the Theory of CTL . . . . . . . . . . . . . . 1348.5 Implementation and Optimizations . . . . . . . . . . . . . . . . . 1368.5.1 Symmetry Breaking . . . . . . . . . . . . . . . . . . . . 1368.5.2 Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . 1378.5.3 Wildcard Encoding for Concurrent Programs . . . . . . . 1378.6 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . 1398.6.1 The Original Clarke-Emerson Mutex . . . . . . . . . . . 1408.6.2 Mutex with Additional Properties . . . . . . . . . . . . . 1418.6.3 Readers-Writers . . . . . . . . . . . . . . . . . . . . . . 1439 Conclusions and Future Work . . . . . . . . . . . . . . . . . . . . . 146Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148A The MONOSAT Solver . . . . . . . . . . . . . . . . . . . . . . . . . 169B Monotonic Theory of Bitvectors . . . . . . . . . . . . . . . . . . . . 173C Supporting Materials . . . . . . . . . . . . . . . . . . . . . . . . . . 179C.1 Proof of Lemma 4.1.1 . . . . . . . . . . . . . . . . . . . . . . . . 179C.2 Proof of Correctness for Algorithm 7 . . . . . . . . . . . . . . . . 181C.3 Encoding of Art Gallery Synthesis . . . . . . . . . . . . . . . . . 184D Monotonic Predicates . . . . . . . . . . . . . . . . . . . . . . . . . . 187D.1 Graph Predicates . . . . . . . . . . . . . . . . . . . . . . . . . . 187D.1.1 Reachability . . . . . . . . . . . . . . . . . . . . . . . . 187D.1.2 Acyclicity . . . . . . . . . . . . . . . . . . . . . . . . . . 188D.1.3 Connected Components . . . . . . . . . . . . . . . . . . 189D.1.4 Shortest Path . . . . . . . . . . . . . . . . . . . . . . . . 190viD.1.5 Maximum Flow . . . . . . . . . . . . . . . . . . . . . . . 190D.1.6 Minimum Spanning Tree Weight . . . . . . . . . . . . . . 192D.2 Geometric Predicates . . . . . . . . . . . . . . . . . . . . . . . . 193D.2.1 Area of Convex Hulls . . . . . . . . . . . . . . . . . . . . 193D.2.2 Point Containment for Convex Hulls . . . . . . . . . . . . 194D.2.3 Line-Segment Intersection for Convex Hulls . . . . . . . . 195D.2.4 Intersection of Convex Hulls . . . . . . . . . . . . . . . . 196D.3 Pseudo-Boolean Constraints . . . . . . . . . . . . . . . . . . . . 197D.4 CTL Model Checking . . . . . . . . . . . . . . . . . . . . . . . . 198viiList of TablesTable 6.1 Reachability constraints in terrain generation. . . . . . . . . . 79Table 6.2 Runtime results for shortest paths constraints in Diorama. . . . 80Table 6.3 Runtime results for connected components constraints in Dio-rama. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81Table 6.4 Runtime results for maximum flow constraints in Diorama. . . 82Table 6.5 Maze generation using minimum spanning tree constraints. . . 84Table 6.6 Multi-layer escape routing with blind vias. . . . . . . . . . . . 92Table 6.7 Multi-layer escape routing with through-vias. . . . . . . . . . . 93Table 6.8 Data centers with tree topologies. . . . . . . . . . . . . . . . . 101Table 6.9 Data centers with FatTree and BCube Topologies. . . . . . . . 103Table 6.10 Hadoop VDCs allocated on data centers. . . . . . . . . . . . . 106Table 7.1 Art gallery synthesis results. . . . . . . . . . . . . . . . . . . 120Table 8.1 Results on the mutual exclusion example. . . . . . . . . . . . . 140Table 8.2 Results on the mutual exclusion with additional properties . . . 142Table 8.3 Results on the readers-writers instances. . . . . . . . . . . . . 144viiiList of FiguresFigure 2.1 Example of Tseitin transformation . . . . . . . . . . . . . . . 5Figure 2.2 A first order formula and associated Boolean skeleton. . . . . 18Figure 3.1 A finite symbolic graph. . . . . . . . . . . . . . . . . . . . . 27Figure 5.1 A finite symbolic graph and satisfying assignment. . . . . . . 56Figure 5.2 A finite symbolic graph, under assignment. . . . . . . . . . . 60Figure 5.3 Summary of our theory solver implementation of reachability. 61Figure 5.4 Results on reachability constraints. . . . . . . . . . . . . . . . 64Figure 5.5 Summary of our theory solver implementation of acyclicity. . 65Figure 5.6 Results on acyclicity constraints. . . . . . . . . . . . . . . . . 67Figure 5.7 Maximum flow predicate example. . . . . . . . . . . . . . . . 69Figure 5.8 Summary of our theory solver implementation of maximum flow. 71Figure 5.9 Results on maximum flow constraints. . . . . . . . . . . . . . 73Figure 6.1 Generated terrain . . . . . . . . . . . . . . . . . . . . . . . . 78Figure 6.2 Random Mazes . . . . . . . . . . . . . . . . . . . . . . . . . 83Figure 6.3 An example multi-layer escape routing. . . . . . . . . . . . . 85Figure 6.4 Multi-layer escape routing with simultaneous via-placement. . 87Figure 6.5 Symbolic nodes and edges. . . . . . . . . . . . . . . . . . . . 88Figure 6.6 Constraints enforcing different via models. . . . . . . . . . . 90Figure 7.1 Convex hull of a symbolic point set. . . . . . . . . . . . . . . 109Figure 7.2 Under- and over-approximative convex hulls. . . . . . . . . . 109Figure 7.3 Point containment for convex hulls . . . . . . . . . . . . . . . 113ixFigure 7.4 Line-segment intersection for convex hulls . . . . . . . . . . 114Figure 7.5 Intersection of convex hulls . . . . . . . . . . . . . . . . . . 116Figure 7.6 A separating axis between two convex hulls. . . . . . . . . . . 118Figure 7.7 Art gallery synthesis. . . . . . . . . . . . . . . . . . . . . . . 120Figure 8.1 An example CTL formula and Kripke structure. . . . . . . . . 123xAcknowledgmentsAlan and Holger, for giving me both the freedom and the support that I needed. Thestudents I have collaborated with, including Celina, Mike, Nodir, Tobias, Kuba,and many others who have passed through our lab, for being a continuing sourceof inspiration. My parents and siblings, for everything else.And the many shelves worth of science fiction that got me through it all....xiChapter 1IntroductionConstraint solvers have widespread applications in all areas of computer science,and can be found in products ranging from drafting tools [18] to user interfaces [20].Intuitively, constraint solvers search for a solution to a logical formula, with differ-ent constraint solvers using different reasoning methods and supporting differentlogics.One particular family of constraint solvers, Boolean satisfiability (SAT) solvers,has made huge strides in the last two decades. Fast SAT solvers now form the coreengines of many other important algorithms, with applications to AI planning (e.g.,[138, 174]), hardware verification (e.g., [35, 40, 148]), software verification (e.g.,[50, 57]), and even to program synthesis (e.g., [136, 184]).Building on this success, the introduction of a plethora of Satisfiability ModuloTheories (SMT) solvers (see, e.g., [44, 70, 81, 192]) has allowed SAT solvers toefficiently solve problems from many new domains, most notably from arithmetic(e.g., [80, 100, 190]) and data structures (e.g., [29, 173, 191]). In fact, since theirgradual formalization in the 1990’s and 2000’s, SMT solvers — and, in particular,lazy SMT solvers, which delay encoding the formula into Boolean logic — havebeen introduced for dozens of logics, greatly expanding the scope of problems thatSAT solvers can effectively tackle in practice.Unfortunately, many important problem domains, including elements of graphand automata theory, do not yet have dedicated SMT solvers. In Chapter 2, wereview the literature on SAT and SMT solvers, and the challenges involved in de-1Chapter 1. Introductionsigning SMT solvers for previously unsupported theories.We have identified a class of theories for which one can create efficient SMTsolvers with relative ease. These theories, which we term monotonic theories, areintegral to many problem domains with major real-world applications in computernetworks, computer-aided design, and protocol synthesis. In Chapter 3, we for-mally define what constitutes a monotonic theory, and in Chapter 4 we describethe SAT Modulo Monotonic Theories (SMMT) framework, a comprehensive set oftechniques for building efficient lazy SMT solvers for monotonic theories. (Furtherpractical implementation details can be found in Appendix A.)As we will show, important properties from graph theory (Chapters 5, 6), ge-ometry (Chapter 7), and automata theory (Chapter 8), among many others, canbe modeled as monotonic theories and solved efficiently in practice. Using theSMMT framework, we have implemented a lazy SMT solver (MONOSAT) sup-porting these theories, with greatly improved performance over a range of impor-tant applications, from circuit layout and data center management to protocol syn-thesis — and even to video game design.We do not claim that solvers built using our techniques are always the bestapproach to solving monotonic theories — to the contrary, we can demonstrate atleast some cases where there exist dedicated solvers for monotonic theories thatout-perform our generic approach.1 However, this thesis does make three centralclaims:1. A wide range of important, useful monotonic theories exist.2. Many of these theories can be better solved using the constraint solver de-scribed in this thesis than by any previous constraint solver.3. The SAT Modulo Monotonic Theories framework described in this thesiscan be used to build efficient constraint solvers.1Specifically, in Chapter 5 we discuss some applications in which our graph theory solver is out-performed by other solvers. More generally, many arithmetic theories can be modeled as monotonictheories, for which there already exist highly effective SMT solvers against which our approach isnot competitive. Examples include linear arithmetic, difference logic, and, as we will discuss later,pseudo-Boolean constraints.2Chapter 2BackgroundBefore introducing the SAT Modulo Monotonic Theories (SMMT) framework inChapters 3 and 4, we review several areas of relevant background. First, we re-view the theoretical background of Boolean satisfiability, and then survey the mostrelevant elements of modern conflict-driven clause-learning (CDCL) SAT solvers.Then we review the aspects of many-sorted first-order logic that form the theoreti-cal background of Satisfiability Modulo Theories (SMT) solvers, and describe thebasic components of lazy SMT solvers.2.1 Boolean SatisfiabilityBoolean satisfiability (SAT) is the problem of determining, for a given proposi-tional formula, whether or not there exists a truth assignment to the Boolean vari-ables of that formula such that the formula evaluates to TRUE (we will define thismore precisely shortly). SAT is a fundamental decision problem in computer sci-ence, and the first to be proven NP-Complete [61]. In addition to its importanttheoretical properties, many real-world problems can be efficiently modeled asSAT problems, making solving Boolean satisfiability in practice an important prob-lem to tackle. Surprisingly, following a long period of stagnation in SAT solvers(from the 1960s to the early 1990s), great strides on several different fronts havenow been made in solving SAT problems efficiently, allowing for wide classes ofinstances to be solved quickly enough to be useful in practice. Major advance-3Chapter 2. Backgroundments in SAT solver technology include the introduction of WALKSAT [180] andother stochastic local search solvers in the early 1990s, and the development ofGRASP [146], CHAFF [151] and subsequent CDCL solvers in the late 1990’s andfollowing decade. We will review these in Section 2.2.A Boolean propositional formula is a formula over Boolean variables that, forany given assignment to those variables, evaluates to either TRUE or FALSE. For-mally, a propositional formula φ can be a literal, which is either a Boolean vari-able v or its negation ¬v; or it may be built from other propositional formulasusing, e.g. one of the standard binary logical operators ¬φ0,(φ0 ∧ φ1), (φ0 ∨ φ1),(φ0 =⇒ φ1), . . ., with φi a Boolean formula. A truth assignmentM : v 7→ {T,F}for φ is a map from the Boolean variables of φ to {TRUE,FALSE}. It is also oftenconvenient to treat a truth assignment as a set of literals, containing the literal v iffv 7→ T ∈M, and the literal ¬v iff v 7→ F ∈M. A truth assignmentM is a completetruth assignment for a formula φ if it maps all variables in the formula to assign-ments; otherwise it is a partial truth assignment. A formula φ is satisfiable if thereexists one or more complete truth assignments to the Boolean variables of φ suchthat the formula evaluates to TRUE; otherwise it is unsatisfiable. A propositionalformula is valid or tautological if it evaluates to TRUE under all possible completeassignments;1 a propositional formula is invalid if it evaluates to FALSE under atleast one assignment.A complete truth assignment M that satisfies the formula φ may be called amodel of φ , written M |= φ (especially in the context of first order logic); or itmay be called a witness for the satisfiability of φ (in the context of SAT as anNP-complete problem). A partial truth assignment M is an implicant of φ , writtenM =⇒ φ , if all possible completions of the partial assignmentM into a completeassignment must satisfy φ ; a partial assignment M is a prime implicant of φ ifM =⇒ φ and no proper subset ofM is itself an implicant of φ .2In most treatments, Boolean satisfiability is assumed to be applied only topropositional formulas in Conjunctive Normal Form (CNF), a restricted subset of1In purely propositional logic, there is no distinction between valid formulas and tautologies;however, in most presentations of first order logic there is a distinction between the two concepts.2There is also a notion of minimal models, typically defined as complete, satisfying truth assign-ments of φ in which a (locally) minimal number of variables are assigned TRUE.4Chapter 2. Background(a∧b)∨ c (¬a∨¬b∨x)∧(a∨¬x)∧(b∨¬x)∧(x∨c)Figure 2.1: Left: a formula not in CNF. Right: an equisatisfiable formula inCNF, produced by the Tseitin transformation. The new formula hasintroduced an auxiliary variable, x, which is constrained to be equal tothe subformula (a∧b).propositional formulas. CNF formulas consist of a conjunction of clauses, whereeach clause is a disjunction of one or more Boolean literals. In the literature, a for-mula in conjunctive normal form is often assumed to be provided in the form of aset of clauses {c0,c1,c2, . . .}, as opposed to as a formula. In this thesis, we use thetwo forms interchangeably (and refer to both as ‘formulas’), but it should always beclear from context which form we mean. Similarly, a clause may itself be providedas a set of literals, rather than as a logical disjunction of literals. For example, weuse the set interpretation of a CNF formula in the following paragraph.If a formula φ in conjunctive normal form is unsatisfiable, and φ ′ is also unsat-isfiable, and φ ′ ⊆ φ , then φ ′ is an unsatisfiable core of φ . If φ ′ is unsatisfiable, andevery proper subset of φ ′ is satisfiable, then φ ′ is a minimal unsatisfiable core.3Any Boolean propositional formula φ can be transformed directly into a logi-cally equivalent formula φ ′ in conjunctive normal form through the application ofDe Morgan’s law and similar logical identities, but doing so may require a numberof clauses exponential in the size of the original formula. Alternatively, one canapply transformations such as the Tseitin transformation [197] to construct a for-mula φ ′ in conjunctive normal form (see Figure 2.1). The Tseitin transformationproduces a CNF formula φ ′ with a number of clauses linear in the size of φ , at thecost of also introducing an extra auxiliary variable for every binary logical oper-ator in φ . A formula created through the Tseitin transformation is equisatisfiableto the original, meaning that it is satisfiable if and only if the original formula wassatisfiable (but a formula created this way might not be logically equivalent to theoriginal formula, as it may have new variables not present in the original formula).In different contexts, Boolean satisfiability solvers may be assumed to be restricted3There is a related notion of a globally minimum unsatisfiable core, which is a minimal unsatis-fiable core of φ that has the smallest possible number of clauses of all unsatisfiable cores of φ .5Chapter 2. Backgroundto operating on CNF formulas, or they may be allowed to take non-normal-forminputs. As the Tseitin transformation (and similar transformations, such as thePlaisted-Greenbaum [167] transformation) can be applied in linear time, in manycontexts it is simply left unspecified whether a Boolean satisfiability solver is re-stricted to CNF or not, with the assumption that a non-normal form formula can beimplicitly transformed into CNF if required.4Boolean satisfiability is an NP-complete problem, and all known algorithmsfor solving it require at least exponential time in the worst case. Despite this, aswe will discuss in the next section, SAT solvers often can solve even very largeinstances with millions of clauses and variables quite easily in practice.2.2 Boolean Satisfiability SolversSAT solvers have a long history of development, but at present, successful solverscan be roughly divided into complete solvers (which can both find a satisfyingsolution to a given formula, if such a solution exists, or prove that no such solutionexists) and incomplete solvers, which are able to find satisfying solutions if theyexist, but cannot prove that a formula is unsatisfiable.Incomplete solvers are primarily stochastic local search (SLS) solvers, whichare only guaranteed to terminate on satisfiable instances, originally introduced inthe WALKSAT [180] solver. Most complete SAT solvers can trace their develop-ment back to the resolution rule([66]), and the Davis-Putnam-Logemann-Loveland(DPLL) algorithm [67]. At the present time, DPLL-based complete solvers havediverged into two main branches: look-ahead solvers (e.g., [97]), which performwell on unsatisfiable ‘random’ (phase transition) instances (e.g., [76]), or on diffi-cult ‘crafted’ instances (e.g., [118]); and Conflict-Driven Clause-Learning (CDCL)solvers, which have their roots in the GRASP [146] and CHAFF [151] solvers(along with many other contributing improvements over the years), which performwell on so-called application, or industrial, instances. These divisions have solidi-fied in the last 15 years: in all recent SAT competitions [134], CDCL solvers havewon in the applications/industrial category, look-ahead solvers have won in the4There are also some SAT solvers specifically designed to work on non-normal-form inputs(e.g., [163, 204]), with applications to combinatorial equivalence checking.6Chapter 2. Backgroundcrafted and unsatisfiable/mixed random categories, while stochastic local searchsolvers have won in the satisfiable random instances category.In this thesis, we are primarily interested in CDCL solvers, and their rela-tion to SMT solvers. Other branches of development exist (such as Sta˚lmarck’smethod [188], and Binary Decision Diagrams [45]), but although BDDs in particu-lar are competitive with SAT solvers in some applications (such as certain problemsarising in formal verification [168]), neither of these techniques has been success-ful in any recent SAT competitions. There have been solvers which combine someof the above mentioned SAT solver strategies, such as Cube and Conquer [119],which combines elements of CDCL and lookahead solvers, or SATHYS, whichcombines techniques from SLS and CDCL solvers [17], with mixed success. AsCDCL solvers developed out of the DPLL algorithm, we will briefly review DPLLbefore moving on to CDCL.2.2.1 DPLL SAT SolversThe Davis-Putnam-Logemann-Loveland (DPLL) algorithm is shown in Algorithm1. DPLL combines recursive case-splitting (over the space of truth assignments tothe variables in the CNF formula) with a particularly effective search-space pruningstrategy and deduction technique, unit propagation.5DPLL takes as input a propositional formula φ in conjunctive normal form,and repeatedly applies the unit propagation rule (and optionally the pure-literalelimination rule) until a fixed point is reached. After applying unit propagation, ifall clauses have been satisfied (and removed from φ ), DPLL returns TRUE. If anyclause has been violated (represented as an empty clause), DPLL returns FALSE.Otherwise, DPLL picks an unassigned Boolean variable v from φ , assigning v toeither TRUE or FALSE, and recursively applies DPLL to both cases. These freeassignments are called decisions.The unit propagation rule takes a formula φ in conjunctive normal form, andchecks if there exists a clause in φ that contains only a single literal l (a unit clause).Unit propagation then removes any clauses in φ containing l, and removes fromeach remaining clause any occurrence of ¬l. After this process, φ does not contain5Unit propagation is also referred to as Boolean Constraint Propagation; in the constraint pro-gramming literature, it would be considered an example of the arc consistency rule [199].7Chapter 2. BackgroundAlgorithm 1 The Davis-Putnam-Logemann-Loveland (DPLL) algorithm forBoolean satisfiability. Takes as input a set of clauses φ representing a formulain conjunctive normal form, and returns TRUE iff φ is satisfiable.function DPLL(φ )while φ contains unit clause {l} doApply unit propagationφ ← PROPAGATE(φ , l)if φ = {} thenAll clauses satisfiedreturn TRUEelse if {} ∈ φ thenφ contains empty clause, is unsatisfiablereturn FALSESelect a variable v ∈ vars(φ)return DPLL(φ ∪{v}) or DPLL(φ ∪{¬v})function PROPAGATE(φ , l)φ ′←{}for clause c ∈ φ doif l ∈ c thenDo not add c to φ ′else if ¬l ∈ c thenRemove literal ¬l from cφ ′← φ ′∪ (c/{¬l})elseφ ′← φ ′∪ creturn φ ′any occurrence of l or ¬l.The strength of the unit propagation rule comes from two sources. First, it canbe implemented very efficiently in practice (using high performance data structuressuch as two-watched-literals [151], which have supplanted earlier techniques suchas head/tail lists [216] or unassigned variable counters [146]). Secondly, each timea new unit clause is discovered, the unit-propagation rule can be applied again,potentially making additional deductions. This can be (and is) applied repeatedlyuntil a fixed point is reached and no new deductions can be made. In practice, for8Chapter 2. Backgroundmany common instances, the unit propagation rule ends up very cheaply makinglong chains of deductions, allowing it to eliminate large parts of the search space.Because it makes many deductions, and because it does so inexpensively, the unitpropagation rule is very hard to augment without unduly slowing down the solver.Other deduction rules have been proposed; in particular, the original DPLL algo-rithm also applied pure literal elimination (in which any variable occurring in onlyone polarity throughout the CNF is assigned that polarity), however, modern SATsolvers almost always apply the unit propagation rule exclusively.6In practice, the decision of which variable to pick, and whether to test the posi-tive assignment (v) or the negative assignment (¬v), has an enormous impact on theperformance of the algorithm, and so careful implementations will put significanteffort into good decision heuristics. Although many heuristics have been proposed,most current solvers use the VSIDS heuristic (introduced in Chaff), with a minor-ity using the Berkmin [109] heuristic. Both of these heuristics choose which vari-able to split on next, but do not select which polarity to try first; common polarityselection heuristics include Jeroslow-Wang [135], phase-learning [166], choosingrandomly, or simply always choosing one of TRUE or FALSE first.2.2.2 CDCL SAT SolversConflict-Driven Clause Learning (CDCL) SAT solvers consist of a number of ex-tensions and improvements to DPLL. In addition to the improved decision heuris-tics and unit-propagation data structures discussed in the previous section, a few ofthe key improvements include:1. Clause learning2. Non-chronological backtracking3. Learned-clause discarding heuristics4. Pre-processing5. Restarts6Note that this generalization applies mostly to CDCL SAT solvers. In contrast, solvers thatextend the CDCL framework to other logics, and in particular SMT solvers, sometimes do applyadditional deduction rules.9Chapter 2. BackgroundAlgorithm 2 A simplified Conflict-Driven Clause-Learning (CDCL) solver. Takesas input a set of clauses φ representing a formula in conjunctive normal form, andreturns TRUE iff φ is satisfiable. PROPAGATE applies unit propagation to a partialassignment, adding any unit literals to the assignment, and potentially finding aconflict clause. ANALYZE and BACKTRACK perform clause learning and non-chronological backtracking, as described in the text. SIMPLIFY uses heuristics toidentify and discard some of the learned clauses. Lines marked occasionally areonly applied when heuristic conditions are met.function SOLVE(φ )level← 0assign←{}loopif PROPAGATE(assign) returns a conflict thenif level = 0 thenreturn FALSEc,backtrackLevel← ANALYZE(conflict)φ ← φ ∪ clevel← BACKTRACK(backtrackLevel,assign)elseif All variables are assigned thenreturn TRUEif level > 0 then occasionally restart the solver to level 0level← BACKTRACK(0,assign)if level = 0 thenOccasionally discard some learned or redundant clausesφ ← SIMPLIFY(φ)Select unassigned literal llevel← level +1assign[l]← TRUEOf these, the first two are the most relevant to this thesis, so we will restrict ourattention to those. Clause learning and non-chronological backtracking (sometimes‘back-jumping’) were originally developed independently, but in current solversthey are tightly integrated techniques. In Algorithm 2, we describe a simplifiedCDCL loop. Unlike our recursive presentation of DPLL, this presentation is in astateful, iterative form, in which PROPAGATE operates on an assignment structure,rather than removing literals and clauses from φ .10Chapter 2. BackgroundClause learning, introduced in the SAT solver GRASP [146], is a form of ‘no-good’ learning [74], a common technique in CSP solvers for pruning unsatisfiableparts of the search space. When a conflict occurs in a DPLL solver (that is, whenφ contains an empty clause), the solver simply backtracks past the most recentdecision and then continues exploring the search tree. In contrast, when a clause-learning SAT solver encounters a conflict, it derives a learned clause, c, not presentin φ , that serves to ‘explain’ (in a precise sense) the conflict. This learned clauseis then appended to φ . Specifically, the learned clause is such that if the clausehad been present in the formula, unit propagation would have prevented (at least)one of the decisions made by the solver that resulted in the conflict. By addingthis new clause to φ , the solver is prevented in the future from making the sameset of decisions that led to the current conflict. In this sense, learned clauses canbe seen as augmenting or strengthening the unit propagation rule. Learned clausesare sometimes called redundant clauses, because they can be added to or removedfrom the formula without removing any solutions from the formula.After a conflict, there may be many choices for the learned clause. ModernCDCL solvers learn asserting clauses, which have two additional properties: 1) allliterals in the clause are false in the current assignment, and 2) exactly one literalin the clause was assigned at the current decision level. Although many possiblestrategies for learning clauses have been proposed (see, e.g., [32, 146]), by farthe dominant strategy is the first unique implication point (1-UIP) method [218],sometimes augmented with post-processing to further reduce the size of the clause([84, 186]).Each time a conflict occurs, the solver learns a new clause, eliminating at leastone search path (the one that led to the conflict). In practice, clauses are oftenmuch smaller than the number of decisions made by the solver, and hence eachclause may eliminate many search paths, including ones that have not yet beenexplored by the solver. Eventually, this process will either force the solver to finda satisfying assignment, or the solver will encounter a conflict in which no newclause can be learned (because no decisions have yet been made by the solver).In this second case, the formula is unsatisfiable, and the solver will terminate. Aside benefit of clause learning is that as the clauses serve to implicitly prevent thesolver from exploring previously searched paths, the solver is freed to explore paths11Chapter 2. Backgroundin any order without having to explicitly keep track of previously explored paths.This allows CDCL solvers to apply restarts cheaply, without losing the progressthey have already made.In a DPLL solver, when a conflict is encountered, the solver backtracks upthe search tree until it reaches the most recent decision where only one branchof the search space has been explored, and then continues down the unexploredbranch. In a non-chronological backtracking solver, the solver may backtrack morethan one level when a conflict occurs. Non-chronological backtracking, like clauselearning, has its roots in CSP solvers (e.g. [169]), and was first applied to SAT inthe SAT solver REL SAT [32], and subsequently combined with clause learningin GRASP. When a clause is learned following a conflict, modern CDCL solversbacktrack to the level of the second highest literal in the clause. So long as thelearned clause was an asserting clause (as discussed above), the newly learnedclause will then be unit at the current decision level (after having backtracked), andthis will trigger unit propagation. The solver can then continue the solving processas normal. Intuitively, the solver backtracks to the lowest level at which, if thelearned clause had been in φ from the start, unit propagation would have impliedan additional literal.Further components that are also important in CDCL solvers include heuristicsand policies for removing learned clauses that are no longer useful [14], restartpolicies (see [15] for an up-to-date discussion), and pre-processing [83], to namejust a few. However, the elements already described in this section are the onesmost relevant to this thesis. For further details on the history and implementationof CDCL solvers, we refer readers to [217].12Chapter 2. Background2.3 Many-Sorted First Order Logic & SatisfiabilityModulo TheoriesSatisfiability Modulo Theories (SMT) solvers extend SAT solvers to support logicsor operations that cannot be efficiently or conveniently expressed in pure propo-sitional logic. Although many different types of SMT solvers have been devel-oped, sometimes with incompatible notions of what a theory is, the majority ofSMT solvers have consolidated into a mostly standardized framework built aroundmany-sorted first order logic. We will briefly review the necessary concepts here,before delving into the implementation of Satisfiability Modulo Theories solversin the next section.A many-sorted first order formula φ differs from a Boolean propositional for-mula in two ways. First, instead of just being restricted to Boolean variables, eachvariable in φ is associated with a sort (analogous to the ‘types’ of variables in aprogramming language). For example, in the formula (a ≤ b)∨ (y+ z = 1.5), thevariables a and b might be of sort integer, while variables y,z might be of sort ra-tional (we’ll discuss the mathematical operators in this formula shortly). Formally,a sort σ is a logical symbol, whose interpretation is a (possibly infinite) set ofconstants, with each term of sort σ in the formula having a value from that set.7The second difference is that a first order formula may contain function andpredicate symbols. In many-sorted first order logic, function symbols have typedsignatures, mapping a fixed number of arguments, each of a specified sort, toan output of a specified sort. For example, the function min(integer, integer) 7→integer takes two arguments of sort integer and returns an integer; the functionquotient(integer, integer) 7→ rational takes two integers and returns a rational. Anargument to a function may either be a variable of the appropriate sort, or a func-tion whose output is of the appropriate sort. A term t is either a variable v or aninstance of a function, f (t0, t1, . . .), where each argument ti is a term. The numberof arguments a function takes is called the ‘arity’ of that function; functions of arity7In principle, an interpretation for a formula may select any domain of constants for each sort,so long as it is a superset of the values assigned to the terms of that sort, in the same way that non-standard interpretations can be selected, for example, for the addition operator. However, we willonly consider the case where the interpretations of each sort are fixed in advance; i.e., the sort Zmaps to the set of integer constants in the expected way.13Chapter 2. Background0 are called constants or constant symbols.Functions that map to Boolean variables are predicates. Although predicatesare functions, in first order logic predicates are treated a little differently than otherfunctions, and are associated with special terminology. Intuitively, the reason forthis special treatment of Boolean-valued functions is that they form the bridgebetween propositional (Boolean) logic and first-order logic. A predicate p takesa fixed number of arguments p(t0, t1, . . .), where each ti is a term, and outputs aBoolean. By convention, it is common to write certain predicates, such as equalityand comparison operators, in infix notation. For example, in the formula a ≤ b,where the variables a and b are of sort integer, the ≤ operation is a predicate, asa≤ b evaluates to a truth value.A predicate with arity 0 is called a propositional symbol or a constant. Eachindividual instance of a predicate p in a formula is called an atom of p. For exam-ple, TRUE and FALSE are Boolean constants, and in the formula p∨q, p and q arepropositional symbols. TRUE, FALSE, p, and q are all examples of arity-0 predi-cates. In many presentations of first order logic, an atom may also be any Booleanvariable that is not associated with a predicate; however, to avoid confusion we willalways refer to non-predicate variables as simply ‘variables’ in this thesis. Atomsalways have Boolean sorts; an atom is a term (just as a predicate is a function), butnot all terms are atoms. If a is an atom, then a and ¬a are literals, but ¬a is not anatom.A many-sorted first order formula φ is either a Boolean variable v, a predi-cate atom p(t0, t1, . . .), or a propositional logic operator applied to another formula:¬φ0,(φ0∧φ1), (φ0∨φ1), (φ0 =⇒ φ1), . . . Finally, the variables of a formula φ maybe existentially or universally quantified, or they may be free. Most SMT solversonly support quantifier-free formulas, in which each variable is implicitly treatedas existentially quantified at the outermost level of the formula.Predicate and function symbols in a formula may either be interpreted, or unin-terpreted. An uninterpreted function or predicate term has no associated semantics.For example, in an uninterpreted formula, the addition symbol ‘+’ is not necessar-ily treated as arithmetic addition. The only rule for uninterpreted functions andpredicates is that any two applications of the same function symbol with equalarguments must return the same value.14Chapter 2. BackgroundIn first order logic, the propositional logic operators (¬,∨,∧. etc.) are alwaysinterpreted. However, by default, other function and predicate symbols are typi-cally assumed to be uninterpreted. Many treatments of first order logic also assumethat the equality predicate, ‘=’, is always interpreted; this is sometimes called firstorder logic with equality.A structureM for a formula φ is an assignment of a value of the appropriatesort to each variable, predicate atom, and function term. Unlike an assignmentin propositional logic, a structure must also supply an assignment of a concretefunction (of the appropriate arity and sorts) to each function and predicate symbolin φ . For example, in the formula z= func(x,y), both {x 7→ 1,y 7→ 2,z 7→ 1,‘func’7→min,‘=’ 7→ equality} and {x 7→ 1,y 7→ 2,z 7→ 2,‘func’ 7→ max,‘=’ 7→ equality} areboth structures for φ .Given a structureM, a formula φ evaluates to either TRUE or FALSE. A struc-ture is a complete structure if it provides an assignment to every variable, atom, andterm (and every uninterpreted function symbol); otherwise it is a partial structure.A (partial) structure M satisfies φ (written M |= φ ) if φ must evaluate to TRUEin every possible completion of the assignment inM. A formula φ is satisfiableif there exists a structureM that satisfies φ , and unsatisfiable if no such structureexists. A structure that satisfies formula φ may be called a model, an interpretation,or a solution for φ .In the context of Satisfiability Modulo Theories solvers, we are primarily in-terested in interpreted first order logic.8 Those interpretations are supplied by the-ories. A theory T is a (possibly infinite) set of logical formulas, the conjunction ofwhich must hold in any satisfying assignment of that theory. A structureM is saidto satisfy a formula φ modulo theory T , writtenM =⇒Tφ , iffM satisfies all theformulas in T ∪{φ}.The formulas in a theory serve to constrain the possible satisfying assignmentsto (some of) the function and predicate symbols. For example, a theory of integeraddition may contain the infinite set of formulas (0+1= 1,1+0= 1,1+1= 2, . . .)defining the binary addition function. A structureM satisfies a formula φ modulotheory T if and only if M satisfies (T ∪ φ). In other words, the structure must8With the specific exception of SMT solvers for the theory of uninterpreted functions, SMTsolvers always operate on interpreted formulas.15Chapter 2. Backgroundsimultaneously satisfy both the original formula φ , and all the (possibly infinite)formulas that make up T .Returning to the first order formula (a ≤ b)∨ (y+ z = 1.5) that we saw ear-lier, ≤ and = are predicates, while + is a function. If a,b are integers, and y,zare rationals, then a theory of linear integer arithmetic may constrain the satisfi-able interpretations of the ≤ function so that in all satisfying models, the predicatebehaves as the mathematical operator would be expected to. Similarly, a theoryof linear rational arithmetic might specify the meaning of the ‘+’ predicate, whilethe interpretation of the equality predicate may be assumed to be supplied by animplicit theory of equality (or it might be explicitly provided by the theory of linearrational arithmetic).A theory is a (possibly infinite) set of logical formulas. The signature of atheory, Σ, consists of three things:1. The sorts occurring in the formulas in that theory,2. the predicate symbols occurring in the formulas of that theory, and3. the function symbols occurring in the formulas of that theory.In most treatments, the Boolean sort, and the standard propositional logic opera-tors (¬,∨,∧,etc.) are implicitly available in every theory, without being counted asmembers of their signature. Similarly, the equality predicate is typically also im-plicitly available in all theories, while being excluded from their signatures. Con-stants (0-arity functions) are included in the signature of a theory.For example, the signature of the theory of linear integer arithmetic (LIA) con-sists of:1. Sorts = {Z}2. Predicates = {<,≤,≥,>}3. Functions ={+,−,0,1,−1,2,−2 . . .}A formula φ in which all sorts, predicates, and function symbols appearing in φcan be found in the signature of theory T (aside from Booleans, propositional logic16Chapter 2. Backgroundoperators, and the equality predicate) is said to be written in the language of T . Anexample of a formula in the language of the theory of linear integer arithmetic is:(x > y)∨ ((x+ y = 2)∧ (x <−1))In this section we have presented one internally self-consistent description ofmany-sorted first order logic, covering the most relevant elements to SatisfiabilityModulo Theories solvers. However, in the literature, there are many variations ofthese concepts, both in terminology and in actual semantics. For an introduction tomany-sorted first order logic as it applies to SMT solvers, we refer readers to [71].2.4 Satisfiability Modulo Theories SolversSatisfiability Modulo Theories solvers extend Boolean satisfiability solvers so thatthey can solve many-sorted first order logic formulas written in the language of oneor more theories, with different SMT solvers supporting different theories.Historically, many of the first SMT solvers were ‘eager’ solvers, which converta formula into a purely propositional CNF formula, and then subsequently applyan unmodified SAT solver to that formula; an example of an early eager SMTsolvers was UCLID [46]. For some theories, eager SMT solvers are still stateof the art (in particular, the bitvector solver STP [100] is an example of a state-of-the-art eager solver; another example would be MINISAT+, a pseudo-Booleanconstraint solver, although pseudo-Boolean constraint solvers are not traditionallygrouped together with SMT solvers). However, for many theories (especially thoseinvolving arithmetic), eager encodings require exponential space.The majority of current high-performance SMT solvers are lazy SMT solvers,which attempt to avoid or delay encoding theory constraints into a propositionalformula. While the predecessors of lazy SMT solvers go back to the mid 1990s [27,107], they were gradually formalized into a coherent SMT framework in the earlyto mid 2000s [9, 16, 31, 73, 101, 161, 192].The key idea behind lazy SMT solvers is to have the solver operate on an ab-stracted, Boolean skeleton of the original first order formula, in which each theoryatom has been swapped out for a fresh Boolean literal (see Figure 2.2). The firstlazy SMT solvers were offline [31, 72, 73]. Offline SMT solvers combine an un-17Chapter 2. Background((x > 1)∧ (2x < 5))∨¬(y = 0) (a∧b)∨¬cFigure 2.2: Left: A first order formula in the theory of linear real arithmetic,with three theory atoms. Right: Boolean skeleton of the same formula.Boolean variables a,b,c replace atoms (x > 1),(2x < 5),(y = 0).modified SAT solver with a specialized theory solver for each of the theories inthe formula. Offline lazy SMT solvers still treat the SAT solver as a black box,but unlike an eager solver, they encode the propositional formula incrementally,repeatedly solving and refining the abstracted Boolean version of the first orderformula. Initially, a SAT solver solves the Boolean skeleton of the formula, whichproduces a truth assignment to the fresh literals introduced for each atom.The solver passes the corresponding assignment to the original theory atomsto a specialized theory solver, which checks if there exists a satisfying model forthe theory under that assignment to those theory atoms. If there is a satisfyingassignment in the theory, then the solver terminates, returning SAT. If the formulais unsatisfiable in the theory under that assignment to the atoms, then the theorysolver derives a learned clause to add to the Boolean skeleton which blocks thatassignment, and the process repeats. We will describe theory solvers in more detailshortly.Although there are still some state-of-the-art offline SMT solvers (the bitvectorsolver Boolector [42] is an example), in most cases, offline solvers perform verypoorly [101], producing many spurious solutions to the propositional formula thatare trivially false in the theory solver. The vast majority [179] of modern SMTsolvers are instead online lazy SMT solvers, formalized in [101]. Unlike an eageror offline SMT solver, an online SMT solver is tightly integrated into a modi-fied SAT solver. For example, the Abstract DPLL framework explicitly formalizesCDCL solvers as state machines [161] for the purpose of reasoning about the cor-rectness of SMT solvers. While there has been work extending stochastic localsearch SAT solvers into SMT solvers (e.g., [98, 112, 158]), by far the dominantapproach at this time is to combine CDCL solvers with specialized reasoning pro-cedures. For this reason, we will phrase our discussion in this section explicitly interms of CDCL solvers.18Chapter 2. BackgroundAn online lazy SMT solver consists of two components. The first, as with of-fline lazy solvers, is a specialized theory solver (or ‘T-solver’). The second compo-nent is a modified CDCL solver, which adds a number of hooks to interact with thetheory solver. As before, the SAT solver operates on an abstracted Boolean skele-ton of the first order formula, φ ′. Unlike in offline solvers, an online solver does notwait to generate a complete satisfying assignment to the Boolean formula beforechecking whether the corresponding assignment to the theory atoms is satisfiablein the theory solver, but instead calls the theory solver eagerly, as assignments totheory literals are made in the SAT solver.Theory solvers may support some or all of the following methods:1. T-Propagate (M):The most basic version of T-Propagate (also called T-Deduce) takes a com-plete truth assignmentM to the theory atoms (for a single theory), returningTRUE ifM is satisfiable in the theory, and FALSE otherwise.In contrast to offline solvers, online lazy SMT solvers (typically) make callsto T-Propagate eagerly [101], as assignments are being built in the CDCLsolver, rather than waiting for a complete satisfying assignment to be gen-erated. In this case, T-Propagate takes a partial, rather than a complete,assignment to the theory atoms. When operating on a partial assignment,T-Propagate makes deductions on a best-effort basis: if it can prove that thepartial assignment is unsatisfiable, then it returns FALSE, and otherwise itreturns TRUE. By calling T-Propagate eagerly, assignments that may sat-isfy the abstract Boolean skeleton, but which are unsatisfiable in the theorysolver, can be pruned from the CDCL solvers search space. This techniqueis sometimes called early pruning.All theory solvers must support at least this basic functionality, but sometheory solvers can do more. If M is a partial assignment, efficient theorysolvers may also be able to deduce assignments for some of the unassignedatoms. These are truth assignments to theory literals l that must hold in allsatisfying completions ofM, writtenM =⇒Tl. Along with early pruning,deducing unassigned theory literals may greatly prune the search space ofthe solver, avoiding many trivially unsatisfiable solutions.19Chapter 2. BackgroundAs with unit propagation, the CDCL solver typically continues applying unitpropagation and theory propagation until no further deductions can be made(or the assignment is found to be UNSAT). However, there are many possiblevariations on the exact implementation of theory propagation. For manytheories of interest, testing satisfiability may be very expensive; deducingassignments for unassigned atoms may be an additional cost on top of that,in some cases a prohibitively expensive one. Many theory solvers can onlycheaply make deductions of a subset of the possible forced assignments.Theory solvers that make all possible deductions of unassigned literals arecalled deduction complete.There are many possible optimizations that have been explored in the lit-erature to try to mitigate the cost of expensive theory propagation calls. Asmall selection include: only calling theory propagate when one or moretheory literals have been assigned [108]; pure literal filtering [16, 39], whichcan remove some redundant theory literals from M; or only calling the-ory propagate every k unit propagation calls, rather than for every singlecall [10]. A theory solver that does not directly detect implied literals canstill be used to deduce such literals by testing, for each unassigned literall, whether M∪{¬l} is unsatisfiable. This technique is known as theoryplunging[77], however, it is rarely used in practice due to its cost.2. T-Analyze (M):When an assignment to the theory atoms M is unsatisfiable in the theory,then an important task for a theory solver is to find a (possibly minimal)subset of M that is sufficient to be unsatisfiable. This unsatisfiable subsetmay be called a conflict set or justification set or infeasible set; its negation (aclause of theory literals, at least one of which must be true in any satisfyingmodel) may be called a theory lemma or a learned clause.T-Analyze may always return the entire M as a conflict set; this is knownas the naı¨ve conflict set. Unfortunately, the naı¨ve conflict set is typically avery poor choice, as it blocks only the exact assignment M to the theoryatoms. Conversely, it is also possible to find a (locally) minimal conflict setthrough a greedy search, repeatedly dropping literals fromM and checking20Chapter 2. Backgroundif it is still unsatisfiable. However, in addition to not guaranteeing a globallyminimal conflict set, such a method is typically prohibitively expensive, es-pecially if the theory propagate method is expensive. Finding small conflictsets quickly is a key challenge for theory solvers.3. T-Backtrack(level):In many cases, efficient theory solvers maintain incremental data structures,so that as T-Propagate (M) is repeatedly invoked, redundant computationin the theory solver can be avoided. T-Backtrack is called when the CDCLsolver backtracks, to keep those incremental data structures in sync with theCDCL solver’s assignments.4. T-Decide(M):Some theory solvers are able to heuristically suggest unassigned theory liter-als as decisions to the SAT solver. If there are multiple theory solvers capableof suggesting decisions, the SAT solver may need to have a way of choosingbetween them.5. T-Model(M):Given a theory-satisfiable assignment M to the theory atoms, this methodextends M into a satisfying model to each of the (non-Boolean) variablesand function terms in the formula. Generating a concrete model may requirea significant amount of additional work on top of simply checking the sat-isfiability of a Boolean assignment M, and so may be separated off into aseparate method to be called when the SMT solver finds a satisfying solu-tion. In some solvers (such as Z3 [70]), models can be generated on demandfor specified subsets of the theory variables.Writing an efficient lazy SMT solver involves finding a balance between moreexpensive deduction capabilities in the theory solver and faster searches in theCDCL solver; finding the right balance is difficult and may depend not only onthe implementation of the theory solver but also on the instances being solved. Fora more complete survey of some of the possible optimizations to CDCL solvers forintegration with lazy theory solvers, we refer readers to Chapter 6 of [179].21Chapter 2. Background2.5 Related WorkIn the following chapters, we will introduce the main subject of this thesis: a gen-eral framework for building efficient theory solvers for a special class of theories,which we call monotonic theories. Several other works in the literature have alsointroduced high-level, generic frameworks for building SMT solvers. There arealso previous constraint solvers which exploit monotonicity. Here, we briefly sur-vey some of the most closely related works. Additionally, the applications andspecific theories we consider later in this thesis (Chapters 5, 6, 7, and 8) also havetheir own, subject-specific literature reviews, and we describe related work in therelevant chapters.There have been a number of works proposing generic SMT frameworks thatare not specialized for any particular class of first order theories. The best knownof these are the DPLL(T) and Abstract DPLL frameworks [101, 161]. Addition-ally, an approach based on extending simplex into a generic SMT solver methodis described in [111]. These frameworks have in common that they are not spe-cific to the class of theories being supported by the SMT solver; the techniques de-scribed in them apply generically to all (quantifier free, first order) theories. There-fore, the work in this thesis could be considered to be an instance of the DPLL(T)framework. However, these high-level frameworks do not articulate any conceptof monotonic theories, and provide no guidance for exploiting their properties.As we will see in Chapters 5 and 6, some of our most successful examples ofmonotonic theories are graph theoretic. Two recent works have proposed genericframeworks for extending SMT solvers with support for graph properties:The first of the two, [220], proposed MathCheck, which extends an SMT solverby combining it with an external symbolic math library (such as SAGE [189]). Thisallows the SMT solver great flexibility in giving it access to a wide set of operationsthat can be specified by users in the language of the symbolic math library. Theyfound that they were able to prove bounded cases of several discrete mathemati-cal theorems using this approach. Unfortunately, their approach is fundamentallyan off-line, lazy SMT integration, and as such represents a extremal point in theexpressiveness-efficiency trade-off: their work can easily support many previouslyunsupported theories (including theories that are non-monotonic), however, those22Chapter 2. Backgroundsolvers are typically much less efficient than SMT solvers with dedicated supportfor the same theories.The second of the two, [133], introduced SAT-to-SAT. Like [220], they de-scribe an off-line, lazy SMT solver. However, rather than utilizing an external mathlibrary to implement the theory solver, they utilize a second SAT solver to imple-ment the theory solver. Further, they negate the return value of that second SATsolver, essentially treating it as solving a negated, universally quantified formula.Their approach is similarly structured to the 2QBF solvers proposed in [172] and[185]. Like MathCheck, SAT-to-SAT is not restricted to monotonic theories. How-ever, SAT-to-SAT requires encoding the supported theory in propositional logic,and suffers a substantial performance penalty compared to our approach (for thetheories that both of our solvers support).One of the main contributions of this thesis is our framework for interpretedmonotonic functions in SMT solving. Although there has been work addressinguninterpreted monotonic functions in SMT solvers [23], we believe that no previ-ous work has specifically addressed exploiting interpreted monotonic functions inSMT solvers.Outside of SAT and SMT solving, other techniques in formal methods havelong considered monotonicity. In the space of non-SMT based constraint solvers,two examples include [213], who proposed a generic framework for building CSPsolvers for monotonic constraints; and interval arithmetic solvers, which are oftendesigned to take advantage of monotonicity directly (see, e.g., [122] for a sur-vey). Neither of these techniques directly extend to SMT solvers, nor have theybeen applied to the monotonic theories that we consider in this thesis. Addition-ally, several works have exploited monotonicity within SAT/SMT-based model-checkers [43, 96, 115, 142].23Chapter 3SAT Modulo Monotonic TheoriesOur interest in this work is to develop tools for building solvers for finite-domaintheories in which all predicates and functions are monotonic — i.e., they consis-tently increase (or consistently decrease) as their arguments increase. These arethe finite monotonic theories.1 In this chapter, we formally define monotonic pred-icates (Section 3.1) and monotonic theories (Section 3.2).While this notion of a finite monotonic theory may initially seem limited, wewill show that such theories are natural fits for describing many useful propertiesof discrete, finite structures, such as graphs (Chapter 5) and automata (Chapters8). Moreover, we have found that many useful finite monotonic theories can beefficiently solved in practice using a common set of simple techniques for buildinglazy SMT theory solvers (Section 2.4). We will describe these techniques — theSAT Modulo Monotonic Theories (SMMT) framework — in Chapter 4.3.1 Monotonic PredicatesConceptually, a (positive) monotonic predicate P is one for which if P(x) holds,and y≥ x, then P(y) holds. An example of a monotonic predicate is IsPositive(x) :R 7→ {T,F}, which takes a single real-valued argument x, and returns TRUE iffx > 0. An example of a non-monotonic predicate is IsPrime(x). Formally:1To forestall confusion, note that our concept of a ‘monotonic theory’ here has no direct relation-ship to the concept of monotonic/non-monotonic logics.24Chapter 3. SAT Modulo Monotonic TheoriesDefinition 1 (Monotonic Predicate). A predicate P: {σ1,σ2, . . .σn} 7→ {T,F},over sorts σi is monotonic iff, for each i in 1..n, the following holds:Positive monotonic in argument i:∀s1 . . .sn,∀x≤ y : P(. . . ,si−1,x,si+1, . . .)→ P(. . . ,si−1,y,si+1, . . .)—or—Negative monotonic in argument i:∀s1 . . .sn,∀x≤ y : ¬P(. . . ,si−1,x,si+1, . . .)→¬P(. . . ,si−1,y,si+1, . . .)We will say that a predicate P is positive monotonic in argument i if ∀s1 . . .sn,∀x≤ y : P(. . . ,si−1,x,si+1, . . .)→ P(. . . ,si−1,y,si+1, . . .); we will say that P isnegative monotonic in argument i if ∀s1 . . .sn,∀x≤ y :¬P(. . . ,si−1,x,si+1, . . .)→¬P(. . . ,si−1,y,si+1, . . .). Notice that although P must be monotonic in all ofits arguments, it may be positive monotonic in some, and negative monotonicin others.Notice that Definition 1 does not specify whether ≤ forms a total or a partialorder. In this work, we will typically assume total orders, though we considerextensions to support partial orders in Section 4.4.This work is primarily concerned with monotonic predicates over finite sorts(such as Booleans and bit vectors); we refer to such predicates as finite monotonicpredicates. Two closely related special cases of finite monotonic predicates de-serve attention: monotonic predicates over Booleans, and monotonic predicatesover powerset lattices. Formally, we define Boolean monotonic predicates as:25Chapter 3. SAT Modulo Monotonic TheoriesDefinition 2 (Boolean Monotonic Predicate). A predicate P: {T,F}n 7→ {T,F}is Boolean monotonic iff, for each i in 1..n, the following holds:Positive monotonic in argument i:∀s1 . . .sn : P(. . . ,si−1,F,si+1, . . .)→ P(. . . ,si−1,T,si+1, . . .)—or—Negative monotonic in argument i:∀s1 . . .sn : ¬P(. . . ,si−1,F,si+1, . . .)→¬P(. . . ,si−1,T,si+1, . . .)As an example of a Boolean monotonic predicate, consider the pseudo-Booleaninequality ∑n−1i=0 cibi ≥ cn, with each bi a variable in {T,F}, and each ci a non-negative integer constant. This inequality can be modeled as a positive Booleanmonotonic predicate P over the Boolean arguments bi, such that P is TRUE iff theinequality is satisfied.This definition of monotonicity over Booleans is closely related to a definitionof monotonic predicates of powerset lattices found in [41] and [145]. Given aset S, [41] defines a predicate P : 2S 7→ {T,F} to be monotonic if P(Sx)→ P(Sy)for all Sx ⊆ Sy.2 Slightly generalizing on that definition, we define a monotonicpredicate over set arguments as:Definition 3 (Set-wise Monotonic Predicate). Given sets S1,S2 . . .Sn a pred-icate P: {2S1 ,2S2 . . .2Sn} 7→ {T,F}, is monotonic iff, for each i in 1..n, thefollowing holds:Positive monotonic in argument i:∀s1 . . .sn,∀sx ⊆ sy : P(. . . ,si−1,sx,si+1, . . .)→ P(. . . ,si−1,sy,si+1, . . .)—or—Negative monotonic in argument i:∀s1 . . .sn,∀sx ⊆ sy : ¬P(. . . ,si−1,sx,si+1, . . .)→¬P(. . . ,si−1,sy,si+1, . . .)2Actually, [41] poses a slightly stronger requirement, requiring that P(S) must hold. We relaxthis requirement, in addition to generalizing their definition to support multiple arguments of bothpositive and negative monotonicity.26Chapter 3. SAT Modulo Monotonic TheoriesAs an illustrative example of a finite monotonic predicate of sets, considergraph reachability: reachs,t(E), over a set of edges E ⊆ V ×V , for a finite set ofvertices V . Predicate reachs,t(E) is TRUE iff there is a path from s to t through theedges of E (see example in Figure 3.1).01e02e1e2 3e3e4reach0,3(E)∧¬reach1,3(E)∧¬(e0 ∈ E ∧ e1 ∈ E)Figure 3.1: A finite symbolic graph over set E of five potential edges, anda formula constraining those edges. A satisfying assignment to E isE = {e1,e4}.Notice that reachs,t(E) describes a family of predicates over the edges of the graph:for each pair of vertices s, t, there is a separate reach predicate in the theory. Eachreachs,t(E) predicate is monotonic with respect to the set of edges E: if a node vis reachable from another node u in a given graph that does not contain edgei, thenit must still be reachable in an otherwise identical graph that also contains edgei.Conversely, if a node v is not reachable from node u in the graph, then removingan edge cannot make v reachable.Notice that these set-wise predicates are monotonic with respect to subsetinclusion, a partial-order. However, observe that any finite set-wise predicatep(S0,S1, . . .) can be converted into Boolean predicates, by representing each setargument Si as a vector of Booleans, where Boolean argument bS(i, j) is TRUEiff element j ∈ Si. For example, consider again the set-wise monotonic pred-icate reachs,t(E). As long as E is finite, we can convert reachs,t(E) into alogically equivalent Boolean monotonic predicate over the membership of E,27Chapter 3. SAT Modulo Monotonic Theoriesreachs,t,E(edge1,edge2,edge3, . . .), where the Boolean arguments edgei definewhich edges (via some mapping to the fixed set of possible edges E) are includedin the graph.This transformation of set-wise predicates into Boolean predicates will prove tobe advantageous later, as the Boolean formulation is totally-ordered (with respectto each individual Boolean argument), whereas the set-wise formulation is onlypartially ordered. As we will discuss in Section 4.4, the SMMT framework —while applicable also to partial orders — works better for total orders. Below, weassume that monotonic predicates of finite sets are always translated into logicallyequivalent Boolean monotonic predicates, unless otherwise stated.3.2 Monotonic TheoriesMonotonic predicates and functions — finite or otherwise — are common in theSMT literature, but in most cases are considered alongside collections of non-monotonic predicates and functions. For example, the inequality x+ y > z, withx,y,z real-valued, is an example of an infinite domain monotonic predicate in thetheory of linear arithmetic (positive monotonic in x,y, and negative monotonic inz). However, the theory of linear arithmetic — as with most common theories —can also express non-monotonic predicates (e.g., x = y, which is not monotonic ineither argument).We introduce the restricted class of finite monotonic theories, which are theo-ries over finite domain sorts, in which all predicates and functions are monotonic.Formally, we define a finite monotonic theory as:Definition 4 (Finite Monotonic Theory). A theory T with signature Σ is finitemonotonic if and only if:1. All sorts in Σ have finite domains;2. all predicates in Σ are monotonic; and3. all functions in Σ are monotonic.As is common in the context of SMT solving, we consider only decidable,28Chapter 3. SAT Modulo Monotonic Theoriesquantifier-free, first-order theories. All predicates in the theory must be mono-tonic; atypically for SMT theories, monotonic theories do not include equality (asequality is non-monotonic). Rather, as each sort σ ∈ T is ordered, we assume thepresence of comparison predicates in T : <,≤,≥,>, over σ×σ . Unlike the equal-ity predicate, the comparison predicate is monotonic, and we will take advantageof this property subsequently.Above, we described two finite monotonic predicates: a predicate of pseudo-Boolean comparisons, and a predicate of finite graph reachability. A finite mono-tonic theory might collect together several monotonic predicates that operate overone or more sorts. For example, many common graph properties are monotonicwith respect to the edges in the graph, and a finite monotonic theory of graphsmight include in its signature several predicates, including the above mentionedreachs,t(E), as well as additional monotonic predicates such as acyclic(E), whichevaluates to TRUE iff edges E do not contain a cycle, or planar(E)which evaluatesto TRUE iff edges E induce a planar graph. 3Although almost all theories considered in the SMT literature are non-monotonic, particularly as most theories implicitly include the equality predicate,we will show that many useful properties — including many properties that havenot previously had practical support in SAT solvers — can be modeled as mono-tonic finite theories, and solved efficiently. Moreover, we will show in Chapter4, that building high-performance SMT solvers for such theories is simple andstraightforward. Subsequently, in Chapters 5, 7, and 8 of this thesis, we willdemonstrate state-of-the-art implementations of lazy SMT solvers for several im-portant finite monotonic theories that previously had poor support in SAT solvers.3In previous work [33], we considered the special case of a Boolean monotonic theory, in whichthe only sort in the theory is Boolean (and hence all monotonic predicates are Boolean monotonicpredicates). In this thesis, the notion of a Boolean monotonic theory is subsumed by the more generalnotion of a finite monotonic theory.29Chapter 4A Framework for SAT ModuloMonotonic TheoriesThis chapter introduces a set of techniques — the SAT Modulo Monotonic The-ories (SMMT) framework — taking advantage of the special properties of finitemonotonic theories in order to create efficient SMT solvers, by providing efficienttheory propagation and improving upon naı¨ve conflict analysis.1Many successful SMT solvers follow the lazy SMT design [70, 179], in whicha SAT solver is combined with a set of theory solvers (or T-solvers), which areresponsible for reasoning about assignments to theory atoms (see Section 2.4 formore details about lazy SMT solvers).While T-solvers may support more complex interfaces, every T-solver must atminimum support two procedures:1. T-Propagate (M), which takes a (possibly partial) assignment M to thetheory atoms, and returns FALSE if it can prove that M is unsatisfiable inthe theory, and TRUE if it could not proveM to be unsatisfiable. 22. T-Analyze (M), which, given an unsatisfiable assignment to the theory liter-alsM, returns a subset (a conflict set) that is also unsatisfiable.31This chapter describes the SMMT framework at a high-level; we refer readers to Appendix Afor a discussion of some of the practical considerations required for an efficient implementation.2T-Propagate is sometimes referred to as T-Deduce in the literature.3T-Analyze is sometimes referred to as conflict or justification set derivation, or as lemma learning30Chapter 4. A Framework for SAT Modulo Monotonic TheoriesThe SMMT framework provides a pattern for implementing these two procedures.In principle, T-Propagate is only required to return FALSE when a completeassignment to the theory atoms is unsatisfiable (returning TRUE on partial assign-ments), and T-Analyze may always return the naı¨ve conflict set (i.e., it may returnthe entire assignmentM as the conflict set). However, efficient T-Propagate imple-mentations typically support checking the satisfiability of partial assignments (al-lowing the SAT solver to prune unsatisfiable branches early). Efficient T-Propagateimplementations are often also able to derive unassigned literals l that are impliedby a partial assignment M, such that T ∪M |= l. Any derived literals l can beadded toM, and returned to the SAT solver to extend the current assignment. Effi-cient implementations of T-Analyze typically make an effort to return small (some-times minimal) unsatisfiable subsets as conflict sets. Although there are many otherimplementation details that can impact the performance of a theory solver (such asthe ability for the theory solver to be applied efficiently as an assignment is in-crementally built up or as the solver backtracks), the practical effectiveness of alazy SMT solver depends on fast implementations of T-Propagate and T-Analyze.This implies a design trade-off, where more powerful deductive capabilities in thetheory solver must be balanced against faster execution time.In Sections 4.1 and 4.2, we show how efficient theory propagation and conflictanalysis can be implemented for finite monotonic theories. The key insight is toobserve that many predicates have known, efficient algorithms for evaluating theirtruth value when given fully specified inputs, but not for partially specified or sym-bolic inputs.4 For example, given a concretely specified graph, one can find the setof nodes reachable from u simply using any standard graph reachability algorithm,such as depth-first-search.Given only a procedure for computing the truth-values of the monotonic pred-icates P from complete assignments, we will show how we can take advantage ofthe properties of a monotonic theory to form a complete decision procedure for anyor clause learning (in which case the method returns a clause representing the negation of a conflictset).4Evaluating a function given a partially specified or symbolic input is typically more challengingthan evaluating a function on a fully specified, concrete input, as the return value may be a set ofpossible values, or a formula, rather than a concrete value. Additionally, evaluating a function on apartial or symbolic assignment may entail an expensive search over the remaining assignment space.31Chapter 4. A Framework for SAT Modulo Monotonic Theoriesfinite monotonic theory, and, further, we will show that in many cases the resultingtheory solver performs much better than preexisting solvers in practice.For clarity, in Sections 4.1 and 4.2 we initially describe how to apply the-ory propagation and conflict analysis for the special case of function-free, totally-ordered finite monotonic theories. These are finite monotonic theories that haveno functions in their signature, aside from predicates (of any arity), and aside fromconstant (arity-0) functions, and for which every sort has a total order relation-ship (rather than a partial order relationship). Subsequently, we relax these restric-tions to support formulas with compositions of positive and negative monotonicfunctions (Section 4.3), and to support finite sorts that are only partially-ordered(Section 4.4).4.1 Theory Propagation for Finite Monotonic TheoriesFirst, we show how efficient theory propagation can be performed for function-free, totally ordered, finite monotonic theories. Consider a theory with a monotonicpredicate P(x0,x1, . . .), with xi of finite sort σ . Let σ be totally ordered, over thefinite domain of constants σ⊥,σ1, . . . ,σ>, with σ⊥ ≤ σ1 ≤ . . .≤ σ>.In Algorithm 3, we describe the simplest version of our procedure for perform-ing theory propagation on partial assignments for function-free, totally-ordered fi-nite monotonic theories. For simplicity, we also assume for the moment that allpredicates are positive monotonic in all arguments.5Algorithm 3 is straightforward. Intuitively, it simply constructs conservativeupper and lower bounds (x+,x−) for each variable x, and then evaluates each predi-cate on those upper and lower bounds. Since the predicates are positive monotonic,if p(x−) evaluates to TRUE, then p(x), for any possible assignment of x≥ x−, mustalso evaluate to TRUE. Therefore, if p(x−) evaluates to TRUE, we can concludethat p(x) must be TRUE in any satisfying completion of the partial assignmentM.Similarly, if p(x+) evaluates to FALSE, then p(x) must be FALSE in any satisfyingcompletion ofM. The extension to multiple arguments is obvious, and presentedin Algorithm 3; we discuss further extensions in Sections 4.3 and 4.4.5In the special case of a function-free Boolean monotonic theory, in which all predicates arerestricted to purely Boolean arguments, Algorithm 3 can be simplified slightly by reducing compar-isons to truth assignments (e.g., (x < 1)≡ ¬x, if x ∈ {T,F}.).32Chapter 4. A Framework for SAT Modulo Monotonic TheoriesBecause x− and x+ are constants, rather than symbolic expressions, evaluatingp(x−) or p(x+) does not require any reasoning about p over symbolic arguments.As such, so long as one has any algorithm for computing predicate p on con-crete inputs (not symbolic), one can plug that algorithm into Algorithm 3 to geta complete theory propagation procedure. In the typical case that computing p iswell-studied, one has the freedom to use any existing algorithm for computing pin order to implement theory propagation and theory deduction — so long as p ismonotonic. For example, if p is the graph reachability predicate we considered inChapter 3, one can directly apply depth-first-search — or any other standard graphreachability algorithm from the literature — to evaluate p on the under and overapproximations of its graph.In other words, for finite monotonic theories, the same algorithms that can beused to check the truth value of a predicate in a witnessing model (i.e., a completeassignment of every atom and also of each variable to a constant value) can alsobe used to perform theory propagation and deduction (and, we will demonstrateexperimentally, does so very efficiently in practice). This is in contrast to the usualcase in lazy SMT solvers, in which the algorithms capable of evaluating the satisfi-ability of partial assignments may bear little resemblance to algorithms capable ofevaluating a witness — typically, the latter is trivial, while the former may be verycomplex.The distinction here is well-illustrated by any of the widely supported arith-metic theories, such as difference logic, linear arithmetic, or integer arithmetic.Checking a witness for any of these three arithmetic theories is trivial, requiringlinear time in the formula size. In contrast, finding efficient techniques for check-ing the satisfiability of a (partial or complete) assignment to the atoms in any ofthese three theories bears little resemblance to witness checking — for example,typical linear arithmetic theory solvers check satisfiability of partial assignmentsusing variations of the simplex algorithm(e.g., [80]), while difference logic solversmay perform satisfiability checking on partial assignments by testing for cycles inan associated constraint graph [144, 200]).The correctness of Algorithm 3 relies on Lemma 4.1.1, which relates modelsof atoms which are comparisons to constants, and models of atoms of positivemonotonic predicates. We observe that the following lemma holds:33Chapter 4. A Framework for SAT Modulo Monotonic TheoriesLemma 4.1.1 (Model Monotonicity). Let MA be an assignment only to atomscomparing variables to constants.6 For any given variable x with sort σ , any twoconstants σi ≤ σ j, and any positive monotonic predicate atom p:7MA∪{(x≥ σi)}=⇒Tp → MA∪{(x≥ σ j)}=⇒Tp (4.1)MA∪{(x≤ σ j)}=⇒T¬p → MA∪{(x≤ σi)}=⇒T¬p (4.2)A proof of this lemma can be found in Appendix C.1; the lemma easily gener-alizes for predicates of mixed positive and negative monotonicity.Algorithm 3 constructs, for each variable x, an over-approximation constant x+,and an under-approximation constant x−. It also constructs an under-approximateassignmentM−A in which for each variable x, (x≤ x−)∈M−A and (x≥ x−)∈M−A ,forcing the evaluation of x to be exactly x− inM−A . Similarly, Algorithm 3 createsan over-approximate model M+A forcing each variable x to x+. Note that bothof these assignments contain only atoms that compare variables to constants, andhence match the prerequisites of Lemma 4.1.1.By Lemma 4.1.1, if M−A =⇒T p, then M =⇒T p. Also by Lemma 4.1.1, ifM+A =⇒T ¬p, thenM =⇒T ¬p. By evaluating each predicate inM−A andM+A ,Algorithm 3 safely under- and over-approximates the truth value of each p inM,using only two concrete evaluations of each predicate atom.Lemma 4.1.1 guarantees that Algorithm 3 returns FALSE only if T |= ¬M;however, it does not guarantee that Algorithm 3 returns FALSE for all unsatisfiablepartial assignments. In the case whereMA is a complete and consistent assignmentto every variable x, we have x− = x+ for each x (and henceM−A =M+A ), and so foreach p, it must be the case that either p(x−0 ,x−1 , . . .) 7→ TRUE or p(x+0 ,x+1 , . . .) 7→FALSE. Therefore, ifMA is a complete assignment, Algorithm 3 returns FALSE ifand only if T |= ¬M.6MA contains only atoms of comparisons of variables to constants. Comparisons between twonon-constant variables are treated as monotonic predicates, and are not included inMA. Also, forsimplicity of presentation (but without loss of generality), we assume that all comparison atomsinMA are normalized to the form (x ≥ σi), or (x ≤ σi). In the special case that atom (x < σ⊥)or (x > σ>) (or, equivalently, theory literals ¬(x ≥ σ⊥),¬(x ≤ σ>)) are inMA, the assignment istrivially unsatisfiable, but cannot be normalized to inclusive comparisons.7Where A =⇒TB is shorthand for ‘(A∧¬B) is unsatisfiable in theory T ’.34Chapter 4. A Framework for SAT Modulo Monotonic TheoriesAlgorithm 3 is guaranteed to be complete only whenM+A =M−A , which is onlyguaranteed to occur when M is a complete assignment to the comparison atoms(but not necessarily a complete assignment to the theory predicates). If M+A 6=M−A , then it is possible for Algorithm 3 to fail to deduce conflicts, or unassignedatoms, that are in fact implied by the theory (in the terminology of Section 2.4,Algorithm 3 is not deduction complete). For example, Algorithm 3 may fail todeduce non-comparison atom p1 from a second non-comparison atom p0 in thecase that p0 =⇒Tp1.For this reason, Algorithm 3 must be paired with a procedure for searching thespace of possible assignments to A in order to form a complete decision procedure.For finite domain theories, this does not pose a problem, as one can either performinternal case splitting in the theory solver, or simply encode the space of possibleassignments to each variable in fresh Boolean variables (for example, as a unaryor binary encoding over the space of possible values in the finite domain of σ ), al-lowing the SAT solver to enumerate that (finite) space of assignments (an exampleof splitting-on-demand [28]). In practice, we have always chosen the latter option,and found it to work well across a wide range of theories.As we will show in subsequent chapters, Algorithm 3, when combined with anappropriate conflict analysis procedure, is sufficient to build efficient solvers for awide set of useful theories.Returning to our earlier example from Section 3.1 of a pseudo-Booleanconstraint predicate ∑n−1i=0 cibi ≥ cn, with Boolean arguments bi and constantsc0, . . .cn, Algorithm 3 could be implemented as shown in Algorithm 5.Our presentation of Algorithm 3 above is stateless: the UpdateBounds sec-tion recomputes the upper- and lower-approximation assignments at each call toT-Propagate. In a practical implementation, this would be enormously wasteful, asthe under- and over-approximate assignments will typically be expected to changeonly in small ways between theory propagate calls. In a practical implementation,one would store the upper- and lower-approximation assignments between calls toT-Propagate, and only alter them as theory literals are assigned or unassigned inM. A practical, stateful implementation of Algorithm 3 is described in more detailin Appendix A.For example, the pseudo-code in Algorithm 5 is a nearly complete representa-35Chapter 4. A Framework for SAT Modulo Monotonic Theoriestion of the actual implementation of theory propagation for pseudo-Boolean con-straints that we have implemented in our SMT solver, with the only difference be-ing that in our implementation we maintain running under- and over-approximatesums (∑n−1i=0 cib−i ,∑n−1i=0 cib+i ), updated each time a theory literal is assigned or unas-signed in the SAT solver, rather than re-calculating them at each T-Propagate call.36Chapter 4. A Framework for SAT Modulo Monotonic TheoriesAlgorithm 3 Theory propagation for function-free monotonic theories, assumingall predicates are positive monotonic, and all sorts are totally ordered. Algorithm 3takes a partial assignmentM to the theory atoms of T . If M is found to be un-satisfiable in T , it returns a tuple (FALSE,c), with c a conflict set produced byT-Analyze; else it returns (TRUE,M), with any deduced literals added toM.function T-Propagate(M)UpdateBounds:M−A ←{},M+A ←{}for each sort σ ∈ T dofor each variable x of sort σ doif (x < σ⊥) ∈M or (x > σ>) ∈M thenreturn FALSE, T-Analyze(M)M←M∪{(x≥ σ⊥),(x≤ σ>)}x−←max({σi|(x≥ σi) ∈M}).x+←min({σi|(x≤ σi) ∈M}).if x− > x+ then return FALSE, T-Analyze(M)M−A ←M−A ∪{(x≤ x−),(x≥ x−)}M+A ←M+A ∪{(x≤ x+),(x≥ x+)}PropagateBounds:for each positive monotonic predicate atom p(x0,x1, . . .) doif ¬p ∈M thenif evaluate(p(x−0 ,x−1 , . . .)) 7→ TRUE thenreturn FALSE, T-Analyze(M)elseTighten bounds (optional)M← TIGHTENBOUNDS(p,M,M+A ,M−A )else if p ∈M thenif evaluate(p(x+0 ,x+1 , . . .)) 7→ FALSE thenreturn FALSE, T-Analyze(M)elseTighten bounds (optional)M← TIGHTENBOUNDS(¬p,M,M−A ,M+A )elseDeduce unassigned predicate atoms:if evaluate(p(x−0 ,x−1 , . . .)) 7→ TRUE thenM←M∪{p}else if evaluate(p(x+0 ,x+1 , . . .)) 7→ FALSE thenM←M∪{¬p}return TRUE,M37Chapter 4. A Framework for SAT Modulo Monotonic TheoriesAlgorithm 4 TIGHTENBOUNDS is an optional routine which may be called byimplementations of Algorithm 3. TIGHTENBOUNDS takes a predicate p that hasbeen assigned a truth value, and performs a search over the arguments xi of p tofind tighter bounds on xi that can be added toM.function TIGHTENBOUNDS(p(x0,x1, . . . ,xn),M,M+A ,M−A )for each x0 . . .xn doif p is positive monotonic in argument i (which has sort σ ) thenfor each y ∈ σ ,M−A [xi]< y≤M+A [xi] doif evaluate(p(. . . ,M−A [xi−1],y,M−A [xi+1], . . .)) 7→ TRUE thenM←M∪{(xi < y)}elsefor each y ∈ σ ,M+A [xi]> y≥M−A [xi] doif evaluate(p(. . . ,M+A [xi−1],y,M+A [xi+1], . . .)) 7→ TRUE thenM←M∪{(xi > y)}returnM38Chapter 4. A Framework for SAT Modulo Monotonic TheoriesAlgorithm 5 Instantiation of Algorithm 3 for the theory of pseudo-Boolean con-straints. A practical implementation would store the under- and over-approximatesums between calls to T-Propagate, updating them as theory literals are assigned,rather than recalculating them each time.function T-Propagate(M)UpdateBounds:for each argument bi dob−i ← FALSE, b+i ← TRUEif bi ∈M thenb−i ← TRUEelse if ¬bi ∈M thenb+i ← FALSEPropagateBounds:for each predicate atom p = ∑n−1i=0 cibi ≥ cn doif ¬p ∈M thenif ∑n−1i=0 cib−i ≥ cn thenreturn FALSE, T-Analyze(M)elseTightenBounds:for each b j doif b+j ∧¬b−j ∧∑n−1i=0 cib−i + c j ≥ cn thenM←M∪{¬b j}else if p ∈M thenif ∑n−1i=0 cib+i < cn thenreturn FALSE, T-Analyze(M)elseTightenBounds:for each b j doif b+j ∧¬b−j ∧∑n−1i=0 cib+i − c j < cn thenM←M∪{b j}elseif ∑n−1i=0 cib−i ≥ cn thenM←M∪{p}else if ∑n−1i=0 cib+i < cn thenM←M∪{¬p}return TRUE,M39Chapter 4. A Framework for SAT Modulo Monotonic Theories4.2 Conflict Analysis for Finite Monotonic TheoriesIn Section 4.1, we described a technique for theory propagation in function-free,totally-ordered finite monotonic theories. As discussed at the opening of this chap-ter, the other function that efficient SMT solvers must implement is T-Analyze,which takes a partial assignmentM that is unsatisfiable in theory T , and returns asubset ofM (a ‘conflict set’) which remains unsatisfiable. Ideally, this conflict setwill be both small, and efficient to derive.All SMT solvers have the option of returning the naı¨ve conflict set, which is justto return the entireM as the conflict set. However, for the special case of conflictsderived by Algorithm 3, we describe an improvement upon the naı¨ve conflict set,which we call the default monotonic conflict set. So long as Algorithm 3 is used,returning the default monotonic conflict set is always an option (and is never worsethan returning the naı¨ve monotonic conflict set).While in many cases we have found that theory specific reasoning allows oneto find even better conflict sets than this default monotonic conflict set, in somecases, including the theory of pseudo-Booleans described in Algorithm 5, as wellas some of the predicates mentioned in subsequent chapters, our implementationdoes in fact fall back on the default monotonic conflict set.We describe our algorithm for deriving a default monotonic conflict set inAlgorithm 6. The first for-loop of Algorithm 6 simply handles the trivial conflictsthat can arise when building the over- and under-approximations of each variablex+,x− in Algorithm 3 (for example, ifM is inconsistent with the total order rela-tion of σ ); these cases are all trivial and self-explanatory.In the last 5 lines, Algorithm 6 deals with conflicts in which the under- orover-approximations x−,x+ are in conflict with assignment inM of a single pred-icate atom, p. Consider a predicate atom p(x0,x1), with p positive monotonic ineach argument. Algorithm 3 can derive a conflict involving p in one of only twocases: Either p(x+0 ,x+1 ) evaluates to FALSE, when p(x0,x1) is assigned TRUE inM, or p(x−0 ,x−1 ) evaluates to TRUE, when p(x0,x1) is assigned FALSE inM. Inthe first case, a sufficient conflict set can be formed from just the literals in Mused to set the upper bounds x+0 ,x+1 , . . . (along with the conflicting predicate literalp(x0,x1)). In the second case, only the atoms of M that were used to form the40Chapter 4. A Framework for SAT Modulo Monotonic TheoriesAlgorithm 6 Conflict set generation for finite monotonic theories. Assuming allpredicates are positive monotonic. Algorithm 6 takes M, a (partial) assignmentthat is unsatisfiable in T , and returns a conflict set, a subset of the atoms assignedinM that are mutually unsatisfiable in T .function T-Analyze(M)for each Variable x of sort σ doCompute x−,x+ as in T-Propagate.case (x < σ⊥) ∈Mreturn {(x < σ⊥)}case (x > σ⊥) ∈Mreturn {(x > σ⊥)}case @σi : (σi ≥ x−)∧ (σi ≤ x+)return {(x≥ x−),(x≤ x+)}for each monotonic predicate atom p(t0, t1, . . .) docase ¬p ∈M, evaluate(p(x−0 ,x−1 , . . .)) 7→ TRUEreturn {¬p,(x0 ≥ x−0 ),(x1 ≥ x−1 ), . . .}case p ∈M, evaluate(p(x+0 ,x+1 , . . .)) 7→ FALSEreturn {p,(x0 ≤ x+0 ),(x1 ≤ x+1 ), . . .}under-approximations x−0 ,x−1 need to be included in the conflict set (along with thenegated predicate atom). In other words, when Algorithm 3 finds a non-trivial con-flict on monotonic predicate atom p, the default monotonic conflict set can excludeeither all of the over-approximate atoms, or all of the under-approximate atoms,from M (except for the atom of the conflicting predicate atom p) — improvingover the naı¨ve conflict set, which does not drop these atoms.Formally, the correctness of this conflict analysis procedure follows fromLemma 4.1.1, in the previous section. As before, letMA be an assignment that con-tains only the atoms ofM comparing variables to constants: {(x≤ σi),(x≥ σi)}.By Lemma 4.1.1, if M−A =⇒T p, then the comparison atoms (xi ≥M−A [xi]) inMA, for the arguments x0,x1, . . . of p, are sufficient to imply p by themselves.Also by Lemma 4.1.1, ifM+A =⇒T ¬p, then the comparison atoms (xi ≤M+A [xi])inMA, for the arguments x0,x1, . . . of p, are sufficient to imply ¬p by themselves.Therefore, the lower-bound (resp. upper-bound) comparison atoms inMA that are41Chapter 4. A Framework for SAT Modulo Monotonic Theoriesarguments of p can safely form justification sets for p (resp. ¬p).8The default monotonic conflict set is always available when Algorithm 3 isused, and improves greatly on the naı¨ve conflict set, but it is not required (or evenrecommended) in most cases. In practice, it is often possible to improve on thisdefault monotonic conflict set, for example by discovering that looser constraintsthan M−A [xi] or M+A [xi] would be sufficient to imply the conflict, or that somearguments of p are irrelevant.Many common algorithms are constructive in the sense that they not only com-pute whether p is true or false, but also produce a witness (in terms of the inputs ofthe algorithm) that is a sufficient condition to imply that property. In many cases,the witness will be constructed strictly in terms of inputs corresponding to atomsthat are assigned TRUE (or alternatively, strictly in terms of atoms that are assignedFALSE). This need not be the case — the algorithm might not be constructive, or itmight construct a witness in terms of some combination of the true and false atomsin the assignment — but, as we will show in Chapters 5, 7, and 8, it commonly isthe case for many theories of interest. For example, if we used depth-first search tofind that node v is reachable from node u in some graph, then we obtain as a sideeffect a path from u to v, and the theory atoms corresponding to the edges in thatpath imply that v can be reached from u.Any algorithm that can produce a witness containing only (x ≥ σi) atoms canbe used to produce conflict sets for any positive predicate atom assignments prop-agated from the under-approximation arguments above. Similarly, any algorithmthat can produce a witness of (x ≤ σi) atoms can also produce conflict sets forany negative predicate atom assignments propagated from over-approximation ar-guments above. Some algorithms can produce both. In practice, we have oftenfound standard algorithms that produce one, but not both, types of witnesses. Incases where improved conflict analysis procedures are not available, one can al-ways safely fall back on the default monotonic conflict analysis procedure of Al-gorithm 6.8Note that we can safely drop any comparisons (x≥ σ⊥) or (x≤ σ>) from the conflict set.42Chapter 4. A Framework for SAT Modulo Monotonic Theories4.3 Compositions of Monotonic FunctionsIn the previous two sections, we described algorithms for T-Propagate and T-Analyze, for the special case of function-free finite theories with totally orderedsorts and positive monotonic predicates. We now extend that approach in two ways:first, to support both positive and negative predicates, and second, to support bothpositive and negative monotonic functions. In fact, the assumption in Algorithm3 that all predicates are positive monotonic in all arguments was was simply forease of presentation, as handling predicates that are negative monotonic, or havemixed monotonicity (positive monotonic in some arguments, negative in others)is straightforward. However, dealing with compositions of positive and negativemonotonic multivariate functions (rather than predicates) is more challenging, as acomposition of positive and negative monotonic functions is not itself guaranteedto be monotonic.For example, consider the functions f (x,y) = x+ y, g(x) = 2x, h(x) = −x3.Even though each of these is either a positive or a negative monotonic function,the composition f (g(x),h(x)) is non-monotonic in x. As a result, a naı¨ve approachof evaluating each term in the over- or under-approximative assignments might notproduce safe upper and lower bounds (e.g., f (g(M+A [x]),h(M+A [x])) may not be asafe over-approximation of f (g(x),h(x))).One approach to resolve this difficulty would be to flatten function compo-sitions through the introduction of fresh auxiliary variables, and then to replaceall functions with equivalent predicates relating the inputs and outputs of thosefunctions (this is the approach we have taken to support bitvectors, as describedin Chapter 5). After such a transformation, the resulting formula no longer hasany non-predicate functions, and as each of the newly introduced predicates arethemselves monotonic, the translated formula can be handled by Algorithm 3. Adrawback to this approach is that Algorithm 3 only considers each predicate atomon its own, and does not directly reason about interactions among these predicates,instead repeatedly propagating derived upper and lower bounds on arguments ofthe predicates back to the SAT solver. This may result in a large number of re-peated calls to Algorithm 3 asM is gradually refined by updating the bounds onthe arguments of each predicate until a fixed point is reached, which may result in43Chapter 4. A Framework for SAT Modulo Monotonic TheoriesAlgorithm 7 Recursively, builds up a safe, approximate evaluation of composedpositive and negative monotonic functions and predicates. If M+A is an over-approximative assignment, andM−A an under-approximative assignment, then AP-PROX returns a safe over-approximation of the evaluation of φ . IfM+A is insteadan under-approximation, andM−A an over-approximation, then APPROX returns asafe under-approximation of the evaluation of φ .function APPROX(φ ,M+A ,M−A )φ is a formula,M+A ,M−A are assignments.if φ is a variable or constant term thenreturnM+A [φ ]else φ is a function term or predicate atom f (t0, t1, . . . , tn)for 0≤ i≤ n doif f is positive monotonic in ti thenxi = APPROX(ti,M+A ,M−A )else argumentsM+A ,M−A are swappedxi = APPROX(ti,M−A ,M+A )return evaluate( f (x0,x1,x2, . . . ,xn))unacceptably slow performance for large formulas.As an alternative, we introduce support for compositions of monotonic func-tions by evaluating compositions of functions approximately, without introducingauxiliary variables, in such a way that the approximation remains monotonic underpartial assignments, while under a complete assignment the approximation con-verges to the correct value.To do so, we introduce the function approx(φ ,M+A ,M−A ), presented in Algo-rithm 7. This function can be used to form both safe over-approximations and safeunder-approximations of compositions of both positive and negative monotonicfunctions. Function approx(φ ,M+A ,M−A ) takes a term φ , which is either a variable,a constant, or a monotonic function or predicate (either positive or negative mono-tonic, or a mixture thereof); and in whichM+A ,M−A are both complete assignmentsto the variables in φ . Intuitively, ifM+A is a safe over-approximation to the vari-ables of φ , and M−A a safe under-approximation, then approx(φ ,M+A ,M−A ) willreturn a safe over-approximation of φ . Conversely, approx(φ ,M−A ,M+A ) (swap-ping the 2nd and 3rd arguments) will return a safe under-approximation of φ . Fur-ther, if both assignments are identical, then approx(φ ,MA,MA) returns an exact44Chapter 4. A Framework for SAT Modulo Monotonic TheoriesAlgorithm 8 Replacement for PropagateBounds section of Algorithm 3. Sup-porting compositions of mixed positive and negative monotonic predicates andfunctions.function PROPAGATEBOUNDS(M,M+A ,M−A )M,M+A ,M−A are assignments, withM+A ,M−A computed byUpdateBounds.for each monotonic predicate atom p(t0, t1, . . .) doif ¬p ∈M thenif APPROX(p,M−A ,M+A ) 7→ TRUE thenreturn FALSE,T-Analyze(M)elseTighten bounds (optional)M← TIGHTENBOUNDS(p,M,M+A ,M−A )else if p ∈M thenif APPROX(p,M+A ,M−A ) 7→ FALSE thenreturn FALSE,T-Analyze(M)elseTighten bounds (optional)M← TIGHTENBOUNDS(¬p,M,M−A ,M+A )elseif APPROX(p,M−A ,M+A ) 7→ TRUE thenM←M∪{p}else if APPROX(p,M+A ,M−A ) 7→ FALSE thenM←M∪{¬p}return TRUE,Mevaluation of φ inMA.Formally, if we have someM∗A such that ∀x,M+A [x] ≥M∗A[x] ≥M−A [x], thenapprox(φ ,M+A ,M−A )≥M∗A[φ ]≥ approx(φ ,M−A ,M+A ). A proof can be found inAppendix C.2.As approx(φ ,M+A ,M−A ) and approx(φ ,M−A ,M+A ) form safe, monotonic,upper- and lower-bounds on the evaluation of MA[φ ], we can directly substituteapprox into Algorithm 3 in order to support compositions of mixed positive andnegative monotonic functions and predicates. These changes are described in Al-gorithm 8.In Algorithm 9, we introduce corresponding changes to the conflict set gener-ation algorithm in order to support Algorithm 8. Algorithm 9 makes repeated calls45Chapter 4. A Framework for SAT Modulo Monotonic TheoriesAlgorithm 9 Conflict set generation for finite monotonic theories with functions.Allowing mixed positive and negative monotonic predicates and functions.function T-Analyze(M)M is a partial assignment.ComputeM−A ,M+A as in UpdateBounds.for each Variable x of sort σ docase x < σ⊥ ∈Mreturn {x < σ⊥}case x > σ⊥ ∈Mreturn {x > σ⊥}case @σi : (σi ≥M−A [x])∧ (σi ≤M+A [x])return {x≥M−A [x],x≤M+A [x]}for each monotonic predicate atom p docase ¬p ∈M, APPROX(p,M−A ,M+A ) 7→ TRUEreturn {¬p}∪ANALYZEAPPROX(p,≤,FALSE,M−A ,M+A )case p ∈M, APPROX(p,M+A ,M−A ) 7→ FALSEreturn {p}∪ANALYZEAPPROX(p,≥,TRUE,M+A ,M−A )to a recursive procedure, analyzeApprox (shown in Algorithm 10). Algorithm 10produces a conflict set in terms of the upper and lower bounds of the variables x inM+A andM−A . Algorithm 10 recursively evaluates each predicate and function onits respective upper and lower bounds, swapping the upper and lower bounds whenevaluating negatively monotonic arguments.Notice that if p is positive monotonic, and ti is a ground variable, then Algo-rithm 9 simply returns the same atom ti ≥ σi or ti ≤ σi, for some constant σi, as itwould in Algorithm 6. As such, Algorithm 9 directly generalizes Algorithm 6.46Chapter 4. A Framework for SAT Modulo Monotonic TheoriesAlgorithm 10 The analyzeApprox function provides conflict analysis for the re-cursive approx function.function ANALYZEAPPROX(φ ,op,k,M+A ,M−A )φ is a formula, op is a comparison operator, k is a constant,M+A ,M−A areassignments.if φ is a constant term thenreturn {}else if φ is a variable thenif op is ≤ thenreturn {φ ≤ k}elsereturn {φ ≥ k}else φ is a function or predicate term f (t0, t1, . . . , tn)c←{}for 0≤ i≤ n doif f is positive monotonic in ti thenxi = ANALYZEAPPROX(ti,M+A ,M−A )c← c∪ ANALYZEAPPROX(ti,op,xi,M+A ,M−A )else argumentsM+A ,M−A are swappedxi = ANALYZEAPPROX(ti,M−A ,M+A )c← c∪ ANALYZEAPPROX(ti,−op,xi,M−A ,M+A )return c47Chapter 4. A Framework for SAT Modulo Monotonic Theories4.4 Monotonic Theories of Partially Ordered SortsAlgorithms 3 and 8, as presented above, assume finite, totally ordered sorts. How-ever, observing that Lemma 4.1.1 applies also to partial-orders, Algorithms 3 and8 can also be modified to handle finite, partially-ordered sorts.Conceptually, theory propagation applied to partial orders proceeds similarlyto theory propagation as applied to total orders, with two major changes. Firstly,checking whether the comparison atoms inM are consistent is slightly more in-volved (and, correspondingly, conflict analysis for the case where the comparisonatoms are inconsistent is also more involved). Secondly, even in cases where thecomparison atoms are consistent, in a partial order there may not exist a constantx+ to form the upper bound for variable x (or there may not exist a lower bound x−);or those bounds may exist, but be very ‘loose’, and hence the theory propagationprocedure may fail to make deductions that are in fact implied by the comparisonatoms ofM.For example, if σ is partially ordered, ¬(x ≥ y) does not necessarily implyx < y, as it may also be possible for x and y to be assigned incomparable values;consequently, it may be possible for both atoms ¬(x ≥ y),¬(x ≤ y) to be inMAwithout implying a conflict. This causes complications for conflict analysis, andalso for the forming of upper and lower bounds needed by Algorithms 3 and 8.In Algorithm 11, we describe modifications to the UpdateBounds section ofAlgorithm 3 to support partially ordered sorts, by forming conservative lower andupper bounds x− and x+, where possible, and returning without making deduc-tions when such bounds do not exist. These changes are also compatible with thechanges to PropagateBounds described in Algorithm 8, and by combining both,one can support theory propagation for finite monotonic theories with functionsand partially ordered sorts.Unfortunately, if σ is only a partial order, then it is possible that either or bothof meet(max(X−)), join(min(X+)) may not exist. In the case that either or both donot exist, Algorithm 11 simply returns, making no deductions at all. Consequently,for partial orders, Algorithm 11 may fail to make deductions that are in fact impliedby the comparisons of MA (here, we define MA as we did in previous sections:MA is a subset ofM containing assignments only to atoms comparing variables48Chapter 4. A Framework for SAT Modulo Monotonic Theoriesto constants ((x ≤ σi), (x ≥ σi)). In fact, even in the case where safe upper andlower bounds do exist, there may exist deductionsMA =⇒Tp (orMA =⇒T¬p)that this algorithm fails to discover. In the special case that σ is totally ordered,X+ and X− can always be represented as singletons (or the empty set, if there isa conflict), max(X−) and min(X+) will always return single constants, the meetand join will always be defined, and the resulting construction of x+,x− becomescompletely equivalent to our presentation in Algorithm 3.The imprecision of Algorithm 11’s deductions is a consequence of reducing theupper and lower bounds to singletons, which allows the upper and lower bounds ofeach predicate p to be checked with only two evaluations of p (one for the upperbound, and one for the lower bound). An alternative approach, which would dis-cover all deductions MA =⇒Tp (or MA =⇒T¬p), would be to search over thespace of (incomparable) maximal and minimal elements max(x+) and min(x−) foreach variable x. In principle, by evaluating all possible combinations of maximaland minimal assignments to each argument of p, one could find the absolute maxi-mum and minimum evaluations of p. However, while there are general techniquesfor optimizing functions over posets (see ordinal optimization, e.g. [78]), imple-menting such a search would likely be impractical unless the domain of σ is verysmall.Above, we described a general purpose approach to handling monotonic theo-ries over finite-domain partially ordered sorts. An alternative approach to support-ing finite partially ordered sorts would be to convert the partial order into a totalorder (e.g., using a topological sorting procedure), after which one could directlyapply the approach of Section 4.1. Unfortunately, strengthening the order relation-ship between elements of the sort may force the solver to learn weaker clauses thanit otherwise would. For example, if a ≤ b holds in the total order, then Algorithm6 may in some cases include that (spurious) comparison atom in learned clauses,even if a and b are in fact incomparable in the original partial order.An important special case of a partially ordered sort is a finite powerset lattice.If the sort is a powerset lattice, then we can translate the space of possible subsetsinto a bit string of Boolean variables, with each Boolean determining the presenceof each possible element in the set. In this case, the size of the represented set ismonotonic with respect to not only the value of the bit string, but also with respect49Chapter 4. A Framework for SAT Modulo Monotonic Theoriesto each individual Boolean making up that string. This transformation allows oneto model the powerset lattice as a totally ordered, monotonic theory over the indi-vidual Booleans of the bit string, and hence to perform theory propagation usingAlgorithms 3 or 8. This is the approach that we take in our implementation of thegraph and finite state machine theories in Chapters 5 and 8, which operate oversets of graphs and Kripke structures, respectively. We have found it to performvery well in practice.50Chapter 4. A Framework for SAT Modulo Monotonic TheoriesAlgorithm 11 Replacement for UpdateBounds section of Algorithm 3, supportingtheory propagation for finite monotonic theories over partially ordered sorts.function UPDATEBOUNDS(M)M−A ←{},M+A ←{}for each variable x of sort σ doX+←{σ⊥,σ1,σ2 . . .σ>}X−←{σ⊥,σ1,σ2 . . .σ>}for each (x > σ j) inM doX− = X− \{∀σi|σi ≤ σ j}for each (x≥ σ j) inM doX− = X− \{∀σi|σi < σ j}for each ¬(x < σ j) inM doX− = X− \{∀σi|σi < σ j}for each ¬(x≤ σ j) inM doX− = X− \{∀σi|σi ≤ σ j}for each (x < σ j) inM doX+ = X+ \{∀σi|σi ≥ σ j}for each (x≤ σ j) inM doX+ = X+ \{∀σi|σi > σ j}for each ¬(x > σ j) inM doX+ = X+ \{∀σi|σi > σ j}for each ¬(x≥ σ j) inM doX+ = X+ \{∀σi|σi ≥ σ j}if X+∩X− = {} then return FALSE,T-Analyze(M)Note: As the minimal (resp. maximal) elements of X− (resp. X+) may notbe unique, min (resp. max) returns a set of minimal (resp. maximal) elements.x−← meet(min(X−))x+← join(max(X+))if x− or x+ does not exist, or x− and x+ are incomparable thenreturn TRUE,Mif x− > x+ then return FALSE,T-Analyze(M)M−A ←M−A ∪{(x≤ x−),(x≥ x−)}M+A ←M+A ∪{(x≤ x+),(x≥ x+)}51Chapter 4. A Framework for SAT Modulo Monotonic Theories4.5 Theory Combination and Infinite DomainsTwo additional concerns should be touched upon before continuing, both related tothe finite domain requirement of finite monotonic theories. First, we can ask, whyare the techniques introduced in this section restricted to finite domain sorts?If the domain is infinite, then in the general case the conflict analysis proce-dure (Algorithm 6 or Algorithm 9) may not be sufficient to guarantee termination,as it may introduce a non-converging sequence of new comparison atoms with in-finitesimally tighter bounds. Unfortunately, we have found no obvious way — inthe general case — to extend support to monotonic theories of infinite domains thatwould resolve this concern.The second concern is theory combination. In general, finite domain sorts —including Booleans, bit vectors, and finite sets — violate the stably-infinite re-quirement of Nelson-Oppen [156] theory combination, which requires that a the-ory have, for every sort, an infinite number of satisfying assignments (models). Assuch, Nelson-Oppen theory combination is not directly applicable to finite mono-tonic theories.9Finite domain theories can, in general, be combined at the propositional levelby introducing fresh variables for each shared variable in the two theories, andthen asserting equality at the propositional level. This can either be accomplishedby passing equality predicates over shared constants between the two theories (as inNelson-Oppen), or by enumerating over the space of possible values each variablecan take and asserting that exactly the same assignment is chosen for those freshvariables.Either of the above theory combination approaches are poor choices for mono-tonic theories, for the reason that it provides a very poor mechanism for communi-cating derived bounds on the satisfiable ranges of each variable between each the-ory solver, when operating on partial assignments. For example, when operatingon a partial assignment, one theory solver may have derived the bounds 3≤ x≤ 5,9Note that, even if we were to consider monotonic theories with infinite domains, we would stillencounter a difficulty, which is that the theory of equality is non-monotonic, and hence monotonictheories do not directly support equality predicates. It seems hopeful that Nelson-Oppen style theorycombination could be applied to theories supporting only comparisons, for example by emulatingequality through conjunctions of ≤,≥ constraints. However, we present no proof of this claim.52Chapter 4. A Framework for SAT Modulo Monotonic Theorieshowever, until x is forced to a specific value (e.g., x = 3), the above approach pro-vides no way to communicate those bounds on x to other theory solvers.Instead, we recommend a form of delayed-theory combination, over the spaceof comparison atoms. This approach is based on Algorithm 12, which iterativelypropagates comparison atoms between two finite monotonic theory solvers forT0,T1. By itself, Algorithm 12 only performs theory propagation to the combinedtheory T0∪T1; it must be combined with a complete decision procedure (for exam-ple, by integrating it into an SMT solver as a theory solver), to become a completetheory combination procedure.10Consider two finite theories, T0 and T1 with shared variables x0,x1, . . ., and withT0 a monotonic theory. So long as both theories support comparisons to constants(x ≥ σi,x ≤ σi), the upper and lower bounds generated by Algorithm 3 can bepassed from one theory solver to the other (by lazily introducing new comparison-to-constant atoms), each time bounds are tightened during theory propagation.11 Iftheory T1 happens also to be able to compute upper and lower bounds during theorypropagation (which may be the case whether or not T1 is a monotonic theory), thenthose upper and lower bounds can also be added toMA in T .It is easy to see that Algorithm 12 must terminate, as at each step in whicha conflict does not occur, M must grow to include a new comparison atom, orelse the algorithm will terminate. As σ is finite, there are only a finite number ofcomparison atoms that can be added toM, so termination is guaranteed. That Al-gorithm 12 converges to a satisfying model in which the shared variables have thesame assignment in each theory solver is also easy to see: WhenM is a completeassignment, it must either be the case that either T0 or T1 derives a conflict andAlgorithm 12 returns FALSE, or it must be the case that, for each shared variable x,∃σi,(x≥ σi) ∈M∧ (x≤ σi) ∈M∧ (x′ ≥ σi) ∈M∧ (x′ ≤ σi) ∈M.As we will describe in Chapter 5, we have used Algorithm 12 to combine ourgraph theory with a non-wrapping theory of bitvectors, and found it to work wellin practice.10Note that the theory combination technique we describe is not a special case of Nelson-Oppentheory combination, as the theories in question are neither signature-disjoint nor stably-infinite.11As the domain of σ is finite, there are only a finite number of new comparison atoms that canbe introduced, and so we can safely, lazily introduce new comparison atoms as needed during theorypropagation without running into termination concerns.53Chapter 4. A Framework for SAT Modulo Monotonic TheoriesAlgorithm 12 Apply theory propagation to the combination of two finite mono-tonic theories, T0 ∪ T1. As in Nelson-Oppen, we purify expressions over sharedvariables by introducing a fresh variable x′i for each shared variable xi, such thatxi only appears in expressions over predicates or functions from a single theory,with x′i replacing xi in expressions over predicates or functions from the other the-ory. We assume both T0 and T1 both support comparisons to constant predicates(x≤ σi),(x≥ σi), and the shared set of constants {σ⊥,σ1, . . .σ>}.function T-Propagate(M,T0,T1)M is a (partial) assignment; T0,T1 are finite, monotonic theories with sharedvariables x0,x′0,x1,x′1, . . . of sort σ .changed← TRUEwhile changed dochanged← FALSEfor Ti ∈ {T0,T1} dostatus,M′← Ti-PROPAGATE(M)if status = FALSE thenstatus is FALSE,M′ ⊆M is a conflict setreturn FALSE,M′elsestatus is TRUE,M′⊇M is a (possibly strengthened) assignmentM←M′for each shared variable x j of Ti dofor each atom (x j ≥ σk) ∈M doif (x′j ≥ σk) /∈M thenchanged← TRUEM←M∪{(x′j ≥ σk)}for each atom (x j ≤ σk) ∈M doif (x′j ≤ σk) /∈M thenchanged← TRUEM←M∪{(x′j ≤ σk)}return TRUE,M54Chapter 5Monotonic Theory of GraphsChapter 4 described a set of techniques for building lazy SMT solvers for finite,monotonic theories. In this chapter, we introduce a set of monotonic graph pred-icates collected into a theory of graphs, as well as an implementation of an SMTsolver for this theory, built using the SMMT framework.Many well-known graph properties — such as reachability, shortest path,acyclicity, maximum s-t flow, and minimum spanning tree weight — are mono-tonic with respect to the edges or edge weights in a graph. We describe our supportfor predicates over these graph properties in detail in Section 5.2.The corresponding graph theory solver for these predicates forms the largestpart of our SMT solver, MONOSAT (described in Appendix A). In Section 5.1, wegive an overview of our implementation, using the techniques described in Chapter4. In Section 5.3, we extend support to bitvector weighted edges, by combining thetheory of graphs with a theory of bitvectors.In later chapters, we will describe how we have successfully applied this theoryof graphs to problems ranging from procedural content generation (Chapter 6.1) toPCB layout (Chapter 6.2), in the latter case solving complex, real-world graph con-straints with more than 1,000,000 nodes — a massive improvement in scalabilityover comparable SAT-based techniques.55Chapter 5. Monotonic Theory of Graphs01e02e1e2 3e3e401e02e1e2 3e3e4reach0,3(E)∧¬reach1,3(E)∧ (¬(e0 ∈ E)∨¬(e1 ∈ E))Figure 5.1: Left: A finite symbolic graph over four nodes and set E of fivepotential edges, along with a formula constraining that graph. A typi-cal instance solved by MONOSAT may have constraints over multiplegraphs, each with hundreds of thousands of edges. Right: A satisfyingassignment (disabled edges dashed, enabled edges solid), correspondingto {(e0 /∈ E),(e1 ∈ E),(e2 /∈ E),(e3 /∈ E),(e4 ∈ E)}.5.1 A Monotonic Graph SolverThe theory of graphs that we introduce supports predicates over several commongraph properties (each of which is monotonic with respect to the set of edges inthe graph). Each graph predicate is defined for a directed graph with a finite set ofvertices V and a finite set of potential edges E ⊆V×V , where a potential edge is anedge that may (or may not) be included in the graph, depending on the assignmentchosen by the SAT solver.1 For each potential edge of E, we introduce an atom(e ∈ E), such that the edge e is enabled in E if and only if theory atom (e ∈ E) isassigned TRUE. In order to distinguish the potential edges e from the edges withedge literals (e ∈ E) that are assigned TRUE, we will refer to an element for which(e ∈ E) is assigned TRUE as enabled in E (or enabled in the corresponding graphover edges E), and refer to an element for which (e ∈ E) is assigned FALSE asdisabled in E, using the notation (e /∈ E) as shorthand for ¬(e ∈ E).Returning to our earlier example of graph reachability (Figure 5.1), the pred-1Many works applying SAT or SMT solvers to graphs have represented the edges in the graph ina similar manner, using a literal to represent the presence of each edge, and sometimes also a literalto control the presence of each node, in the graph. See, e.g., [90, 106, 133, 220].56Chapter 5. Monotonic Theory of Graphsicate reachs,t(E) is TRUE if and only if node t is reachable from node s in graphG, under a given assignment to the edge literals (ei ∈ E). As previously observed,given a graph (directed or undirected) and some fixed starting node s, enablingan edge in E can increase the set of nodes that are reachable from s, but cannotdecrease the set of reachable nodes. The other graph predicates we consider areeach similarly positive or negative monotonic with respect to the set of edges in thegraph; for example, enabling an edge in a graph may decrease the weight of theminimum spanning tree of that graph, but cannot increase it.Theory propagation as implemented in our theory of graphs is described in Al-gorithm 13, which closely follows Algorithm 3. Algorithm 13 representsM+A ,M−Ain the form of two concrete graphs, G− and G+. The graph G− is formed from theedge assignments in MA: only edges e for which the atom (e ∈ E) is in M areincluded in G−. In the second graph, G+, we include all edges for which (e /∈ E)is not inMA.Algorithm 13 makes some minor improvements over Algorithm 3. The first isthat we re-use the data structures for G−,G+ across separate calls to T-Propagate,updating them by adding or removing edges as necessary. As the graphs can bevery large (e.g., in Section 6.2 we will consider applications with hundreds ofthousands or even millions of edges), and there are typically only a few edgeseither added or removed between calls to T-Propagate, this saves a large amount ofredundant effort that would otherwise be caused by repeatedly creating the graphs.A second improvement is to check whether, under the current partial assign-ment, either G− or G+ is unchanged from the previous call to T-Propagate. If thesolver has only enabled edges in E (resp. only disabled edges in E) since the lasttheory propagation call, then the graph G+ (resp. G−) will not have changed, andso we do not need to recompute properties for that graph.Each time an assignment is made to a graph theory atom, Algorithm 13 eval-uates each predicate atom on both the under- and over-approximative graphs G−and G+. The practical performance of this scheme can be greatly improved by us-ing partially or fully dynamic graph algorithms, which can be efficiently updatedas edges are removed or added to G− or G+. Similarly, care should be taken torepresent G− and G+ using data structures that can be cheaply updated as edgesare removed and added. In our implementation, we use an adjacency list in which57Chapter 5. Monotonic Theory of GraphsAlgorithm 13 Theory propagation for the theory of graphs, adapted from Algo-rithm 3. M is a (partial) assignment. E−,E+ are sets of edges; G−,G+ are under-and over-approximate graphs. T-Propagate returns a tuple (FALSE, conflict) ifMis found to be unsatisfiable, and returns tuple (TRUE,M) otherwise.function T-Propagate(M)UpdateBounds:E−←{},E+←{E}for each finite symbolic graph G = (V,E) dofor each edge ei of E doif (ei /∈ E) ∈M thenE+← E+ \{ei}if (ei ∈ E) ∈M thenE−← E−∪{ei}G−← (V,E−),G+← (V,E+)PropagateBounds:for each predicate atom p(E) doIf p is negative monotonic, swap G−,G+ below.if ¬p ∈M thenif evaluate(p,G−) thenreturn FALSE,analyze(p,G−)else if p ∈M thenif not evaluate(p,G+) 7→ FALSE thenreturn FALSE,analyze(¬p,G+)elseif evaluate(p,G−) thenM←M∪{p}else if not evaluate(p,G+) thenM←M∪{¬p}return TRUE,Mevery possible edge of E has a unique integer ID. These IDs map to correspondingedge literals (e ∈ E) and are useful for conflict analysis, in which we often mustwork backwards from graph analyses to identify the theory literals correspond-ing to a subset of relevant edges. Each adjacency list also maintains a history ofadded/removed edges, which facilitates the use of dynamic graph algorithms in thesolver.58Chapter 5. Monotonic Theory of Graphs5.2 Monotonic Graph PredicatesMany of the most commonly used properties of graphs are monotonic with respectto the edges in the graph. A few well-known examples include:1. Reachability2. Acyclicity3. Connected component count4. Shortest s-t path5. Maximum s-t flow6. Minimum spanning tree weightThe first three properties are unweighted, while the remainder operate overweighted graphs and are monotonic with respect to both the set of edges and theweights of those edges. For example, the length of the shortest path from node sto node t in a graph may increase as the length of an edge is increased, but cannotdecrease.These are just a few examples of the many graph predicates that are monotonicwith respect to the edges in the graph; there are many others that are also monotonicand which may be useful (a few examples: planarity testing/graph skew, graphdiameter, minimum global cut, minimum-cost maximum flow, and Hamiltonian-circuit). Some examples of non-monotonic graph properties include Eulerian-circuit (TRUE iff the graph has an Eulerian-circuit) and Graph-Isomorphism (TRUEiff two graphs are isomorphic to each other).For each predicate p in the theory solver we implement three functions:1. evaluate(p,G), which takes a concrete graph (it may be the under-, or theover-approximative graph) and returns TRUE iff p holds in the edges of G,2. analyze(p,G), which takes a concrete graph G in which p evaluates to TRUEand returns a justification set for p, and3. analyze(¬p,G), which takes a concrete graph G in which p evaluates toFALSE and returns a justification set for ¬p.59Chapter 5. Monotonic Theory of GraphsEach of these functions is used by Algorithm 13 to implement theory propaga-tion and conflict analysis. In Sections 5.2.1 and 5.2.2, we describe the implementa-tion of our theory solvers for reachability and acyclicity predicates; subsequently,in Section 5.3.1, we describe our implementation of a theory solver for maximums-t flow over weighted graphs. In Appendix D.1, we list our implementations ofthe remaining supported graph properties mentioned above.5.2.1 Graph Reachability01e02e1e2 3e3e401e0,∞2e1,1e2,1 3e3,1e4,∞Figure 5.2: Left: A finite symbolic graph of edges E under assignment{(e0 ∈ E),(e1 ∈ E),(e2 ∈ E),(e3 /∈ E),(e4 ∈ E)} in which reach0,3(E)holds. Edges e1,e4 (bold) form a shortest-path from node 0 to node 3.Right: The same graph under the assignment {(e0 ∈ E),(e1 /∈ E),(e2 /∈E),(e3 /∈ E),(e4 ∈ E)}, in which reach0,3(E) does not hold. A cut (red)of disabled edges e1,e2,e3 separates node 0 from node 3.The first monotonic graph predicate we consider in detail is reachability in a finitegraph (from a fixed starting node to a fixed ending node). A summary of ourimplementation can be found in Figure 5.3.Both directed and undirected reachability constraints on explicit graphs arisein many contexts in SAT solvers.2 A few applications of SAT solvers in which2As with all the other graph predicates we consider here, this reachability predicate operates ongraphs represented in explicit form, for example, as an adjacency list or matrix. This is in contrastto safety properties as considered in model checking [35, 47, 58], which can be described as testingthe reachability of a node in a (vast) graph represented in an implicit form (i.e., as a BDD or SATformula). Although SAT solvers are commonly applied to implicit graph reachability tasks such asmodel checking and planning, the techniques we describe in this section are not appropriate for modelchecking implicitly represented state graphs; nor are techniques such as SAT-based bounded [35] orunbounded [41] model checking appropriate for the explicit graph synthesis problems we consider.60Chapter 5. Monotonic Theory of GraphsMonotonic Predicate: reachs,t(E) , true iff t can be reached from u in thegraph formed of the edges enabled in E.Implementation of evaluate(reachs,t(E),G): We use the dynamic graphreachability/shortest paths algorithm of Ramalingam and Reps [171]to test whether t can be reached from s in G. Ramalingam-Reps is adynamic variant of Dijkstra’s Algorithm [79]); our implementation fol-lows the one described in [48]. If there are multiple predicate atomsreachs,t(E) sharing the same source s, then Ramalingam-Reps onlyneeds to be updated once for the whole set of atoms.Implementation of analyze(reachs,t(E),G−) Node s reaches t in G−, butreachs,t(E) is assigned FALSE in M (Figure 5.2, left). Let e0,e1, . . .be an s − t path in G−; return the conflict set {(e0 ∈ E),(e1 ∈E), . . . ,¬reachs,t(E)}.Implementation of analyze(¬reachs,t(E),G+) Node t cannot be reachedfrom s in G+, but reachs,t(E) is assigned TRUE. Let e0,e1, . . . be acut of disabled edges (ei /∈ E) separating u from v in G+. Return theconflict set {(e0 /∈ E),(e1 /∈ E), . . . ,reachs,t(E)}.We find a minimum separating cut by creating a graph containing alledges of E (including both the edges of G+ and the edges that are dis-abled in G+ in the current assignment), in which the capacity of eachdisabled edge of E is 1, and the capacity of all other edges is infin-ity (forcing the minimum cut to include only edges that correspond todisabled edge atoms). Any standard maximum s-t flow algorithm canthen be used to find a minimum cut separating s from t (see Figure 5.2,right, for an example of such a cut). In our implementation, we use thedynamic Kohli-Torr [140] minimum cut algorithm for this purpose.Decision Heuristic: (Optional) If reachs,t(E) is assigned TRUE in M, butthere does not yet exist a s−t path in G−, then find a s−t path in G+ andpick the first unassigned edge in that path to be assigned true as the nextdecision. In practice, such a path has typically already been discovered,during the evaluation of reachs,t on G+ during theory propagation.Figure 5.3: Summary of our theory solver implementation for reachs,t .61Chapter 5. Monotonic Theory of Graphsreachability constraints play an integral role include many variations of routingproblems (e.g., FPGA and PCB routing [153, 154, 177, 203] and optical switchrouting [5, 102]), Hamiltonian cycle constraints [133, 220], and product line con-figuration [6, 149, 150]. The SAT-based solver for the Alloy relational modelinglanguage[130] encodes transitive closure into CNF as a reachability constraint overan explicit graph [129, 131].There are several common ways to encode reachability constraints into SATand SMT solvers described in the literature. Some works encode reachability di-rectly into CNF by what amounts to unrolling the Floyd-Warshall algorithm sym-bolically (e.g., [177]); unfortunately, this approach requires O(|E| · |V |2) clausesand hence scales very poorly.SMT solvers have long had efficient, dedicated support for computing congru-ence closure over equality relations[19, 157, 159, 160], which amounts to fast all-pairs undirected reachability detection.3 Undirected reachability constraints can beencoded into equality constraints (by introducing a fresh variable for each vertex,and an equality constraint between neighbouring vertices that is enforced only ifan arc between neighbouring vertices is enabled). The resulting encoding is ef-ficient and requires only O(|V |+ |E|) constraints, however, the encoding is onlyone-sided: it can enforce that two nodes must not be connected, but it cannot en-force that they must be connected. Further, this equality constraint encoding cannotenforce directed reachability constraints at all.While equality constraints cannot encode directed reachability, directed reach-ability constraints can still be encoded into SMT solvers using arithmetic theories.An example of such an approach is described in [90]. This approach introduces anon-deterministic distance variable for each node other than source (which is set toa distance of zero), and then adds constraints for every node other than the sourcenode, enforcing that each node’s distance is one greater than its nearest-to-sourceneighbor, or is set to some suitably large constant if none of its neighbours arereachable from source. This approach requires O(|E| · |V | · log |V |) constraints ifdistances are encoded in bit-blasted bitvectors, or O(|E| · |V |) constraints if dis-We will consider applications related to model checking in Chapter 8.3Note: Some writers distinguish undirected reachability from directed reachability using the termconnectivity for the undirected variant.62Chapter 5. Monotonic Theory of Graphstances are instead encoded in linear or integer arithmetic. However, while thisproduces a concise encoding, it forces the solver to non-deterministically guess thedistance to each node, which in practice seems to work very poorly (see Figure5.4). 4In addition to SAT and SMT solvers, many works have recently used answerset programming (ASP) [25] constraint solvers to enforce reachability constraints.Modern ASP solvers are closely related to CDCL solvers, however, unlike SAT,ASP can encode reachability constraints on arbitrary directed graphs in linearspace, and ASP solvers can solve the resulting formulas efficiently in practice.The ASP solver CLASP [104], which is implemented as an extended CDCL solver,has been particularly widely used to enforce reachability constraints for procedu-ral content generation applications (e.g. [125, 182, 183, 219] all apply CLASP toprocedural content generation using reachability constraints).Considered on its own, the directed reachability predicate implemented inMONOSAT’s theory of graphs scales dramatically better than any of the aboveapproaches for directed reachability, as can be seen in Figure 5.4.5 In Section 6.1,we will see that this translates into real performance improvements in cases wherethe constraints are dominated by a small number of reachability predicates. How-ever, in cases where a large number of reachability predicates are used (and inparticular, in cases where those reachability predicates are mutually unsatisfiableor nearly unsatisfiable), this approach to reachability predicates scales poorly.4If only one-sided directed reachability constraints are required (i.e., the formula is satisfiableonly if node s reaches node t), then more efficient encodings are possible, by having the SAT solvernon-deterministically guess a path. Some examples of one-sided reachability encodings are describedin [106].5All experiments in this chapter were run on a 2.67GHz Intel x5650 CPU (12Mb L3, 96 GbRAM), on Ubuntu 12.04, restricted to 10,000 seconds of CPU time, and 16 GB of RAM.63Chapter 5. Monotonic Theory of Graphs0 5000 10000 15000 20000 25000 30000 35000 40000# of nodes0110100100010000Time to solve (s)MonoSATClaspSATFigure 5.4: Run-times of MONOSAT, SAT, and ASP solvers on randomlygenerated, artificial reachability constraint problems of increasing size.The graphs are planar and directed, with 10% of the edges randomlyasserted to be pair-wise mutually exclusive. In each graph, it is assertedthat exactly one of (the bottom right node is reachable from the topleft), or (the top right node is reachable from the bottom left) holds,using a pair of two-sided reachability constraints. The SAT solver re-sults report the best runtimes found by Lingeling (version ‘bbc’) [34],Glucose-4 [14], and MiniSat 2.2 [84]. The SAT solvers have erraticruntimes, but run out of memory on moderately sized instances. CLASPeventually runs out of memory as well (shown with asterisk), but onmuch larger instances than the SAT solvers. We also tested the SMTsolver Z3 [70] (version 4.3.2, using integer and rational linear arith-metic and bitvector encodings), however, we exclude it from the graphas Z3 timed out on all but the smallest instances.64Chapter 5. Monotonic Theory of GraphsMonotonic Predicate: acyclic(E), true iff there are no (directed) cycles inthe graph formed by the enabled edges in E.Implementation of evaluate(acyclic(E),G): Apply the PK dynamic topo-logical sort algorithm (as described in [164]). The PK algorithm isa fully dynamic graph algorithm that maintains a topological sort of adirected graph as edges are added to or removed from the graph; as aside effect, it also detects directed cycles (in which case no topologicalsort exists). Return FALSE if the PK algorithm successfully produces atopological sort, and return TRUE if it fails (indicating the presence of adirected cycle).Implementation of ¬analyze(acyclic(E),G−) There is a cycle in E, butacyclic(E) is assigned TRUE. Let e0,e1, . . . be the edges that makeup a directed cycle in E; return the conflict set {(e0 ∈ E),(e1 ∈E), . . . ,acyclic(E)}.Implementation of analyze(acyclic(E),G+) There is no cycle in E, butacyclic(E) is assigned FALSE. Let e0,e1, . . . be the set of all edges not inG+; return the conflict set {(e0 /∈ E),(e1 /∈ E), . . . ,¬acyclic(E)}. (Notethat this is the default monotonic conflict set.)Figure 5.5: Summary of our theory solver implementation for acyclic.5.2.2 AcyclicityA second monotonic predicate we consider is acyclicity. Predicate acyclic(E) isTRUE iff the (directed) graph over the edges enabled in E contains no cycles. Theacyclicity of a graph is negative monotonic in the edges of E, as enabling an edgein E may introduce a cycle, but cannot remove a cycle. We describe our theorysolver implementation in Figure 5.5. In addition to the directed acyclicity pred-icate above, there is an undirected variation of the predicate (in which the edgesof E are interpreted as undirected) — however, for the undirected version, the PKtopological sort algorithm cannot be applied, and so we fall back on detecting cy-cles using depth-first search. While depth-first search is fast for a single run, it issubstantially slower than the PK topological sort algorithm for repeated calls asedges are removed or added to the graph.65Chapter 5. Monotonic Theory of GraphsSeveral works have explored pure CNF encodings of acyclicity constraints(e.g., [175] applied acyclicity constraints as part of encoding planning problemsinto SAT, and [65] applied acyclicity in their encoding of Bayesian networks intoSAT). Typical CNF encodings requireO(|E| · |V |) orO(|E| · log |V |) clauses [106].Acyclicity constraints can also be encoded into several existing SMT logics: thetheories of linear arithmetic, integer arithmetic, and difference logic, as well as themore restrictive theory of ordering constraints ( [103]) can all express acyclicity ina linear number of constraints. Recent work [105, 106] has shown that specializedSMT theories directly supporting (one-sided) acyclicity predicates can outperformSAT and arithmetic SMT encodings of acyclicity.In Figure 5.6, we compare the performance of MONOSAT’s acyclicity con-straint to both SAT and SMT encodings. We can see that our approach greatlyout-performs standard SAT encodings, but is only slightly faster than the ded-icated, one-sided acyclicity SMT solvers ACYCGLUCOSE/ACYCMINISAT (de-scribed in [106]). In fact, for the special case of acyclicity constraints that areasserted to TRUE at the ground level, the implementation of the ACYCMINISATsolver, with incremental mode enabled and edge propagation disabled, essentiallymatches our own (with the exception that we detect cycles using a topological sort,rather than depth-first search).Many variations of the implementations of the above two predicates could alsobe considered. For example, the Ramalingam-Reps algorithm we use for reachabil-ity is appropriate for queries in sparsely connected directed and undirected graphs.If the graph is densely connected, then other dynamic reachability algorithms (suchas [116]) may be more appropriate. If there are many reach predicates that do notshare a common source in the formula, then an all-pairs dynamic graph algorithm(such as [75]) might be more appropriate. In many cases, there are also specializedvariations of dynamic graph algorithms that apply if the underlying graph is knownto have a special form (for example, specialized dynamic reachability algorithmshave been developed for planar graphs [117, 194].)The reachability predicate theory solver implementation that we describe hereis particularly appropriate for instances in which most reachability predicates sharea small number of common sources or destinations (in which case multiple pred-icate atoms can be tested by a single call to the Ramalingam-Reps algorithm).66Chapter 5. Monotonic Theory of Graphs0 5000 10000 15000 20000 25000 30000 35000# of nodes000110100100010000Time to solve (s)MonoSATAcycglucoseSATFigure 5.6: Run-times of SAT and SMT solvers on randomly generated, ar-tificial acyclicity constraint problems of increasing size. We considergrids of edges, constrained to be partitioned into two graphs such thatboth graphs are disjoint and acyclic. The graphs are planar and directed,with 0.1% of the edges randomly asserted to be pair-wise mutually ex-clusive. The plain SAT entry represents the best runtimes obtained bythe solvers Lingeling (version ‘bbc’) [34], Glucose-4 [14], or MiniSat2.2 [84]. The SAT solvers run out of memory at 5000 edges, so wedo not report results for SAT for larger instances. The ACYCGLUCOSEentry represents the best runtimes for ACYCGLUCOSE and ACYCMIN-ISAT, with and without pre-processing, incremental mode, and edgepropagation. While both MONOSAT and ACYCGLUCOSE greatly out-perform a plain CNF encoding, the runtimes of MONOSAT and ACY-CGLUCOSE are comparable (with MONOSAT having a slight edge).Two recent works [90, 152] introduced SMT solvers designed for VLSI and clockrouting (both of which entail solving formulas with many reachability or shortestpath predicates with distinct sources and destinations). In [152] the authors pro-vide an experimental comparison against our graph theory solver (as implementedin MONOSAT), demonstrating that their approach scales substantially better thanours for VLSI and clock routing problems. In the future, it may be possible to in-tegrate the sophisticated decision heuristics described in those works into MONO-SAT. In Section 6.2, we also describe a third type of routing problem, escaperouting, for which we obtain state-of-the-art performance using the maximum flowpredicate support that we discuss in the next section.67Chapter 5. Monotonic Theory of Graphs5.3 Weighted Graphs & BitvectorsIn Section 5.2, we considered two predicates that take a single argument (the setof edges, E), operating on an unweighted, directed graph. These two predicates,reachability and acyclicity, have many applications ranging from procedural con-tent generation to circuit layout (for example, in a typical circuit layout application,one must connect several positions in a graph, while ensuring that the wires areacyclic and non-crossing).However, many common graph properties, such as maximum s-t flow, shortests-t path, or minimum spanning tree weight, are more naturally posed as comparisonpredicates operating over graphs with variable edge weights (or edge capacities).6As described in Section 4.5, a monotonic theory over a finite sort σ can be com-bined with another theory over σ , so long as that other theory supports comparisonoperators, with the two theory solvers communicating only through exchanges ofatoms comparing variables to constants: (x ≤ c),(x < c). Extending Algorithm13 to support bitvector arguments in this way requires only minor changes to the-ory propagation in the graph solver (shown in Algorithm 14); it also requires usto introduce a bitvector theory solver capable of efficiently deriving tight compari-son atoms over its bitvector arguments. We describe our bitvector theory solver inAppendix B.5.3.1 Maximum FlowThe maximum s-t flow in a weighted, directed graph, in which each edge of E hasan associated capacity c, is positive monotonic with respect to both the set of edgesenabled in E, and also with respect to the capacity of each edge. We introducea predicate maxFlows,tE,m,c0,c1, . . ., with E a set of edges, and m,c0,c1, . . . fixedwidth bit-vectors, which evaluates to TRUE iff the maximum s-t flow in the directedgraph induced by edges E, with edge capacities c0,c1, . . ., is greater or equal to m.76It is also possible to consider these predicates over constant edge-weight graphs. We describedsuch an approach in [33]; one can recover that formulation from our presentation here by simplysetting each bitvector variable to a constant value.7There is also a variation of this predicate supporting strictly greater than comparisons, but wedescribe only the ‘greater or equal to’ variant. It is also possible to describe this predicate insteadas a function that returns the maximum flow, constrained by a bitvector comparison constraint. Fortechnical reasons this function-oriented presentation is less convenient in our implementation.68Chapter 5. Monotonic Theory of Graphs01e0,1/22e1,1/2e2,1/1 3e3,0/3e4,2/20123∞∞ 1 1211101Figure 5.7: Left: Over-approximate graph G+ for the partial assignment{(e1 ∈ E),(e2 ∈ E),(e3 /∈ E),(e4 ∈ E)}, with (e0 ∈ E) unassigned. Themaximum 0-3 flow in this graph is 2; each edge has its assigned flowand capacity shown as f/c. Right: The cut graph Gcut for this assign-ment, with corresponding minimum 0-3 cut (red): {e3,e4}. Either edgee3 must be included in the graph, or the capacity of edge e4 must beincreased, in order for the maximum flow to be greater than 2.This predicate is positive monotonic with respect to the edge literals and theircapacity bitvectors: Enabling an edge in E can increase but cannot decrease themaximum flow; increasing the capacity of an edge can increase but cannot decreasethe maximum flow. It is negative monotonic with respect to the flow comparisonbitvector m. We summarize our theory solver implementation in Figure 5.8.Maximum flow constraints appear in some of the same routing-related applica-tions of SAT solvers as reachability constraints, in particular when bandwidth mustbe reasoned about; some examples from the literature include FPGA layout [4],virtual data center allocation [212] and routing optical switching networks [5, 102].However, whereas many applications have successfully applied SAT solversto reachability constraints, encoding maximum flow constraints into SAT or SMTsolvers scales poorly (as we will show), and has had limited success in practice.In contrast, constraints involving maximum flows can be efficiently encoded intointeger-linear programming (ILP) and solved using high-performance solvers, suchas CPLEX [64] or Gurobi [162]. Many examples of ILP formulas including flowconstraints (in combination with other constraints) can be found in the literature;examples include PCB layout [93, 123], routing optical switching networks [205,214], virtual network allocation [38, 126], and even air-traffic routing [198].69Chapter 5. Monotonic Theory of GraphsMonotonic Predicate: maxFlows,t(E,m,c0,c1, . . .), true iff the maximum s-tflow in G with edge capacities c0,c1, . . . is ≥ m.Implementation of evaluate(maxFlows,t ,G,m,c0,c1, . . .): We apply the dy-namic minimum-cut/maximum s-t algorithm by Kohli and Torr [140] tocompute the maximum flow of G, with edge capacities set by ci. ReturnTRUE iff that flow is greater or equal to m.Implementation of analyze(maxFlows,t ,G−,m+,c−0 ,c−1 , . . .): The maximums-t flow in G− is f , with f ≥ m+). In the computed flow, eachedge ei is either disabled in G−, or it has been allocated a (possiblyzero-valued) flow fi, with fi ≤ c−i . Let ea,eb, . . . be the edges en-abled in G− with non-zero allocated flows fa, fb, . . .. Either one ofthose edges must be disabled in the graph, or one of the capacities ofthose edges must be decreased, or the flow will be at least f . Returnthe conflict set {(ea ∈ E),(eb ∈ E), . . . ,(ca ≥ fa),(cb ≥ fb) . . . ,(m ≤f ),maxFlows,t(E,m,c0,c1, . . .)}.Implementation of analyze(¬maxFlows,t ,G+,m−,c+0 ,c+1 , . . .) The maxi-mum s-t flow in G+ is f , with f < m− (see Figure 5.7 left). In thecomputed flow, each edge that is enabled in G+ has been allocated a(possibly zero-valued) flow fi, with fi ≤ c+i .If f is a maximum flow in G+, then there must exist a cut of edges in G+whose flow assignments equal their capacity. Our approach to conflictanalysis is to discover such a cut, by constructing an appropriate graphGcut , as described below.Create a graph Gcut (see Figure 5.7 right, for an example of such agraph). For each edge ei = (u,v) in G+, with fi < c+i , add a forwardedge (u,v) to Gcut with infinite capacity, and also a backward edge (v,u)with capacity fi. For each edge ei = (u,v) in G+ with fi = c+i , add aforward edge (u,v) to Gcut with capacity 1, and also a backward edge(v,u) with capacity fi. For each edge ei = (u,v) that is disabled in G+,70Chapter 5. Monotonic Theory of Graphsadd only the forward edge (u,v) to Gcut , with capacity 1.Compute the minimum s-t cut of Gcut . Some of the edges along thiscut may have been edges disabled in G+, while some may have beenedges enabled in G+ with fully utilized edge capacity. Let ea,eb, . . .be the edges of the minimum cut of Gcut that were disabled in G+.Let cc,cd , . . . be the capacities of edges in the minimum cut for whichthe edge was included in G+, with fully utilized capacity. Returnthe conflict set {(ea /∈ E),(eb /∈ E), . . . ,(cc ≤ fc),(cd ≤ fd), . . . ,(m >f ),¬maxFlows,t(E,m,c0,c1, . . .)}.In practice, we maintain a graph Gcut for each maximum flow predicateatom, updating its edges only lazily when needed for conflict analysis.Decision Heuristic: (Optional) If maxFlows,t is assigned TRUE in M, butthere does not yet exist a sufficient flow in G−, then find a maximumflow in G+, and pick the first unassigned edge with non-zero flow to beassigned TRUE as the next decision. If no such edge exists, then pickthe first unassigned edge capacity and assign its capacity to its flowG+, as the next decision. In practice, such a flow is typically alreadydiscovered, during the evaluation of maxFlows,t on G+ during theorypropagation.Figure 5.8: Summary of our theory solver implementation for maxFlows,t .Maximum flow constraints on a graph G= (V,E) in which each edge (u,v)∈ Ehas an associated capacity c(u,v) can be encoded into arithmetic SMT theories intwo parts. The first part of the encoding introduces, for each edge (u,v), a freshbitvector, integer, or real flow variable f (u,v), constrained by the standard networkflow equations [62]:71Chapter 5. Monotonic Theory of Graphs∀u,v ∈V : f (u,v)≤ c(u,v)∀u ∈V/{s, t},∑v∈Vf (u,v) = ∑w∈Vf (v,w)∑v∈Vf (s,v)− ∑w∈Vf (w,s) = ∑w∈Vf (w, t)−∑v∈Vf (t,v)The second part of the encoding non-deterministically selects an s-t cut byintroducing a fresh Boolean variable a(v) for each node v ∈V , with a(v) TRUE iffv is on the source side of the cut. The sum of the capacities of the edges passingthrough that cut are then asserted to be equal to the flow in the graph:a(s)∧¬a(t) (5.1)∑v∈Vf (s,v)− ∑w∈Vf (w,s) = ∑(u,v)∈E,a(u)∧¬a(v)c(u,v) (5.2)By the max-flow min-cut theorem [86], an s-t flow in a graph is equal to ans-t cut if and only if that flow is a maximum flow (and the cut a minimum cut).This encoding is concise (requiring O(|E|+ |V |) arithmetic SMT constraints, orO(|E| · log |V |+ |V | · log |V |) Boolean variables if a bitvector encoding is used);unfortunately, the encoding depends on the solver non-deterministically guessingboth a valid flow and a cut in the graph, and in practice scales very poorly (seeFigure 5.9).72Chapter 5. Monotonic Theory of Graphs0 5000 10000 15000 20000 25000 30000 35000 40000# of nodes0110100100010000Time to solve (s)MonoSATClaspFigure 5.9: Run-times of MONOSAT and CLASP on maximum flow con-straints. The SMT solver Z3 (using bitvector, integer arithmetic, andlinear arithmetic encodings) times out on all but the smallest instance,so we omit it from this figure. We can also see that CLASP is an or-der of magnitude or more slower than MONOSAT.8In these constraints,randomly chosen edge-capacities (between 5 and 10 units) must be par-titioned between two identical planar, grid-graphs, such that the maxi-mum s-t flow in both graphs (from the top left to the bottom right nodes)is exactly 5. This is an (artificial) example of a multi-commodity flowconstraint, discussed in more detail in Section 6.3.2.In Figure 5.9 we compare the performance of MONOSAT’s maximum flowpredicate to the performance of Z3 (reporting the best results from linear arith-metic, integer arithmetic, and bitvector encodings as described above), and to theperformance of CLASP, using a similar encoding into ASP [181].In this case, Z3 is only able to solve the smallest instance we consider withina 10,000 second cutoff. We can also see that the encoding into ASP performssubstantially better than Z3, while also being orders of magnitude slower than theencoding in MONOSAT.8CLASP is unexpectedly unable to solve some of the smallest instances (top left of Figure 5.9).While I can speculate as to why this is, I do not have a definitive answer.73Chapter 5. Monotonic Theory of Graphs5.4 ConclusionIn this chapter we described our implementations of three important graph predi-cates in MONOSAT: reachability, acyclicity, and maximum s-t flow. Each of thesepredicates is implemented following the techniques described in Chapter 4, and foreach we have demonstrated in this chapter state-of-the-art performance on large,artificially generated constraints. In Chapter 6, we will describe a series of appli-cations for these graph predicates, demonstrating that the SMMT framework, asembodied in MONOSAT, can achieve great improvements in scalability over com-parable constraint solvers in realistic, rather than artificial, scenarios. MONOSATalso provides high performance support for several further important graph predi-cates, including shortest path, connected component count, and minimum spanningtree weight. The implementations of the theory solvers for these predicates are sim-ilar to implementations discussed in this chapter, and are summarized in AppendixD.1.74Chapter 5. Monotonic Theory of GraphsAlgorithm 14 Theory propagation, as implemented for the theory of graphs incombination with the theory of bitvectors. M is a (partial) assignment. E−,E+are sets of edges; G−,G+ are under- and over-approximate graphs. T-Propagatereturns a tuple (FALSE, conflict) ifM is found to be unsatisfiable, and returns tuple(TRUE,M) otherwise.function T-Propagate(M)E−←{},E+←{E}for each finite symbolic graph G = (V,E) dofor each edge ei of E doif (ei /∈ E) ∈M thenE+← E+ \{ei}if (ei ∈ E) ∈M thenE−← E−∪{ei}G−← (V,E−),G+← (V,E+)for each bitvector variable x of width n doif (x < 0) ∈M or (x > 2n−1) ∈M then return FALSEM←M∪{(x≥ 0),(x≤ 2n−1)}x−←max({σi|(x≥ σi) ∈M}).x+←min({σi|(x≤ σi) ∈M}).if x− > x+ then return FALSEfor each predicate atom p(E,x0,x1, . . .) doE is a set of edge literals, and xi are bitvectors. If p is negative monotonicin argument E, swap G−,G+ below; if p is negative monotonic in argument xi,swap x−i ,x+i .if ¬p ∈M thenif evaluate(p,G−,x−0 ,x−1 , . . .) thenreturn FALSE,analyze(p,G−,x−0 ,x−1 , . . .)else if p ∈M thenif not evaluate(p,G+,x+0 ,x+1 , . . .) 7→ FALSE thenreturn FALSE,analyze(¬p,G+,x+0 ,x+1 , . . .)elseif evaluate(p,G−,x0−,x1−, . . .) thenM←M∪{p}else if not evaluate(p,G+,x+0 ,x+1 , . . .) thenM←M∪{¬p}return TRUE,M75Chapter 6Graph Theory ApplicationsOver the last several chapters, we have claimed that theory solvers implementedusing the techniques we have described in Chapter 4 can have good performancein practice. Here we describe applications of our theory of graphs (as implementedin our SMT solver MONOSAT, described in Appendix A) to three different fields,along with experimental evidence demonstrating that, for the theory of graphs de-scribed above, we have achieved — and in many cases greatly surpassed — state-of-the-art performance in each of these different domains.The first application we consider, procedural content generation, presents re-sults testing each of the graph predicates described in Chapter 5, demonstratingMONOSAT’s performance across a diverse set of content generation tasks. Thesecond and third applications we consider rely on the maximum flow predicate,and demonstrate MONOSAT’s effectiveness on two industrial applications: circuitlayout and data center allocation. These latter two application scenarios will showthat MONOSAT can effectively solve real-world instances over graphs with hun-dreds of thousands of nodes and edges — in some cases, even instances with morethan 1 million nodes and edges.6.1 Procedural Content GenerationThe first application we consider is procedural content generation, in which artisticobjects, such as landscapes or mazes are designed algorithmically, rather than by76Chapter 6. Graph Theory Applicationshand. Many popular video games include procedurally generated content, lead-ing to a recent interest in content generation using declarative specifications (see,e.g., [37, 183]), in which the artifact to be generated is specified as the solution toa logic formula.Many procedural content generation tasks are really graph generation tasks. Inmaze generation, the goal is typically to select a set of edges to include in a graph(from some set of possible edges that may or may not form a complete graph) suchthat there exists a path from the start to the finish, while also ensuring that when thegraph is laid out in a grid, the path is non-obvious. Similarly, in terrain generation,the goal may be to create a landscape which combines some maze-like propertieswith other geographic or aesthetic constraints.For example, the open-source terrain generation tool Diorama1 considers a setof undirected, planar edges arranged in a grid. Each position on the grid is as-sociated with a height; Diorama searches for a height map that realizes a com-plex combination of desirable characteristics of this terrain, such as the positionsof mountains, water, cliffs, and players’ bases, while also ensuring that all posi-tions in the map are reachable. Diorama expresses its constraints using answerset programming (ASP) [25]. As we described in Section 5.2.1, ASP solvers areclosely related to SAT solvers, but unlike SAT solvers can encode reachabilityconstraints in linear space. Partly for this reason, ASP solvers are more commonlyused than SAT solvers in declarative procedural content generation applications.For instance, Diorama, Refraction [183], and Variations Forever [182] all use ASP.Below, we provide comparisons of our SMT solver MONOSAT against thestate-of-the-art ASP solver CLASP 3.04 [104] (and, where practical, also to MINI-SAT 2.2 [84]) on several procedural content generation problems. These experi-ments demonstrate the effectiveness of our reachability, shortest paths, connectedcomponents, maximum flow, and minimum spanning tree predicates. All experi-ments were conducted on Ubuntu 14.04, on an Intel i7-2600K CPU, at 3.4 GHz(8MB L3 cache), limited to 900 seconds and 16 GB of RAM. Reported runtimesfor CLASP do not include the cost of grounding (which varies between instanta-neous and hundreds of seconds, but in procedural content generation applications1http://warzone2100.org.uk77Chapter 6. Graph Theory ApplicationsFigure 6.1: Generated terrain. Left, a height map generated by Diorama,and right, a cave (seen from the side, with gravity pointing to the bot-tom of the image) generated in the style of 2D platformers. Numbersin the height map correspond to elevations (bases are marked as ‘B’),with a difference greater than one between adjacent blocks creating animpassable cliff. Right, an example Platformer room, in which playersmust traverse the room by walking and jumping — complex movementdynamics that are modeled as directed edges in a graph.is typically a sunk cost that can be amortized over many runs of the solver).Reachability: We consider two applications for the theory of graph reachabil-ity. The first is a subset of the cliff-layout constraints from the terrain generatorDiorama.2The second example is derived from a 2D side-scrolling video game. Thisgame generates rooms in the style of traditional Metroidvania platformers. Reach-ability constraints are used in two ways: first, to ensure that the air and groundblocks in the map are contiguous, and secondly, to ensure that the player’s on-screen character is capable of reaching each exit from any reachable position onthe map. This ensures not only that there are no unreachable exits, but also thatthere are no traps (i.e., reachable locations that the player cannot escape from, suchas a steep pit) in the room. In this instance, although there are many reachabil-2Because we had to manually translate these constraints from ASP into our SMT format, we useonly a subset of these cliff-layout constraints. Specifically, we support the undulate, sunkenBase,geographicFeatures, and everythingReachable options from cliff-layout, with near=1, depth=5,and 2 bases (except where otherwise noted).78Chapter 6. Graph Theory ApplicationsReachability MONOSAT CLASP MINISATPlatformer 16×16 0.8s 1.5s TimeoutPlatformer 24×24 277s Timeout n/aDiorama 16×16 6s < 0.1s TimeoutDiorama 32×32 58.9s 0.2s n/aDiorama 48×48 602.6s 7.9s n/aTable 6.1: Runtime results on reachability constraints in terrain generationtasks using MONOSAT and CLASP. We can see that for the Platformerinstances, which are dominated by reachability constraints from just foursource nodes, MONOSAT is greatly more scalable than CLASP; however,for the Diorama constraints, which contain many reachability predicatesthat do not share common source or destination nodes, CLASP greatlyoutperforms MONOSAT.ity predicates, there are only four distinct source nodes among them (one sourcenode to ensure the player can reach the exit; one source node to ensure the playercannot get trapped, and one source node each to enforce the ‘air’ and ‘ground’connectedness constraints). As a result, the reachability predicates in this instancecollapse down to just four distinct reachability theory solver instances in MONO-SAT, and so can be handled very efficiently. In contrast, the Diorama constraintscontain a large number of reachability predicates with distinct source (and desti-nation) nodes, and so MONOSAT must employ a large number of distinct theorysolvers to enforce them. Example solutions to small instances of the Diorama andplatformer constraints are shown in Figure 6.1.Runtime results in Table 6.1 show that both MONOSAT and CLASP can solvemuch larger instances than MINISAT (for which the larger instances are not evenpractical to encode, indicated as ‘n/a’ in the table). The comparison betweenMONOSAT and CLASP is mixed: On the one hand, CLASP is much faster thanMONOSAT on the undirected Diorama instances. On the other hand, MONOSAToutperforms CLASP on the directed Platformer constraints.Given that ASP supports linear time encodings for reachability constraints andis widely used for that purpose, CLASP’s strong performance on reachability con-straints is not surprising. Below, we combine the Diorama constraints with ad-ditional graph constraints for which the encoding into ASP (as well as CNF) is79Chapter 6. Graph Theory ApplicationsSize Range MONOSAT CLASP MINISAT8×8 8-16 <0.1s <0.1s 9s16×16 16-32 4s 7s >3600s16×16 32-48 4s 23s 2096s16×16 32-64 4s 65s >3600s16×16 32-96 4s >3600s >3600s16×16 32-128 4s >3600s >3600s24×24 48-64 46s 30s >3600s24×24 48-96 61s 1125s >3600s32×32 64-128 196s >3600s >3600sTable 6.2: Runtime results for shortest paths constraints in Diorama. Here,we can see that both as the size of the map is increased, and as the lengthsof the shortest path constraints are increased, MONOSAT outperformsCLASP.non-linear, and in each case MONOSAT outperforms CLASP, as well as MINISATdramatically.Shortest Paths: We consider a modified version of the Diorama terrain genera-tor, replacing the reachability constraint with the constraint that the shortest pathbetween the two bases must fall within a certain range (‘Range’, in Figure 6.2). Wetested this constraint while enforcing an increasingly large set of ranges, and alsowhile testing larger Diorama graph sizes (8× 8,16× 16,24× 24,32× 32). Onecan see that while ASP is competitive with MONOSAT in smaller graphs and withsmaller shortest-path range constraints, both as the size of the shortest path rangeconstraint increases, and also as the size of the graph itself increases, the encodingsof shortest path constraints in both SAT and ASP scale poorly.The two-sided encoding for shortest paths into CNF that we compare to hereis to symbolically unroll the Floyd-Warshall algorithm (similar to encoding two-sided reachability constraints). A shortest path predicate shortestPaths,t(G) ≤ L,which evaluates to TRUE iff the shortest path in G is less than or equal to constantL, can be encoded into CNF using O(L · |E| · |V |) clauses (with E the set of edges,and V the set of vertices, in G). This encoding is practical only for very smallgraphs, or for small, constant values of L (as can be seen in the performance of80Chapter 6. Graph Theory ApplicationsComponents MONOSAT CLASP8 Components 6s 98s10 Components 6s Timeout12 Components 4s Timeout14 Components 0.82s Timeout16 Components 0.2s TimeoutTable 6.3: Runtime results for connected components constraints in Diorama.Here CLASP can solve only the smallest instances.MINISAT in Table 6.2).Whereas ASP solvers have linear space encodings for reachability, the encod-ings of shortest path constraints into ASP are the same as for SAT solvers. Thereare alsoO(|E| · |V | · log |V |) encodings of shortest paths into the theory of bitvectors(see, e.g., [90]), and O(|E| · |V |) encodings into the theories of linear arithmetic orinteger arithmetic, using comparison constraints. However, while these SMT en-codings are concise, they essentially force the solver to non-deterministically guessthe minimum distances to each node in the graph, and perform very poorly in prac-tice (as shown in [90]).Table 6.2 shows large performance improvements over CLASP and MINISAT.Connected Components: We modify the Diorama constraints such that the gen-erated map must consist of exactly k different terrain ‘regions’, where a regionis a set of contiguous terrain positions of the same height. This produces terrainwith a small number of large, natural-looking, contiguous ocean, plains, hills, andmountain regions. We tested this constraint for the 16×16 size Diorama instance,with k= {8,10,12,14,16}. In MONOSAT, this encoding uses two connected com-ponent count predicates: components≥(E,k)∧¬components≥(E,k+ 1), over anadditional undirected graph in which adjacent grid positions with the same terrainheight are connected.The Diorama instances with connected component constraints are significantlyharder for both solvers, and in fact neither solver could solve these instances inreasonable time for Diorama instances larger than 8×8. Additionally, for these in-stances, we disabled the ‘undulate’ constraint, as well as the reachability constraint,81Chapter 6. Graph Theory ApplicationsMaximum Flow MONOSAT CLASP MINISAT8×8, max-flow=16 2s 2s 1s16×16, max-flow=8 9s 483s >3600s16×16, max-flow=16 8s 27s >3600s16×16, max-flow=24 14s 26s >3600s24×24, max-flow=16 81s >3600s >3600s32×32, max-flow=16 450s >3600s >3600sTable 6.4: Runtime results for maximum flow constraints in Diorama. Wecan see that as the instance size increases, MONOSAT greatly outper-forms CLASP.as neither CLASP nor MONOSAT could solve the problem with these constraintscombined with the connected components constraint. Results are presented in Ta-ble 6.3, showing that MONOSAT scales well as the number of required connectedcomponents is increased, whereas for CLASP, the constraints are only practicalwhen the total number of constrained components is small.3Maximum Flow: We modify the Diorama constraints such that each edge hasa capacity of 4, and then enforce that the maximum flow between the top nodesand the bottom nodes of the terrain must be 8,16, or 24. This constraint preventschokepoints between the top and bottom of the map.Maximum flow constraints can be encoded into ASP using the built-in unaryarithmetic support inO(|E| · |V |2) constraints. As discussed in Section 5.3.1, max-imum flow constraints can also be encoded via the theory of bitvectors into pureCNF using O(|E| · log |V |+ |V | · log |V |) constraints. However, neither of theseencodings performs well in practice. In Table 6.4, we show that MONOSAT canhandle maximum flow constraints on much larger graphs that CLASP or MINISAT.In fact, MONOSAT’s maximum flow predicate is highly scalable; we will havemore to say about this in Sections 6.2 and 6.3.Minimum Spanning Trees: A common approach to generating random, tradi-tional, 2D pen-and-paper mazes, is to find the minimum spanning tree of a ran-3The entries begin at 8, as forcing smaller numbers of connected components than 8 was unsatis-fiable.82Chapter 6. Graph Theory ApplicationsFigure 6.2: Random Mazes. Mazes generated through a combination ofminimum spanning tree edge and weight constraints, and a constraint onthe length of the path from start to finish. On the left, an un-optimizedmaze, with awkward, comb-structured walls (circled). On the right, anoptimized maze, generated in seconds by MONOSAT.domly weighted graph. Here, we consider a related problem: generating a randommaze with a shortest start-to-finish path of a certain length.We model this problem with two graphs, G1 and G2. In the first graph, wehave randomly weighted edges arranged in a grid. The random edge weights makeit likely that a minimum spanning tree of this graph will form a visually complexmaze. In the second graph we have all the same edges, but unweighted; these un-weighted edges will be used to constrain the length of the shortest path through themaze. Edges in G2 are constrained to be enabled if and only if the correspondingedges in G1 are elements of the minimum spanning tree of G1. We then enforcethat the shortest path in G2 between the start and end nodes is within some specifiedrange. Since the only edges enabled in G2 are the edges of the minimum spanningtree of G1, this condition constrains that the path length between the start and endnode in the minimum spanning tree of G1 be within these bounds. Finally, we con-strain the graph to be connected. Together, the combined constraints on these twographs will produce a maze with guaranteed bounds on the length of the shortestpath solution.44While one could use the connected component count predicate to enforce connectedness, as weare already computing the minimum spanning tree for this graph in the solver, we can more efficientlyjust enforce the constraint that the minimum spanning tree has weight less than infinity.83Chapter 6. Graph Theory ApplicationsSpanning Tree MONOSAT CLASPMaze 5×5 < 0.1s 15sMaze 8×8 1.5s TimeoutMaze 16×16 32s TimeoutTable 6.5: Runtime results for maze generation using minimum spanning treeconstraints, comparing MONOSAT and CLASP. In these instances, theshortest path through the maze was constrained to a length between 3and 4 times the width of the maze.The solver must select a set of edges in G1 to enable and disable such thatthe minimum spanning tree of the resulting graph is a) connected and b) resultsin a maze with a shortest start-to-finish path within the requested bounds. By it-self, these constraints can result in poor-quality mazes (see Figure 6.2, left, andnotice the unnatural wall formations circled in red); by allowing the edges in G1to be enabled or disabled freely, any tree can become the minimum spanning tree,effectively eliminating the effect of the random edge weight constraints.Instead we convert this into an optimization problem, by combining it with anadditional constraint: that the minimum spanning tree of G1 must be ≤ to someconstant, which we then lower repeatedly until it cannot be lowered any furtherwithout making the instance unsatisfiable.5 This produces plausible mazes (seeFigure 6.2, right), while also satisfying the shortest path constraints, and can besolved in reasonable time using MONOSAT (Figure 6.5).5MONOSAT has built-in support for optimization problems via linear or binary search, andCLASP supports minimization via the “#minimize” statement.84Chapter 6. Graph Theory ApplicationsFigure 6.3: A multi-layer escape routing produced by MONOSAT for the TITMS320C, an IC with a 29×29 ball grid array.6.2 Escape RoutingThe procedural content generation examples provide a good overview of the per-formance of many of the graph predicates described in Chapter 5 on graph-basedprocedural content generation tasks. We next turn to a real-world, industrial ap-plication for our theory of graphs: escape routing for Printed Circuit Board (PCB)layout.In order to connect an integrated circuit (IC) to a PCB, traces must be routedon the PCB to connect each pin or pad on the package of the IC to its appropriatedestination. PCB routing is a challenging problem that has given rise to a largebody of research (e.g., [127, 137, 211]). However, high-density packages withlarge pin-counts, such as ball grid arrays, can be too difficult to route globally in asingle step. Instead, initially an escape routing is found for the package, and onlyafterward is that escape routing connected to the rest of the PCB. Escape routingarises in particular when finding layouts for ball grid arrays (BGAs), which are ICswith dense grids of pins or pads covering an entire face of the IC (Figure 6.3).85Chapter 6. Graph Theory ApplicationsIn escape routing, the goal of connecting each signal pin on the package toits intended destination is (temporarily) relaxed. Instead, an easier initial problemis considered: find a path from each signal pin to any location of the PCB thatis on the perimeter of the IC (and may be on any layer of the PCB). Once suchan escape routing has been found, each of those escaped traces is routed to itsintended destination in a subsequent step. That subsequent routing is not typicallyconsidered part of the escape routing process.Many variants of the escape routing problem have been considered in the liter-ature. Single-layer escape routing can be solved efficiently using maximum-flowalgorithms, and has been explored in many studies [52, 93, 94, 201, 207, 210]; agood survey of these can be found in [208]. [143] uses a SAT solver to performsingle-layer escape routing under additional ‘ordering constraints’ on some of thetraces. SAT and SMT solvers have also been applied to many other aspects of cir-cuit layout (e.g., [59, 82, 88, 89, 152] applied SAT, SMT, and ASP solvers to rec-tilinear or VLSI wire routing, and [90] applied an SMT solver to clock-routing).To the best of our knowledge, we are the first to apply SAT or SMT solvers tomulti-layer escape routing.For our purposes, a printed circuit board consists of one or more layers, withthe BGA connected to the top-most layer. Some layers of the PCB are reservedjust for ground or for power connections, while the remaining layers, sometimesincluding the top-most layer that the package connects to, are routable layers. Sig-nals along an individual layer are conducted by metal traces, while signals crossinglayers are conducted by vias. Typically, vias have substantially wider diametersthan traces, such that the placement of a via prevents the placement of neighbour-ing traces. Different manufacturing processes support traces or vias with differentdiameters; denser printing capabilities can allow for multiple traces to fit betweenadjacent BGA pads (or, conversely, can support tighter spacing between adjacentBGA pads).However, because the placement of vias between layers occludes the placementof nearby traces on those layers, multi-layer escape routing cannot be modeledcorrectly as a maximum flow problem. Instead, multi-layer escape routing hastypically been solved using a greedy, layer-by-layer approach(e.g., [202]). Below,we show how multi-layer escape routing can be modeled correctly by combining86Chapter 6. Graph Theory ApplicationsFigure 6.4: Multi-layer escape routing with simultaneous via-placement. On-grid positions are shown as large nodes, while 45-degree traces passthrough the small nodes. This is a symbolic graph, in which some of thenodes or edges in this graph are included only if corresponding Booleanvariables in an associated formula φ are assigned TRUE. We constructφ such that nodes connecting adjacent layers (the central black node,representing a via) are included in the graph only if the nodes surround-ing the via (marked in gray) are disabled. The via node (center, black)is connected to all nodes around the periphery of the gray nodes, as wellas to the central node (interior to the gray nodes).the maximum flow predicate of Chapter 5 with additional Boolean constraints, andsolved efficiently using MONOSAT.6.2.1 Multi-Layer Escape Routing in MONOSATFigure 6.4 illustrates the symbolic flow graph we use to model multi-layer escaperouting. Each layer of this flow-graph is similar to typical single-layer networkflow-based escape routing solutions (see, e.g., [207]), except that in our graphall positions are potentially routable, with no spaces reserved for vias or pads.Each node in the graph has a node capacity of 1 (with node capacities enforced byintroducing pairs of nodes connected by a single edge of capacity 1).We also include potential vias in the graph, spaced at regular intervals, thatthe solver may choose to include or exclude from the graph. The potential viais shown as a black node in Figure 6.4, with neighbouring nodes shown in gray,indicating that they would be blocked by that via. In Figure 6.4, we show in graythe nodes that would be blocked by a via with a radius roughly 1.5 times the widthof a trace. However, different via widths can be easily supported by simply alteringthe pattern of gray nodes to be blocked by the via.87Chapter 6. Graph Theory Applicationsbbbbbbcc c c c c cbaaaaaaaavia2via3via2→¬aivia3→¬ai(via2∧¬via3)↔ bi(¬via2∧ via3)↔ ciFigure 6.5: Detail from layer 2 of Fig. 6.4, showing some symbolic nodes andedges controlled by Boolean variables (ai,bi,ci,via2,via3) in formula φ .To avoid clutter, the picture shows multiple edges with labels a, b, andc, but formally, each edge will have its own Boolean variable ai, bi,and ci. All nodes and edges have capacity 1, however, nodes and edgeswith associated variables are only included in the graph if their variableis assigned TRUE. Two via nodes are shown, one connecting from thelayer above, and one connecting to the layer below. Nodes in the layerthat are occluded if a via is placed in this position are shown in gray (inthis case, the via has a diameter twice the width of a trace, but any widthof via can be modeled simply by adjusting the pattern of nodes blockedby the via). The first two constraints shown enforce that if either vianode is included in G, then the nodes in the layer that would be occludedby the via (in gray) must be disabled. The remaining constraints allowthe nodes surrounding the blocked nodes to connect to the via if andonly if the via begins or ends at this layer (rather than passing throughfrom an upper to a lower layer). These constraints are included in φ foreach potential via location at each layer in G.Each via node is connected to the nodes surrounding the gray nodes that areblocked by the via, as well as the central node interior to the gray nodes (see Figure6.5). These represent routable connections on the layer if the via is placed, allowingtraces to be routed from the via nodes to the nodes surrounding the blocked nodes,or allowing the via to route through the layer and down to the next layer below.Each via node is associated with a Boolean variable (via2 in Figure 6.5), suchthat the via node is included in the graph if and only if via2 is TRUE in φ . Thepotentially blocked nodes around each via are also associated with variables (forclarity, drawn as a in Figure 6.5, however in φ each edge will actually have a uniquevariable ai). For each via, we include constraints via→¬ai in φ , disabling all theimmediate neighbouring nodes if the via is included in the graph. Any satisfiable88Chapter 6. Graph Theory Applicationsassignment to these constraints in φ selects a subset of the nodes of G representinga compatible, non-overlapping set of vias and traces.Four configurations are possible for the two via nodes shown in Figure 6.5: (a)neither via node is enabled, allowing traces to be routed through the gray nodeson this layer, (b) the via enters from above, and connects to this layer, (c) the viabegins at this layer, connecting to a layer below, and (d) the via passes through thislayer, connecting the layer above to the layer below. By adding constraints restrict-ing the allowable configurations of vias, as described in Figure 6.6, our approachcan model all the commonly used via types: through-hole vias, buried vias, blindvias, or any-layer micro-vias. With minor adjustments to the constraints, these dif-ferent via types can be combined into a single model or can be restricted to specificlayers of the PCB, allowing a wide variety of PCB manufacturing processes to besupported.We add an additional source node s and sink node t to the graph, with di-rected, capacity-1 edges connecting the s to each signal in the graph, and directed,capacity-1 edges connecting all nodes on the perimeter of each layer to t. Finally,a single flow constraint max f lows,t(E) ≥ |signals| is added to φ , ensuring that inany satisfying assignment, the subset of edges included in the graph must admit aflow corresponding to a valid escape routing for all of the signal pins.A solution to this formula corresponds to a feasible multi-layer escape routing,including via and trace placement. However, as MONOSAT only supports maxi-mum flow constraints, and not minimum-cost maximum flow constraints, the tracerouting in this solution is typically far from optimal (with traces making completelyunnecessary detours, for example). For this reason, once MONOSAT has produceda feasible escape routing, including a placement of each via, we then apply an off-the-shelf minimum-cost maximum flow solver [165] to find a corresponding locallyoptimal trace routing for each individual layer of the feasible routing. This can besolved using completely standard linear-programming encodings of minimum-costmaximum flow, as the vias are already placed and the layer that each signal is to berouted on is already known.89Chapter 6. Graph Theory ApplicationsVia Type ConstraintThrough-hole via j→ (¬a1i ∧¬a2i ∧ . . .¬ani )Blind via j→ (¬a1i ∧¬a2i ∧ . . .¬a ji )Buried via j→ (¬asi ∧¬as+1i ∧ . . .¬ati)Micro —Figure 6.6: Constraints enforcing different via models in φ , for via1 . . .vian,where variables aki control the potentially blocked nodes of via j. (Forvariable aki , the index k indicates the layer number, and the constraint isenforced for all values of i of neighbouring nodes to the via.) Through-hole vias are holes drilled through all layers of the PCB; the corre-sponding constraints block the placement of traces at the position ofthe through-hole on all layers of the PCB. Buried and blind vias drillthrough a span of adjacent layers, with blind vias always drilling allthe way to either the topmost or bottom layer of the PCB (we show thecase for blind vias starting at the top layer, and continuing to layer j,above). In the constraints for buried vias, s and t are the allowable startand end layers for the buried via, determined by the type of buried via.Micro-vias allow any two adjacent layers to be connected and require noadditional constraints (the default behaviour); if only a subset of the lay-ers support micro-vias, then this can be easily enforced. Each of thesevia types can also be combined together in one routing or (exceptingthrough-holes) restricted to a subset of the layers.6.2.2 EvaluationWe evaluate our procedure on a wide variety of dense ball grid arrays from fourdifferent companies, ranging in size from a 28×28 pad ARM processor with 382routable signal pins to a 54×54 pad FPGA with 1755 routable signal pins. Theseparts, listed in Tables 6.6 and 6.7, include 32-bit and 64-bit processors, FPGAs, andSoCs. The first seven of these packages use 0.8mm pitch pads, while the remainderuse 1mm pitch pads. Each part has, in addition to the signal pins to be escaped, aroughly similar number of power and ground pins (for example, the 54×54 XilinxFPGA has 1137 power and ground pins, in addition to the 1755 signal pins). Mostparts also have a small number of disconnected pins, which are not routed at all.Typically, power and ground pins are routed to a number of dedicated power andground layers in the PCB, separately from the signal traces; we assume that the90Chapter 6. Graph Theory Applicationsbottom-most layers of the PCB contain the power and ground layers, and route allpower and ground pins to those layers with through-hole vias. This leaves only theroutable signal pins to be escaped in each part on the remaining layers.For comparison, we implemented a simple network-flow-based single-layerescape routing algorithm, similar to the one described in [207]. We then imple-mented a greedy, layer-by-layer router by routing as many signals as possible on thetop-most layer (using maximum flow), and, while unrouted signals remain, addinga new layer with vias connecting to each unrouted signal. This process repeats untilno unrouted signals remain. As can be seen in Table 6.6, this layer-by-layer routingstrategy is simple but effective, and has been previously suggested for multi-layerescape routing in several works in the literature (for example, [202] combinesthis strategy with their single-layer routing heuristic to create a multi-layer escaperouting method).In Table 6.6, we compare our approach to the layer-by-layer strategy usingblind vias, and, in Table 6.7, using through-hole vias. All experiments were run ona 2.67GHz Intel x5650 CPU (12Mb L3, 96 Gb RAM), in Ubuntu 12.04. Althoughour approach supports buried and micro-vias, we found that all of these instancescould be solved by MONOSAT with just 2 or 3 signal layers, even when using themore restrictive through-hole and blind via models, and so we omit evaluationsfor these less restrictive models (which, in two or three layer PCBs, are nearlyequivalent to blind vias).In Tables 6.6 and 6.7, MONOSAT finds many solutions requiring fewer layersthan the layer-by-layer strategy (and in no case requires more layers than the layer-by-layer approach). For example, in Table 6.6, MONOSAT finds a solution usingblind vias (for the TI AM5K2E04 processor, packaged in a dense, 33× 33 pad,0.8mm pitch BGA) which requires only 3 signal layers, whereas the layer-by-layerapproach requires 4 signal layers for the same part. In this case, MONOSAT wasalso able to prove that no escape routing using 2 or fewer layers was possible forthis circuit (assuming the same grid-model is used). In Table 6.7, using more re-strictive through-vias, there are several examples where MONOSAT finds solutionsusing 1 or even 2 fewer signal layers than the layer-by-layer approach.The runtimes required by MONOSAT to solve these instances are reasonable,spanning from a few seconds to a little less than 2 hours for the largest instance91Chapter 6. Graph Theory ApplicationsPart Size Layer-by-Layer MONOSATLayers Time (s) Layers Time (s)TI AM5716 28×28 3 5.4s + 63.6s 2* 54.4s + 81.2sTI AM5718 28×28 3 5.4s + 70.4s 2* 47.4s + 118.4sTI AM5726 28×28 3 5.3s + 49.7s 3 53.3s + 492.8sTI AM5728 28×28 3 5.3s + 48.6s 3 53.8s + 387.2sTI TMS320C 29×29 4 7.7s + 81.5s 3 75.0s + 497.7sTI AM52E02 33×33 4 9.5s + 107.7s 3 103.8 + 921.8TI AM5K2E04 33×33 4 9.2s + 96.8s 3* 114.7s + 962.0sTI 66AK2H1 39×39 3 24.0s + 508.0s 2* 338.4s + 878.1sLattice M25 32×32 2* 10.4s + 160.5s 2* 140.1s + 306.7sLattice M40 32×32 3 114.6s + 205.6s 2* 194.0s + 364.4sLattice M40 34×34 3 18.5s + 300.0s 2* 254.3s + 425.2sLattice M80 34×34 3 17.8s + 266.5s 2* 411.3s + 505.8sLattice M80 42×42 3 27.0s + 499.0s 3 810.3s + 882.4sLattice M115 34×34 3 16.9s + 274.1s 2* 392.9s + 578.2sLattice M115 42×42 3 27.4s + 461.4s 3 242.5s + 254.7sAltera 10AX048 28×28 2* 8.0s + 109.3s 2* 85.1s + 183.4sAltera 10AX066 34×34 2* 13.1s + 218.3s 2* 151.1s + 371.2sAltera 10AX115 34×34 2* 13.9s + 286.8s 2* 168.5s + 501.9sAltera 10AT115 44×44 3 29.6s + 579.5s 3 384.8s + 928.8sAltera EP4S100 44×44 3 31.0s + 698.6s 3 401.8s + 1154.6sXilinx XCVU160 46×46 3 34.3s + 617.9s 3 414.3s + 977.5sXilinx XCVU440 49×49 4 52.2s + 1167.5s 3* 1246.9s + 2133.7sXilinx XCVU440 54×54 4 60.9s + 1438.1s 3* 1597.3s + 2726.9sTable 6.6: Multi-layer escape routing with blind vias. Run-times are re-ported as a+b, where a is the time to find a feasible multi-layer routing,and b is the time to post-process that feasible solution using minimum-cost maximum flow routing. Boldface highlights when our approach re-quired fewer layers; solutions that use a provably minimal number oflayers are marked with *. Length shows the average trace length in mm.These solutions ignore differential pair constraints (routing differentialsignals as if they were normal signals).considered. These runtimes are all the more impressive when considering that thegraphs encoded in some of these instances are very large by SAT solver standards,with more than 1,000,000 nodes (and nearly 4,000,000 edges) in the formula forthe 54×54 Xilinx FPGA.92Chapter 6. Graph Theory ApplicationsPart Size Layer-by-Layer MONOSATLayers Time (s) Layers Time (s)TI AM5716 28×28 3 5.4s + 58.9s 2* 35.8s + 62.9sTI AM5718 28×28 3 5.1s + 61.8s 2* 46.0s + 69.1sTI AM5726 28×28 4 7.7s + 63.1s 3 97.6s + 80.2sTI AM5728 28×28 4 6.8s + 56.7s 3 110.1s + 85.4sTI TMS320C 29×29 5 10.9s + 106.1s 3 109.6s + 128.4sTI AM52E02 33×33 5 13.4s + 133.3s 3 203.2s + 180.4sTI AM5K2E04 33×33 5 12.7s + 125.4s 3* 291.3s + 187.8sTI 66AK2H1 39×39 3 24.9s + 495.1s 2* 347.9s + 510.7sLattice M25 32×32 2* 10.7s + 143.7s 2* 132.3s + 278.7sLattice M40 32×32 3 15.3s + 183.1s 2* 161.5s + 311.3sLattice M40 34×34 3 18.0s + 270.2s 2* 183.9s + 405.6sLattice M80 34×34 3 17.7s + 238.4s 3 304.8s + 638.2sLattice M80 42×42 3 26.2s + 478.4s 3 810.3s + 882.4sLattice M115 34×34 3 16.8s + 227.1s 3 364.2s + 358.7sLattice M115 42×42 3 27.8s + 457.3s 3 945.9s + 1500.3sAltera 10AX048 28×28 2* 8.2s + 98.2s 2* 109.2s + 115.8sAltera 10AX066 34×34 2* 14.0s + 212.3s 2* 203.5s + 282.9sAltera 10AX115 34×34 2* 13.3s + 235.1s 2* 198.2s + 293.7sAltera 10AT115 44×44 3 28.9s + 455.2s 3 616.1s + 992.9sAltera EP4S100 44×44 3 28.5s + 589.2s 3 733.5s + 834.5sXilinx XCVU160 46×46 3 32.5s + 538.3s 3 646.6s + 1216.4sXilinx XCVU440 49×49 4 53.7s + 1051.1s 3 3457.5s + 1284.5sXilinx XCVU440 54×54 4 600s + 1373.6s 3 6176.9s + 1861.9sTable 6.7: Multi-layer escape routing with through-vias. Run-times are re-ported as a+b, where a is the time to find a feasible multi-layer routing,and b is the time to post-process that feasible solution using minimum-cost maximum flow routing. Boldface highlights when our approach re-quired fewer layers; solutions that use a provably minimal number oflayers are marked with *. Length shows the average trace length in mm.These solutions ignore differential pair constraints (routing differentialsignals as if they were normal signals).93Chapter 6. Graph Theory Applications6.3 Virtual Data Center AllocationThe final application we consider is virtual data center allocation. Virtual datacenter allocation [26, 113] is a challenging network resource allocation problem,in which instead of allocating a single virtual machine to the cloud, a connectedvirtual data center (VDC) consisting of several individual virtual machines mustbe allocated simultaneously, with guaranteed bandwidth between some or all ofthe virtual machines.While allocating individual VMs to a data center is a well-studied problem, al-locating virtual data centers remains an open research problem, with current com-mercial approaches lacking end-to-end bandwidth guarantees [1–3]. Solutions thatdo provide end-to-end bandwidth guarantees lack scalability [212], are restrictedto data centers with limited or artificial topologies [22, 176, 212], or are incom-plete [113], meaning that they may fail to find allocations even when feasible al-locations exist, especially as load increases, resulting in under-utilized data centerresources.As we will show, we can formulate the VDC allocation problem as a multi-commodity flow problem, and solve it efficiently using a conjunction of maximumflow predicates from our theory of graphs. Using this approach, MONOSAT canallocate VDCs of up to 15 VMs to physical data centers with thousands of servers,even when those data centers are nearly saturated. In many cases, MONOSAT canallocate 150%− 300% as many total VDCs to the same physical data center asprevious methods.6.3.1 Problem FormulationFormally, the VDC allocation problem6 is to find an allocation of VMs to servers,and links in the virtual network to links in the physical network, that satisfies thecompute, memory and network bandwidth requirements of each VM across theentire data center infrastructure, including servers, top-of-rack (ToR) switches andaggregation switches.The physical network consists of a set of servers S, switches N, and a di-6There are a number of closely related formalizations of the VDC allocation problem in theliterature; here we follow the definition in [212].94Chapter 6. Graph Theory Applicationsrected graph (S∪N,L), with capacities c(u,v) for each link (u,v) ∈ L. The virtualdata center consists of a set of virtual machines V M and a set of directed band-width requirements R ⊆ V M×V M×Z+. For each server s ∈ S, we are givenCPU core, RAM, and storage capacities cpu(s),ram(s),storage(s), and each vir-tual machine v ∈ V M has corresponding core, RAM, and storage requirementscpu(v),ram(v),storage(v).Given a PN and VDC defined as above, the multi-path VDC allocation problemis to find an assignment A : V M 7→ S of virtual machines v ∈ V M to servers s ∈ S,and, for each bandwidth requirement (u,v,b) ∈ R, an assignment of non-negativebandwidth Bu,v(l) to links l ∈ L, such that the following sets of constraints aresatisfied:(L) Local VM allocation constraints. These ensure that each virtual ma-chine is assigned to exactly one server in the physical network (multi-ple VMs may be assigned to each server), and that each server has suf-ficient CPU core, RAM, and storage resources available to serve the sumtotal of requirements of the VMs allocated to it. Let V (s) = {v ∈ V M |A(v) = s}; then ∀s ∈ S : ∑V (s) cpu(v)≤ cpu(s)∧∑V (s) ram(v)≤ ram(s)∧∑V (s) storage(v)≤ storage(s). We model resource requirements using inte-ger values, and assume that no sharing of resources between allocated VMsis allowed.(G) Global bandwidth allocation constraints. These ensure that sufficientbandwidth is available in the physical network to satisfy all VM to VMbandwidth requirements in R simultaneously. Formally, we require that∀(u,v,b) ∈ R, the assignments Bu,v(l) form a valid A(u)−A(v) network flowgreater or equal to b, and that we respect the capacities of each link l in thephysical network: ∀l ∈ L :∑(u,v,b)∈R Bu,v(l)≤ c(l). We model bandwidths us-ing integer values and assume that communication bandwidth between VMsallocated to the same server is unlimited.Prior studies [54, 114, 193, 209] observed that if path-splitting is allowed, thenthe global bandwidth allocation constraints correspond to a multi-commodity flowproblem, which is NP-complete even for undirected integral flows [92], but has95Chapter 6. Graph Theory Applicationspoly-time solutions via linear programming if real-valued flows are allowed.7Since multi-commodity flow can be reduced to the multi-path VDC allocationproblem, for the case of integer-valued flows, the multi-path VDC allocation prob-lem is NP-hard [54].Next, we will show how multi-commodity integral flow problems can be en-coded as a conjunction of maximum flow constraints over graphs with symbolicedge weights. We will then provide a solution to the full multi-path VDC allo-cation problem by combining our multi-commodity flow encoding for global con-straints G with a pseudo-Boolean encoding of local constraints L.6.3.2 Multi-Commodity Flow in MONOSATWe model multi-path, end-to-end bandwidth guarantees in MONOSAT as a multi-commodity flow problem. In this subsection, we describe how we model integer-value multi-commodity flow in terms of the built-in maximum flow predicates thatMONOSAT supports; in the next subsection, we show how to use these multi-commodity flow constraints to express VDC allocation.Before we introduce our encoding for integer-value multi-commodity flow, it ishelpful to provide some context. SMT solvers have not traditionally been applied tolarge multi-commodity flow problems; rather multi-commodity flow problems areusually solved using integer-arithmetic solvers, or are approximated using linearprogramming. MONOSAT does not directly provide support for multi-commodityflows, but as we will show below, by expressing multi-commodity flows as a con-junction of single-commodity maximum flow predicates (which MONOSAT doessupport), we can use MONOSAT to solve large multi-commodity flow problems –a first for SMT solvers.We consider this formulation of integer-value multi-commodity flows in termsof combinations of maximum flow predicates to be a key contribution of this sec-tion. While there are many obvious ways to encode multi-commodity flows in SMTsolvers, the one we present here is, to the best of our knowledge, the only SMT en-7Note that while linear programming supports global bandwidth constraints, it does not supportthe local server constraints. Therefore, approaches that model the global constraints as a linearprogram either include additional steps to perform local server allocation [209], or use mixed integerprogramming [54].96Chapter 6. Graph Theory Applicationscoding to scale to multi-commodity flow problems with thousands of nodes. Asthere are many applications to which SMT solvers are better suited than integerarithmetic solvers (and vice-versa), this SMT formulation has many potential ap-plications beyond virtual data center allocation.Given a directed graph G = (V,E), an integer capacity c(u,v) for each edge(u,v) ∈ E, and a set of commodity demands K, where a commodity demand i ∈ Kis a tuple (si, ti,di), representing an integer flow demand of di between source si ∈Vand target ti ∈ V . The integral multi-commodity flow problem is to find a feasibleflow such that each demand di is satisfied, while for each edge (u,v) the total flowof all capacities (summed) is at most c(u,v):fi(u,v)≥ 0, ∀(u,v) ∈ E, i ∈ K∑i∈Kfi(u,v)≤ c(u,v), ∀(u,v) ∈ E∑v∈Vfi(u,v)−∑v∈Vfi(v,u) =0, if u /∈ {si, ti}d, if u = si−d, if u = ti,∀i ∈ KTo encode multi-commodity integral flow constraints in MONOSAT, we in-stantiate directed graphs G1..|K| with the same topology as G. For each edge(u,v)i ∈ Gi, we set its capacity to be a fresh bitvector c(u,v)i, subject to the con-straint 0 ≤ c(u,v)i ≤ c(u,v). We then assert for each edge (u,v) that ∑i c(u,v)i ≤c(u,v) – that is, the capacities of each edge (u,v) in each commodity graph Gimust sum to at most the original capacity of edge (u,v). Finally, for each commod-ity demand (si, ti,di), we assert that the maximum si–ti flow in Gi is ≥ di, usingMONOSAT’s built-in maximum flow constraints.If the multi-commodity flow is feasible, the solver will find a partitioning of thecapacities among the graphs Gi such that the maximum si–ti flow in Gi is at leastdi for each commodity constraint i. We can then force each commodity flow in thesolution to be exactly di by adding an extra node ni to each graph Gi, an edge (ti,ni)with capacity di, and replacing the commodity demand (si, ti,di) with (si,ni,di).97Chapter 6. Graph Theory Applications6.3.3 Encoding Multi-Path VDC AllocationThe local constraints of the multi-path VDC allocation problem can be modeledas a set of pseudo-Boolean constraints, for which many efficient and direct encod-ings into propositional satisfiability (SAT) are known [21, 85]. The part of con-straint set L (see Section 6.3.2) enforcing that each VM is assigned to at most oneserver is a special case of a so-called ‘at-most-one’ pseudo-Boolean constraint [53],which can be handled even more efficiently (in fact, MONOSAT has built-in the-ory support for large ‘at-most-one’ constraints). Constraint set G can be encodedas a multi-commodity flow as described above, with up to |V M|2 commodity de-mands (one for each bandwidth tuple (u,v,bandwidth) ∈ R). However, we cangreatly improve on this by grouping together bandwidth constraints that sharea common source, and merging them into a single commodity demand: Givena set of bandwidth constraints (u,vi,bandwidthi) ∈ R with the same source u,we can convert these into a single commodity demand by adding an extra nodew 6∈ V M, along with edges (vi,w) with capacity bandwidthi. The commodity de-mands (u,vi,bandwidthi) can then be replaced by a single commodity demand(u,w,∑i bandwidthi).Since there are at most |V M| distinct sources in R, this reduces the number ofcommodity demands from |V M|2 in the worst case to |V M|. In cases where theVDC is undirected, we can improve on this further, by swapping sources and sinksin communication requirements so as to maximize the number of requirementswith common sources. To do so, we construct the undirected graph of communica-tion requirements, with an undirected edge of weight (u,v) = bandwidth for eachbandwidth requirement, and find an approximate minimum-cost vertex cover (us-ing the 2-opt approximation from [24]). This can be done efficiently (in polynomialtime) even for large networks. Necessarily, each edge, and hence each communi-cation requirement, will have at least one covering vertex. For each requirement(u,v,bandwidth), if v is a covering vertex and u is not, we replace the requirementwith (v,u,bandwidth), swapping u and v. After swapping all un-covered sourcevertices in this way, we then proceed to merge requirements with common sourcesas above. For cases where the VDC is directed, we skip this cover-finding opti-mization, and only merge together connection requirements that happen to have98Chapter 6. Graph Theory Applicationsthe same (directed) source in the input description.Given this optimized set of commodity demands, we construct a directed graphG consisting of the physical network (S∪N,L), and one node for each virtualmachine in V M. If any VDC communication requirements (u,vi,bandwidthi) havebeen merged into combined requirements (u,w,∑bandwidthi) as above, we addadditional, directed edges (vi,w) with capacity bandwidthi to G.For each v ∈ V M and each server s ∈ S, we add a directed symbolic edge evsfrom v to s with unlimited capacity to G; this edge controls the server to which eachVM is allocated. Note that only the VM allocation edges evs have to be symbolic;all remaining edges in G have known constant capacities and can be asserted to bein G.We assert (using MONOSAT’s theory of pseudo-Boolean constraints, as de-scribed in Algorithm 5 of Chapter 4) that for each VM v, exactly one edge evs isenabled, so that the VM is allocated to exactly one server: ∀v ∈V M : ∑s evs =1. For each server s, we assert ∑v cpu(v) ≤ cpu(s)∧∑v RAM(v) ≤ RAM(s)∧∑v storage(v)≤ storage(s), i.e. that the set of VMs allocated to each server (whichmay be more than one VM per server) will have sufficient CPU core, RAM, andstorage resources available on that server. Together these assertions enforce con-straint set L from our problem definition.Finally, using the multi-commodity flow encoding described above, we assertthat the multi-commodity flow in G satisfies (u,v,bandwidth) for each optimizedcommodity requirement. When constructing the individual commodity flow graphsGi from G, we add an assertion that each VM-server allocation edge evs is containedin Gi if, and only if, that edge is included in G; this ensures that the same VM-serverallocation edges are enabled in each commodity flow graph.6.3.4 EvaluationWe compare the performance of MONOSAT to that of two previous VDC tools:SecondNet’s VDCAlloc algorithm [113] — the seminal VDC allocation toolwith sound, end-to-end bandwidth guarantees — and the Z3-based abstraction-refinement technique from [212], the tool most similar to our own contribution.99Chapter 6. Graph Theory ApplicationsSecondNet’s VDCAlloc algorithm [113]. SecondNet’s VDCAlloc algorithm(‘SecondNet’, except where ambiguous) is an incomplete, heuristic-driven algo-rithm based on bipartite matching, that is fast (much faster than MONOSAT)and scales well to physical networks with even hundreds of thousands of servers(whereas MONOSAT scales only to a few thousand servers).However, while SecondNet scales very well, it also has major limitations. Asit is based on bipartite matching, it fundamentally cannot allocate more than oneVM in each VDC to any given server. SecondNet also performs allocation in anincomplete, greedy fashion: it commits to a node allocation before attempting linkmapping, and it maps links from the virtual network one at a time. In heavilyutilized networks, this greedy process can fail to find a feasible allocation of avirtual data center, even when a feasible allocation exists. Below we will showthat in many realistic circumstances, SecondNet allocates less than half of the totalfeasible allocations, and sometimes less than a third.Abstraction-refinement technique based on Z3 [212]. The authors of [212]introduced two approaches for performing single-path VDC allocation with band-width guarantees that use the SMT solver Z3 [70]. Unlike the SMT solver MONO-SAT used by MONOSAT, Z3 has no built-in support for graph predicates. There-fore, a major challenge tackled by [212] was to efficiently represent the globalbandwidth and connectivity constraints in the low-level logic of Z3. For thespecial case of physical data-centers with proper tree topologies, the authors intro-duced an efficient encoding of these constraints that can be represented in Z3 usingonly a linear number of constraints in the number of servers, which they show per-forms well enough to scale to several hundred servers (but can only be applied totree topologies).The first approach from [212](which we call Z3-generic) uses an encoding thatcan handle any data center topology, but as was shown in [212], scales extremelypoorly. The second approach (which we call Z3-AR) is an optimized abstraction-refinement technique with Z3 as its back-end solver. This approach is more scal-able than the generic encoding, but is restricted to data centers with tree topologies.In our experiments we found that Z3-generic performed poorly, often failing to findany allocations within a 1-hour timeout. For brevity we do not report results forZ3-generic.100Chapter 6. Graph Theory ApplicationsVDC Instance Structurevn3.1 vn3.2 vn3.3 vn5.1 vn5.2 vn5.3#VDC Time #VDC Time #VDC Time #VDC Time #VDC Time #VDC TimeTree Physical Data Center, 200 servers with 4 cores eachSecondNet† 88 < 1 88 < 1 87 < 1 48 2.1 51 2.2 52 2.2[212]-AR 88 17.9 88 33.5 88 34.7 53 61.9 53 59.9 53 55.1MONOSAT 88 4.7 88 6.8 88 6.3 53 7.3 53 7.1 53 9.0Tree Physical Data Center, 200 servers with 16 cores eachSecondNet† 313 < 1 171 < 1 129 < 1 56 2.6 59 2.9 57 2.7[212]-AR 355 76.0 301 (3600) 300 (3600) 88 (3600) 50 (3600) 24 (3600)MONOSAT 355 15.1 353 22.4 352 20.45 201 22.1 201 22.6 202 25.6Tree Physical Data Center, 400 servers with 16 cores eachSecondNet† 628 2.1 342 1.0 257 1.2 109 18.6 117 21.5 114 18.7[212]-AR 711 209.9 678 (3600) 691 3547.9 72 (3600) 43 (3600) 77 (3600)MONOSAT 711 55.4 709 86.44 705 78.6 404 90.3 405 85.4 405 99.8Tree Physical Data Center, 2000 servers with 16 cores eachSecondNet† 3140 48.3 1712 23.8 1286 30.7 539 2487.7 582 2679.0 567 2515.5[212]-AR 3555 2803.7 741 (3600) 660 (3600) 76 (3600) 86 (3600) 204 (3600)MONOSAT 3554 1558.2 3541 2495.7 3528 2375.1 958 (3600) 1668 (3600) 1889 (3600)Table 6.8: Total Number of Consecutive VDCs Allocated on Data Centerswith Tree Topologies. The six major columns under “VDC InstanceStructure” give results for serial allocation of the six different VDC typesfrom [212]. ‘#VDC’ is the number of VDCs allocated until failure tofind an allocation, or until the 1h timeout; ‘Time’ is the total runtime forall allocations, in seconds; ‘Mdn’ is the median runtime per allocation,in seconds. ‘(3600)’ indicates that allocations were stopped at the time-out. Largest allocations are in boldface. In this table, ‘SecondNet†’ isan implementation of SecondNet’s VDCAlloc algorithm [113], modifiedto handle tree topologies by the authors of [212]. MONOSAT is muchfaster than [212]-AR but slower than SecondNet. In most cases where itdoes not time out, MONOSAT is able to allocate several times as manyVDCs as SecondNet, in a reasonable amount of time per VDC (usuallyless than one second per VDC, except on the largest instances).To ensure a fair comparison, we used the original implementations of thesetools, with minor bug fixes and latest enhancements, obtained from the originalauthors.101Chapter 6. Graph Theory ApplicationsComparison on Trees from [212] Our first experiment reproduces and extendsan experiment from [212], in which a series of identically structured VDCs areallocated one-by-one to tree-structured data centers until the solver is unable tomake further allocations (or a timeout of 1 CPU hour is reached). We obtained theoriginal implementation Z3-AR from the authors for this experiment, along with aversion of SecondNet they implemented with support for the tree-structured datacenters considered here. In this experiment, the VDCs being allocated always haveidentical structure; this is a limitation introduced here for compatibility with thesolvers from [212]. In our subsequent experiments, below, we will consider morerealistic allocation scenarios. Except where noted, all experiments were conductedon a 2.66GHz (12Mb L3 cache) Intel x5650 processor, running Ubuntu 12.04, andlimited to 16GB RAM.We started with the 200-server/4-cores-per-server physical data centerfrom [212], but then considered larger versions of the original benchmark to studyperformance scaling. The larger data centers are also more representative of currentdata centers, using 16-core servers rather than the 4-core servers from the originalpaper.Table 6.8 summarizes our results. SecondNet, being heuristic and incomplete,is much faster than MONOSAT, but MONOSAT is much faster and more scalablethan Z3-AR. Importantly, we see that MONOSAT is able to scale to thousandsof servers, with typical per-instance allocation times of a few seconds or less perVDC. Furthermore, on these tree-structured data centers, MONOSAT is typicallyable to allocate two or even three times as many VDCs as SecondNet onto the sameinfrastructure.Comparison on BCube and FatTree from [113] The second experiment we con-ducted is a direct comparison against the original SecondNet implementation, withthe latest updates and bug fixes, obtained from its original authors [113] (this is alsothe version of SecondNet we use for all subsequent comparisons in this section).Note that the implementation of Z3-generic, and both the theory and implementa-tion of Z3-AR, are restricted to tree topologies, so they could not be included inthese experiments.The SecondNet benchmark instances are extremely large — in one case ex-102Chapter 6. Graph Theory ApplicationsVDC Instance Structurevn3.1 vn3.2 vn3.3 vn5.1 vn5.2 vn5.3#VDCTime (s)#VDC Time #VDC Time #VDC Time #VDC Time #VDC TimeFatTree Physical Data Center, 432 servers with 16 cores eachSecondNet 624 < 0.1 360 < 0.1 252 < 0.1 132 < 0.1 132 < 0.1 120 < 0.1MONOSAT 768 148.3 762 257.9 760 247.1 448 317.3 438 315.9 436 323.26BCube Physical Data Center, 512 Servers with 16 cores eachSecondNet 833 < 0.1 486 < 0.1 386 < 0.1 157 < 0.1 195 < 0.1 144 < 0.1MONOSAT 909 201.0 881 347.3 869 310.4 473 502.7 440 423.3 460 412.8FatTree Physical Data Center, 1024 servers with 16 cores eachSecondNet 1520 < 0.1 848 < 0.1 608 < 0.1 272 < 0.1 288 < 0.1 278 < 0.1MONOSAT 1820 787.6 1812 1435.2 1806 1437.5 1064 1941.5 1031 1925.2 963 (3600)BCube Data Center, 1000 Servers with 16 cores eachSecondNet 1774 < 0.1 1559 1.4 1188 2.0 302 < 0.1 356 < 0.1 293 < 0.1MONOSAT 1775 746.6 1713 1350.4 1677 1513.8 938 1741.8 884 1747.1 912 1697.0Table 6.9: Total Number of Consecutive VDCs Allocated on Data Centerswith FatTree and BCube Topologies. Table labels are the same as in Ta-ble 6.8, but in this table, ‘SecondNet’ refers to the original implementa-tion. As before, SecondNet is much faster, but MONOSAT scales well upto hundreds of servers, typically allocating VDCs in less than a second.And in most cases, MONOSAT allocated many more VDCs to the samephysical data center. (An interesting exception is the lower-left corner.Although MONOSAT is complete for any specific allocation, it is allo-cating VDCs one-at-a-time in an online manner, as do other tools, so theoverall allocation could end up being suboptimal.)ceeding 100 000 servers — but also extremely easy to allocate: the available band-width per link is typically ≥ 50× the requested communication bandwidths in theVDC, so with only 16 cores per server, the bandwidth constraints are mostly ir-relevant. For such easy allocations, the fast, incomplete approach that Second-Net uses is the better solution. Accordingly, we scaled the SecondNet instancesdown to 432–1024 servers, a realistic size for many real-world data centers. Forthese experiments, we generated sets of 10 VDCs each of several sizes (6, 9, 12and 15 VMs), following the methodology described in [212]. These VDCs haveproportionally greater bandwidth requirements than those originally considered bySecondNet, requiring 5–10% of the smallest link-level capacities. The resultingVDC instances are large enough to be representative of many real-world use cases,103Chapter 6. Graph Theory Applicationswhile also exhibiting non-trivial bandwidth constraints. For each of these sets ofVDCs, we then repeatedly allocate instances (in random order) until the data centeris saturated.Table 6.9 shows allocations made by SecondNet and MONOSAT on two datacenters, one with a BCube topology with 512 servers, and one with a FatTree topol-ogy with 432 servers. As in our previous experiment, SecondNet is much fasterthan MONOSAT, but MONOSAT is fast enough to be practical for data centerswith hundreds of servers, with typical allocation times of a few seconds per VDC(however, in a minority of cases, MONOSAT did require tens or even hundreds ofseconds for individual allocations). In many cases, MONOSAT was able to allocatemore than twice as many VDCs as SecondNet on these data centers — a substantialimprovement in data center utilization.Comparison on commercial networks The above comparisons consider howMONOSAT compares to existing VDC allocation tools on several artificial (butrepresentative) network topologies from the VDC literature. To address the ques-tion of whether there are actual real-world VDC applications where MONOSATperforms not only better than existing tools, but is also fast enough to be used inpractice, we also considered a deployment of a standard Hadoop virtual cluster,on a set of actual data center topologies. We collaborated with the private cloudprovider ZeroStack [215] to devise an optimal virtual Hadoop cluster to run Tera-sort [60]. Each Hadoop virtual network consists of a single master VM connectedto 3–11 slave VMs. We consider 5 different sizes of VMs, ranging from 1 CPU and1GB RAM, to 8 CPUs and 16GB of RAM; for our experiments, the slave VMs areselected randomly from this set, with the master VM selected randomly but alwaysat least as large as the largest slave VM. The Hadoop master has tree connectivitywith all slaves, with either a 1 or 2 Gbps network link between the master and eachslave.The physical data center topology was provided by another company, whichrequested to remain anonymous. This company uses a private cloud deployedacross four data centers in two geographic availability zones (AZs): us-west andus-middle. Each data center contains between 280 and 1200 servers, spread across1 to 4 clusters with 14 and 40 racks. Each server has 16 cores, 32 GB RAM,104Chapter 6. Graph Theory Applications20 Gbps network bandwidth (over two 10 Gbps links). The network in each datacenter has a leaf-spine topology, where all ToR switches connect to two distinctaggregation switches over two 20 Gbps links each (a total of 4 links with 80 Gbps;two on each aggregation switch) and aggregation switches are interconnected withfour 40 Gbps links each. For each cluster, there is a gateway switch with a 240Gbps link connected to each aggregation switch. All data centers use equal-costmulti-path (ECMP) to take advantage of multiple paths.A VDC is allocated inside one AZ: VMs in one VDC can be split across twoclusters in an AZ, but not across two AZs. Table 6.10 shows VDC allocation resultsper AZ.We applied SecondNet and MONOSAT in this setting, consecutively allocatingrandom Hadoop master-slave VDCs of several sizes, ranging from 4 to 12 VMs,until no further allocations could be made. Note that, as with the previous experi-ment, Z3-AR is unable to run in this setting as it is restricted to tree-topology datacenters.In Figure 6.10 we show the results for the largest of these data centers (re-sults for the smaller data centers were similar). As with the previous experimentson Tree, FatTree, and BCube topology data centers, although SecondNet is muchfaster than MONOSAT, MONOSAT’s per-instance allocation time is typically justa few seconds, which is realistically useful for many practical applications. Aswith our previous experiments, MONOSAT is able to allocate many more VDCsthan SecondNet - in these examples, allocating between 1.5 and 2 times as manytotal VDCs as SecondNet, across a range of data center and VDC sizes, includinga commercial data center with more than 1000 servers.MONOSAT is not only able to find many more allocations than SecondNet inthis realistic setting, but MONOSAT’s median allocation time, 1-30 seconds, showsthat it can be practically useful in a real, commercial setting, for data centers andVDCs of this size. This provides strong evidence that MONOSAT can find prac-tical use in realistic settings where large or bandwidth-hungry VDCs need to beallocated. It also demonstrates the practical advantage of a (fast) complete algo-rithm like MONOSAT over a much faster but incomplete algorithm like Second-Net: for bandwidth-heavy VDCs, even with arbitrary running time, SecondNet’sVDCAlloc is unable to find the majority of the feasible allocations.105Chapter 6. Graph Theory ApplicationsVDC Instance Structure1G-4 VMs 2G-4 VMs 1G-10 VMs2G-10 VMs1G-15 VMs 2G-15 VMs#VDCTime #VDCTime #VDCTime #VDCTime #VDCTime #VDC TimeUS West 1: 2 clusters, 60 racks, 1200 servers with 32 cores eachSecondNet 2400 0.6 1740 0.4 786 0.2 240 < 0.1 480 0.1 0 < 0.1MONOSAT 2399 261.7 2399 261.7 943 244.1 932 265.7 613 554.5 600 1205.3US West 2: 1 clusters, 14 racks, 280 servers with 32 cores eachSecondNet 560 0.1 406 < 0.1 184 < 0.1 56 0.1 112 < 0.1 0 < 0.1MONOSAT 560 14.5 560 14.2 221 11.3 217 11.3 146 14.9 143 44.7US Mid 1: 4 clusters, 24 racks, 384 servers with 32 cores eachSecondNet 768 0.2 553 0.1 244 < 0.1 73 0.1 191 < 0.1 0 < 0.1MONOSAT 767 28.0 765 27.3 303 21.8 301 20.5 200 30.5 191 79.4US Mid 2: 1 cluster, 40 racks, 800 servers with 32 cores eachSecondNet 1600 0.3 1160 0.3 524 0.12 160 < 0.1 320 0.1 0 < 0.1MONOSAT 1597 111.7 1597 114.9 634 82.8 621 97.9 413 263.7 402 373.3Table 6.10: Total number of consecutive Hadoop VDCs allocated on datacenters. Here we consider several variations of a virtual network de-ployment for a standard Hadoop cluster, on 4 real-world, full-scale com-mercial network topologies. 1G-4 VMs stands for Hadoop virtual clus-ter consisting of 4 VMs (one master and 3 slaves as described Section6.3.4 commercial networks subsection), where master connects with allslaves over 1 Gbps link. As before, SecondNet is much faster thanMONOSAT (though MONOSAT is also fast enough for real-world us-age, requiring typically < 1 second per allocation). However, as thevirtual network becomes even moderately large, MONOSAT is able toallocate many more virtual machines while respecting end-to-end band-width constraints - often allocating several times as many machines asSecondNet, and in extreme cases, finding hundreds of allocations incases where SecondNet cannot make any allocations at all. Similarly,keeping the virtual network the same size but doubling the bandwidthrequirements of each virtual machine greatly decreases the allocationsthat SecondNet can make, while MONOSAT is much more robust tothese more congested settings.This reinforces our observations from our earlier experiments with artificialtopologies: MONOSAT improves greatly on state-of-the-art VDC allocation, forbandwidth-constrained data centers with as many as 1000 servers.106Chapter 6. Graph Theory Applications6.4 ConclusionIn this chapter, we have presented comprehensive experimental results from threediverse fields: procedural content generation, circuit layout, and data center allo-cation. The experiments demonstrate that our graph theory, as implemented in theSMT solver MONOSAT using the techniques described in Chapter 4, can extendthe state-of-the-art. Our results in particular highlight the scalability and effec-tiveness of our support for maximum flow predicates (which are key to the circuitlayout and data center allocation encodings).At the same time, these experiments also give a picture of some of the caseswhere MONOSAT does not perform as well as other approaches. For example, wesaw that for reachability constraints, there are cases where MONOSAT is greatlymore scalable, and also cases where it is greatly less scalable, than the ASP solverCLASP.However, while there do exist cases where MONOSAT’s graph theory maynot outperform existing approaches, we have also shown that in many importantcases — and across diverse fields — MONOSAT’s graph theory does in fact attainstate-of-the-art performance that greatly improves upon existing SAT, SMT, or ASPsolvers.107Chapter 7Monotonic Theory of GeometryThe next monotonic theory that we consider is a departure from the other theorieswe have explored. This is a theory of predicates concerning the convex hulls offinite, symbolic point sets (Figure 7.1). There are many common geometric prop-erties of the convex hulls of point sets that monotonically increase (or decrease) asadditional points are added to the set. For example, the area covered by the convexhull of a point set may increase (but cannot decrease) as points are added to the set.Similarly, given a fixed query point q, adding additional points to the pointset can cause the convex hull of that point set to grow large enough to contain q,but cannot cause a convex hull that previously contained q to no longer containq. Similarly, given the convex hulls of two point sets, adding a point to either setcan cause the two hulls to overlap, but cannot cause overlapping hulls to no longeroverlap.Many other common geometric properties of point sets are also monotonic (ex-amples include the minimum distance between two point sets, the geometric span(i.e., the maximum diameter) of a point set, and the weight of the minimum steinertree of a point set), but we restrict our attention to convex hulls in this chapter. Ourcurrent implementation is limited to 2 dimensional point sets; however, it shouldbe straightforward to generalize our solver to 3 or more dimensions.We implemented a theory solver supporting several predicates of convex hullsof finite point sets, using the SMMT framework, in our SMT solver MONO-SAT (described in Appendix A). Our implementation operates over one or more108Chapter 7. Monotonic Theory of Geometryp1p2p3p0p1p2p3p0Figure 7.1: Left: Convex hull (shaded) of a symbolic point set S ⊆{p0, p1, p2, p3}, with S = {p0, p2, p3}. Right: Convex hull of S ={p0, p1, p2, p3}. Each point pi has a fixed, constant position; the solvermust chose which of these points to include in S. Many properties ofthe convex hulls of a point set S are monotonic with respect to S. Forexample, adding a point to S may increase the area of the convex hull,but cannot decrease it.p7p6p4p5p1p2p3p0 p6p4p5p7p1p2p3p0Figure 7.2: Left: Under-approximative convex hull H− (shaded) of a sym-bolic point set S⊆{p0 . . . p7}, under the partial assignmentM= {(p0 ∈S),(p1 ∈ S),(p2 ∈ S),(p3 ∈ S),(p4 /∈ S),(p6 /∈ S),(p7 ∈ S)}, with p5unassigned (shown in grey). Right: Over-approximate convex hull H+of the same assignment.109Chapter 7. Monotonic Theory of Geometrysymbolic sets of points S0,S1, . . ., with their elements represented in Booleanform as a set of element atoms (p0 ∈ S0),(p1 ∈ S0), . . .. Given a partial assign-ment M = {(p0 ∈ S),(p1 ∈ S),(p2 ∈ S),(p3 ∈ S),(p4 /∈ S),(p6 /∈ S),(p7 ∈ S)}to the element atoms, the theory solver first forms concrete over- and under-approximations of each point set, S+,S−, where S− contains only the elementsfor which (pi ∈ Si) ∈M, and S+i all elements in S−i along with any points thatare unassigned in M. For each set, the theory solver then computes an under-approximative convex hull, H−i , from S−i , and an over-approximative convex hull,H+i , from S+i (see Figure 7.2). Because S−i and S+i are both concrete sets (contain-ing tuples representing 2D points), we can use any standard convex hull algorithmto compute H−i and H+i ; we found that Andrew’s monotone chain algorithm [7]worked well in practice.Having computed H−i and H+i for each set Si, we then iterate through eachpredicate atom p(Si) in the theory and individually compute whether they hold inH−i , or fail to hold in H+i . These individual checks are computed for each predicateas described in Section 7.1.One important consideration of these geometric properties is that numericalaccuracy plays a greater role than in the graph properties we considered above. Solong as the coordinates of the points themselves are rationals, the area of the con-vex hull, along with point containment and intersection queries can be computedprecisely using arbitrary precision rational arithmetic by computing determinants(see e.g. chapters 3.1.6 and 11.5 of [91] for a discussion of using determinants forprecise computation of geometric properties).Our implementation of the geometric theory solver is very close to thegraph solver we described in Chapter 5, computing concrete under- and over-approximations H−,H+ of the convex hull of each point set in the same manneras we did for G− and G+. We also include two additional improvements. First,after computing each H−i ,H+i , we compute axis-aligned bounding boxes for eachhull (bound−i ,bound+i ). These bounding boxes are very inexpensive to compute,and allow us to cheaply eliminate many collision detections with the underlyingconvex hulls (this is especially important when using expensive arbitrary precisionarithmetic).Secondly, once an intersection (or point containment) is detected, we find a110Chapter 7. Monotonic Theory of Geometrysmall (but not necessarily minimal) set of points which are sufficient to produce thatcollision. For example, when a point is found to be contained in a convex hull, wecan find three points from that hull that together form a triangle that contains thatpoint. So long as the points composing that triangle remain in the hull, even if otherpoints are removed from the hull the point will remain contained. These pointscan typically be arranged to be computed as a side effect of collision detection,and their presence in S− of S+ can be checked very cheaply, allowing us to skipmany future collision checks while those points remain in the relevant set. Thisis particularly useful for the geometric properties we consider here, as a) a fixed,small number of points typically constitute a proof of containment (usually 2, 3 or4), in contrast to potentially very long paths in the graph theory solver of Chapter5, and b) the geometric properties we are computing may be much more expensivethan the graph properties being computed in the previous section.Many techniques exist for speeding up the computation of dynamic geometricproperties, especially with regards to collision detection, and we have only imple-mented the most basic of these (bounding boxes); a more efficient implementationcould make use of more efficient data structures (such as hierarchical boundingvolumes or trapezoidal maps [69]) to greatly speed up or obviate many of the com-putations in our solver. As before, because these computations are performed onthe concrete under- and over-approximation point sets, standard algorithms or off-the-shelf collision detection libraries may be used.7.1 Geometric Predicates of Convex HullsIn this section, we describe the predicates of convex hulls supported by our geom-etry theory (each of which is a finite, monotonic predicate). For each predicate,we describe the algorithms used to check the predicates on H− and H+ (as com-puted above), and to perform conflict analysis when a conflict is detected by theorypropagation.7.1.1 Areas of Convex HullsGiven a set of points S, the area of the convex hull of that set of points (shadedarea of Figure 7.1) can increase as more points are added to S, but cannot decrease.111Chapter 7. Monotonic Theory of GeometryThis predicate allows us to constrain the area of the convex hull of a point set. Inprinciple, this predicate could also be set up as a monotonic function (returning thearea of the hull as a rational-value), however as our implementation does not yetsupport linear arithmetic, we will treat the area constraints as predicates comparingthe area to fixed, constant values (expressed as arbitrary precision rationals).Monotonic Predicate: hullArea≥x(S), hullArea>x(S), with S a finite set S ⊆ S′,true iff the convex hull of the points in S has an area ≥ (resp. >) than x(where x is a constant). This predicate is positive monotonic with respect tothe set S.Algorithm: Initially, compute area(bound−),area(bound+). If area(bound−) >x, compute area(H−); if area(bound+) < x, compute area(H+). The areascan be computed explicitly, using arbitrary precision rational arithmetic.Conflict set for hullArea≥x(S): The area of H− is greater or equal to x. Letp0, p1, . . . be the points of S− that form the vertices of the under-approximatehull H− (e.g., points p0, p1, p2, p3 in Figure 7.2), with area(H−) ≥ x. Thenat least one of the points must be disabled for the area of the hull to decreasebelow x. The conflict set is {(p0 ∈ S),(p1 ∈ S), . . . ,¬hullArea≥x(S)}.Conflict set for ¬hullArea≥x(S): The area of H+ is less than x; then at least onepoint pi /∈ S+ that is not contained in H+ must be added to the point setto increase the area of H+ (e.g., only point p4 in Figure 7.2). Let pointsp0, p1, . . . be the points of S′ not contained in H+. The conflict set is{(p0 /∈ S+),(p1 /∈ S+), . . . ,hullArea≥x(S)}, where p0, p1, . . . are the (pos-sibly empty) set of points pi /∈ S+.7.1.2 Point Containment for Convex HullsGiven a set of points S, and a fixed point q (Figure 7.3), adding a point to S cancause the convex hull of S to grow to contain p, but cannot cause a convex hull thatpreviously contained q to no longer contain q.Note that this predicate is monotonic with respect to the set of elements in S,but that it is not monotonic with respect to translating the position of the point q112Chapter 7. Monotonic Theory of Geometryqp1p2p3p0 qp1p2p3p0Figure 7.3: Left: The convex hull (shaded) of a symbolic point set S ⊆{p0, p1, p2, p3}, along with a query point q. Right: The same as theleft, but with point p1 added to set S. After adding this point, the convexhull of S contains point q.(which is why q must be a constant). The points that can be contained in S, alongwith the query point q, each have constant (2D) positions.There are a few different reasonable definitions of point containment (all mono-tonic), depending on whether the convex hull is treated as a closed set (includingboth the area inside the hull and also the edges of the hull) or an open set (only theinterior of the hull is included). We consider the closed variant here.Monotonic Predicate: hullContainsq(S), true iff the convex hull of the 2D pointsin S contains the (fixed point) q.Algorithm: First, check whether q is contained in the under-approximative bound-ing box bound−. If it is, then check if q is contained in H−. We use thePNPOLY [95] point inclusion test to perform this check, using arbitraryprecision rational arithmetic, which takes time linear in the number of ver-tices of H−. In the same way, only if q is contained in bound+, check if q iscontained in H+ using PNPOLY.Conflict set for hullContainsq(S): Convex hull H− contains p. Let p0, p1, p2 bethree points from H− that form a triangle containing p (such a triangle mustexist, as H− contains p, and we can triangulate H−, so one of those trian-gles must contain p; this follows from Carathodory’s theorem for convexhulls [51]). So long as those three points are enabled in S, H− must contain113Chapter 7. Monotonic Theory of Geometryrp1p2p3p0rp1p2p3p0Figure 7.4: Left: The convex hull (shaded) of a symbolic point set S ⊆{p0, p1, p2, p3}, along with a line-segment r. Right: The same as theleft, but with point p1 added to set S. After adding this point, the convexhull of S intersects line-segment r.them, and as they contain p, p must be contained in H−. The conflict set is{(p0 ∈ S),(p1 ∈ S),(p2 ∈ S),¬hullContainsq(S)}.Conflict set for ¬hullContainsq(S): p is outside of H+. If H+ is empty, then theconflict set is the default monotonic conflict set. Otherwise, we will em-ploy the separating axis theorem to produce a conflict set. The separatingaxis theorem [91] states that given any two convex volumes, either thosevolumes intersect, or there exists a separating axis on which the volumesprojections do not overlap. As we gave two non-intersecting convex hulls,by the separating axis theorem, there exists a separating axis between H+and p. Let p0, p1, . . . be the (disabled) points of S whose projection onto thataxis is ≥ the projection of p onto that axis1. At least one of those pointsmust be enabled in S in order for H+ to grow to contain p. The conflict setis {(p0 /∈ S),(p1 /∈ S), . . . ,hullContainsq(S)}.114Chapter 7. Monotonic Theory of Geometry7.1.3 Line-Segment Intersection for Convex HullsGiven a set of points S, and a fixed line-segment r (Figure 7.4), adding a point toS can cause the convex hull of S to intersect r, but cannot cause a convex hull thatpreviously intersected r to no longer intersect r.This predicate directly generalizes the previous predicate (point containment),and we will utilize some of the methods described above in computing it.Monotonic Predicate: hullIntersectsr(S), true iff the convex hull of the 2D pointsin S intersects the (fixed) line-segment r.Algorithm: First, check whether line-segment r intersects bound−. If it does,check if r intersects H−. If H− is empty, is a point, or is itself a line-segment,this is trivial (and can be checked precisely in arbitrary precision arithmeticby computing cross products following [110]). Otherwise, we check if ei-ther end-point of r is contained in H−, using PNPOLY as above for pointcontainment. If neither end-point is contained, we check whether the line-segment intersects H−, by testing each edge of H− for intersection with r(as before, by computing cross-products). If r does not intersect the under-approximation, repeat the above on bound+ and H+.Conflict set for hullIntersectsr(S): Convex hull H− intersects line-segment r. Ifeither end-point or r was contained in H−, then proceed as for the point con-tainment predicate. Otherwise, the line-segment r intersects with at least oneedge (pi, p j) of H−. Conflict set is {(pi ∈ S),(p j ∈ S),¬hullIntersectsr(S)}.Conflict set for ¬hullIntersectsr(S): r is outside of H+. If H+ is empty, then theconflict set is the naı¨ve monotonic conflict set. Otherwise, by the separatingaxis theorem, there exists a separating axis between H+ and line-segment r.Let p0, p1, . . . be the (disabled) points of S whose projection onto that axis is≥ the projection of the nearest endpoint of r onto that axis. At least one of1We will make use of the separating axis theorem several times in this chapter. The standard pre-sentation of the separating axis theorem involves normalizing the separating axis (which we wouldnot be able to do using rational arithmetic). This normalization is required if one wishes to computethe minimum distance between the projected point sets; however, if we are only interested in com-paring distances (and testing collisions), we can skip the normalization step, allowing the separatingaxis to be found and applied using only precise rational arithmetic.115Chapter 7. Monotonic Theory of Geometryp1p2p3p0q0 q1q3q2p1p2p3p0q0 q1q3q2Figure 7.5: Left: Two symbolic point sets S0 ⊆ {p0, p1, p2, p3} and S1 ⊆{q0,q1,q2,q3}, and corresponding convex hulls (shaded areas). Hullsare shown for S0 = {p0, p2, p3}, S1 = {q0,q1,q2}. These convex hullsdo not intersect. Right: Same as the left, but showing hulls after addingpoint p1 into S0. With this additional point in S0, the convex hulls ofS0,S1 now intersect.those points must be enabled in S in order for H+ to grow to contain p. Theconflict set is {(p0 /∈ S),(p1 /∈ S), . . . ,hullIntersectsr(S)}.7.1.4 Intersection of Convex HullsThe above predicates can be considered a special-case of a more general predicatethat tests the intersections of the convex hulls of two point sets: Given two setsof points, S0 and S1, with two corresponding convex hulls hull0 and hull1 (seeFigure 7.5), adding a point to either point set can cause the corresponding hull togrow such that the two hulls intersect; however, if the two hulls already intersect,adding additional points to either set cannot cause previously intersecting hulls tono longer intersect.This convex hull intersection predicate can be used to simulate the previouspredicates (hullContainsq(S) and hullIntersectsr(S)), by asserting all of the pointsof one of the two point sets to constant assignments. For efficiency, we handle pointcontainment, and intersection with fixed line segments or polygons, as special casesin our implementation.As with the other above predicates, there are variations of this predicate de-pending on whether or not convex hulls with overlapping edges or vertices areconsidered to intersect (that is, whether the convex hulls are open or closed sets).116Chapter 7. Monotonic Theory of GeometryWe will consider only the case where both hulls are closed (i.e., overlapping edgesor vertices are considered to intersect).Monotonic Predicate: hullsIntersect(S0,S1), true iff the convex hull of the pointsin S0 intersects the convex hull of the points of S1.Algorithm: If the bounding box for convex hull H0 intersects the bounding boxfor H1, then there are two possible cases to check for:1. 1: A vertex of one hull is contained in the other, or2. 2: an edge of H0 intersects an edge of H1.If neither of the above cases holds, then H0 and H1 do not intersect.Each of the above cases can be tested in quadratic time (in the number ofvertices of H0 and H1), using arbitrary precision arithmetic. In our imple-mentation, we use PNPOLY, as described above, to test vertex containment,and test for pair-wise edge intersection using cross products.Conflict set for hullsIntersect(S0,S1): H−0 intersects H−1 . There are two (not mu-tually exclusive) cases to consider:1. A vertex of one hull is contained within the other hull. Let the point bepa; let the hull containing it be H−b . Then (as argued above) there mustexist three vertices pb1, pb2, pb3 of H−b that form a triangle containingpa. So long as those three points and pa are enabled, the two hullswill overlap. Conflict set is {(pa ∈ Sa),(pb1 ∈ Sb),(pb2 ∈ Sb),(pb3 ∈Sb),¬hullsIntersect(S0,S1)}.2. An edge of H−0 intersects an edge of H−1 . Let p1a, p1b be points of H−0 ,and p2a, p2b points of H−1 , such that line segments p1a, p1b and p2a, p2bintersect. So long as these points are enabled, the hulls of the twopoint sets must overlap. Conflict set is {(p0a ∈ S0),(p0b ∈ S0),(p1a ∈S1),(p1b ∈ S1),¬hullsIntersect(S0,S1)}.Conflict set for ¬hullsIntersect(S0,S1): H+0 do not intersect H+1 . In this case,there must exist a separating axis (Figure 7.6) between H+0 and H+1 . (Such117Chapter 7. Monotonic Theory of Geometryp1p2p3p0q0 q1q3q2Figure 7.6: A separating axis between two convex hulls. Between any twodisjoint convex polygons, there must exist a separating line (dotted line)parallel to one edge (in this case, parallel to edge (p0, p2)), and a sep-arating axis (solid line) normal to that edge. Small circles show thepositions of each point, projected onto the separating axis.an axis can be discovered as a side effect of computing the cross productsof each edge in step 2 above.) Project all disabled points of S0 and S1 ontothat axis (white dots in Figure 7.6). Assume (without loss of generality)that the maximum projected point of H+1 is less than the minimum projectedpoint of H+1 . Let p0a, p0b, . . . be the disabled points of S0 whose projectionsare on the far side of the maximum projected point of H+1 . Let p1a, p1b, . . .be the disabled points of S1 whose projections are near side of the minimumprojected point of H+1 . At least one of these disabled points must be enabled,or this axis will continue to separate the two hulls. The conflict set is {(p0a /∈S0),(p0b /∈ S0), . . . ,(p1a /∈ S1),(p1b /∈ S1), . . . ,hullsIntersect(S0,S1)}.Many other common geometric properties of point sets are also monotonic (ex-amples include the minimum distance between two point sets, the geometric span(i.e., the maximum diameter) of a point set, and the weight of the minimum steinertree of a point set), but we restrict our attention to convex hulls in this chapter. Ourcurrent implementation is limited to 2-dimensional point sets; however, it shouldbe straightforward to generalize our solver to 3 or more dimensions.7.2 ApplicationsHere we consider the problem of synthesizing art galleries, subject to constraints.This is a geometric synthesis problem, and we will use it to demonstrate the convex118Chapter 7. Monotonic Theory of Geometryhull constraints introduced in this chapter. The Art Gallery problem is a classic NP-Hard problem [55, 124, 141], in which (in the decision version) one must determinewhether it is possible to surveil the entire floor-plan of a multi-room building witha limited set of fixed cameras. There are many variations of this problem, and wewill consider a common version in which cameras are restricted to being placed onvertices, and only every vertex of the room must be seen by a camera (as opposed tothe floor or wall spaces between them). This variant is NP-Complete, by reductionfrom the dominating set problem [141].Here, we introduce a related problem: Art Gallery Synthesis. Art Gallery Syn-thesis is the problem of designing an ‘art gallery’ (a non-intersecting set of simplepolygons without holes) that can be completely surveilled by at most a given num-ber of cameras, subject to some other constraints on the allowable design of theart gallery. The additional constraints may be aesthetic or logistic constraints (forexample, that the area of the floor plan be within a certain range, or that there bea certain amount of space on the walls). There are many circumstances where onemight want to design a building or room that can be guarded from a small numberof sightlines. Obvious applications include aspects of video game design, but alsoprisons and actual art galleries.Without constraints on the allowable geometry, the solution to the art gallerysynthesis problem is trivial, as any convex polygon can be completely guardedby a single camera. For this reason, the choice of constraints has a great impacton the practical difficulty of solving this problem. There are many ways that onecould constrain the art gallery synthesis problem; below, we will describe a setof constraints that are intended to exhibit the strengths of our convex hull theorysolver, while also producing visually interesting and complex solutions.We consider the following constrained art gallery synthesis problem: Given abox containing a fixed set of randomly placed 2-dimensional points, we must findN non-overlapping convex polygons with vertices selected from those points, suchthat a) the area of each polygon is greater than some fixed constant (to prevent linesor very small slivers from being created), b) the polygons may meet at an edge, butmay not meet at just one vertex (to prevent forming wall segments of infinitesimalthickness), and c) all vertices of all the polygons, and all 4 corners of the room, canbe seen by a set of at most M cameras (placed at those vertices). Figure 7.7 shows119Chapter 7. Monotonic Theory of GeometryFigure 7.7: Two artificial ‘art galleries’ found by MONOSAT. Cameras arelarge black circles; potential vertices are gray dots. Each convex poly-gon is the convex hull of a subset of those gray dots, selected by thesolver. Notice how the cameras have been placed such that the verticesof all hulls (and the corners of the room) can been seen, some alongvery tight angles, including one vertex completely embedded in threeadjoining polygons. Notice also that some vertices can only be seenfrom one ‘side’, and that some edges cannot be fully observed by cam-eras (as only vertices are required to be observable in this variant of theproblem).Art Gallery Synthesis MONOSAT Z310 points, 3 polygons, ≤3 cameras 2s 7s20 points, 4 polygons, ≤4 cameras 36s 433s30 points, 5 polygons, ≤5 cameras 187s > 3600s40 points, 6 polygons, ≤6 cameras 645s > 3600s50 points, 7 polygons, ≤7 cameras 3531s > 3600s20 points, 10 polygons, ≤5 cameras 3142s Timeout20 points, 10 polygons, ≤3 cameras 16242s MemoutTable 7.1: Art gallery synthesis results.two example solutions to these constraints, found by MONOSAT.In Table 7.1, we compare MONOSAT to an encoding of the above constraintsinto the theory of linear arithmetic (as solved by Z3, version 4.3.1). Unfortunately,the encoding into linear arithmetic is very expensive, using a cubic number ofcomparisons to find the points making up the hull of each convex polygon. We120Chapter 7. Monotonic Theory of Geometrydescribe the encoding in detail in Appendix C.3.For each instance, we list the number of randomly placed points from whichthe vertices of the polygons can be chosen, the number of convex polygons to beplaced, and the maximum number of cameras to be used to observe all the verticesof the room (with cameras constrained to be placed only on vertices of the placedpolygons, or of the bounding box). These experiments were conducted on an Inteli7-2600K CPU, at 3.4 GHz (8MB L3 cache), limited to 10 hours of runtime and 16GB of RAM.The version of art gallery synthesis that we have considered here is an artificialone that is likely far removed from a real-world application. Nonetheless, theseexamples show that, as compared to a straightforward linear arithmetic encoding,the geometric predicates supported by MONOSAT can be greatly more scalable,and show promise for more realistic applications.121Chapter 8Monotonic Theory of CTLComputation Tree Logic (CTL) synthesis [56] is a long-standing problem with ap-plications to synthesizing synchronization protocols and concurrent programs. Weshow how to formulate CTL model checking as a monotonic theory, enabling usto use the SMMT framework of Chapter 4 to build a Satisfiability Modulo CTLsolver (implemented as part of our solver MONOSAT, described in Appendix A).This yields a powerful procedure for CTL synthesis, which is not only faster thanprevious techniques from the literature, but also scales to larger and more diffi-cult formulas. Moreover, our approach is efficient at producing minimal Kripkestructures on common CTL synthesis benchmarks.Computation tree logic is widely used in the context of model checking, wherea CTL formula specifying a temporal property, such as safety or liveness, ischecked for validity in a program or algorithm (represented by a Kripke structure).Both the branching time logic CTL and its application to model checking werefirst proposed by Clarke and Emerson [56]. In that work, they also introduceda decision procedure for CTL satisfiability, which they applied to the synthesisof synchronization skeletons, abstractions of concurrent programs which are no-toriously difficult to construct manually. Though CTL model checking has foundwide success, there have been fewer advances in the field of CTL synthesis due toits high complexity.In CTL synthesis, a system is specified by a CTL formula, and the goal is tofind a model of the formula — a Kripke structure in the form of a transition system122Chapter 8. Monotonic Theory of CTLAG(EF(q∧¬p)) p,¬qp,q ¬p,q¬p,¬qFigure 8.1: An example CTL formula (left), and a Kripke structure (right)that is a model for that formula. The CTL formula can be paraphrasedas asserting that “On all paths starting from the initial state (A) in allstates on those paths (G) there must exist a path starting from that state(E) which contains at least one state (F) in which q holds and p doesnot.” The Kripke structure consists of a finite set of states connectedwith unlabelled, directed edges, and a finite set of state properties. Onestate is distinguished as the initial state in which the CTL formula willbe evaluated, and each state is labelled with a truth value for each atomicproposition {p,q} in the formula.in which states are annotated with sets of atomic propositions (we will refer tothese propositions as state properties). The most common motivation for CTLsynthesis remains the synthesis of synchronization for concurrent programs, suchas mutual exclusion protocols. In this setting, the Kripke structure is interpreted asa global state machine in which each global state contains every process’s internallocal state. The CTL specification in this setting consists of both structural intra-process constraints on local structures, and inter-process behavioral constraints onthe global structure (for instance, starvation freedom). If a Kripke structure isfound which satisfies the CTL specification, then one can derive from it the guardedcommands that make up the corresponding synchronization skeleton [13, 56].We introduce a theory of CTL model checking, supporting a predicateModel(φ), which evaluates to TRUE iff an associated Kripke structure K = (T,P),with transition system T and state property mapping P, is a model for the CTLformula φ . We will then show that this predicate can be used to perform boundedCTL synthesis, by specifying a space of possible Kripke structures of bounded sizeand solving the resulting formula.Due to the CTL small model property [87], in principle a bounded CTL-SATprocedure yields a complete decision procedure for unbounded CTL-SAT, but123Chapter 8. Monotonic Theory of CTLin practice, neither bounded approaches, nor classical tableau approaches, havebeen scalable enough for completeness to be a practical concern. Rather, ourapproach (like similar constraint-solver based techniques for CTL [68, 121] andLTL [132, 178]) is appropriate for the case where a formula is expected to be satis-fiable by a Kripke structure with a modest number of states (∼ 100). Nevertheless,we will show that our approach solves larger and more complex satisfiable CTLformulas, including ones with a larger numbers of states, much faster than existingbounded and unbounded synthesis techniques. This makes our approach particu-larly appropriate for CTL synthesis.In addition to being more efficient than existing techniques, our approach isalso capable of synthesizing minimal models. As we will discuss below, previousCTL synthesis approaches were either incapable of finding minimal models [12,56], or could not do so with comparable scalability to our technique [68, 121].We begin with a review of related work in Section 8.1. In Sections 8.2 and 8.3,we show how to apply the SMMT framework to a theory of CTL model checking.Sections 8.4 and 8.5 explain the most important implementation details and opti-mizations. In Section 8.6, we provide experimental comparisons to state-of-the-arttechniques showing that our SMMT-approach to CTL synthesis can find solutionsto larger and more complex CTL formulas than comparable techniques in two fam-ilies of CTL synthesis benchmarks: one derived from mutual exclusion protocols,and the other derived from readers-writers protocols. Further, our approach doesso without the limitations and extra expert knowledge that previous approachesrequire.8.1 BackgroundThe original 1981 Clarke and Emerson paper introducing CTL synthesis [56] pro-posed a tableau-based synthesis algorithm, and used this algorithm to construct a2-process mutex in which each process was guaranteed mutually exclusive accessto the critical section, with starvation freedom.Subsequently, although there has been steady progress on the general CTLsynthesis problem, the most dramatic gains have been with techniques that arestructurally-constrained, taking a CTL formula along with some additional ‘struc-124Chapter 8. Monotonic Theory of CTLtural’ information about the desired Kripke structure, not specified in CTL, whichis then leveraged to achieve greater scalability than generic CTL synthesis tech-niques. For example, in 1998, Attie and Emerson [11, 12] introduced a CTL syn-thesis technique for the case where the Kripke structure is known to be composedof multiple similar communicating processes. They used this technique to synthe-size a Kripke structure for a specially constructed 2-process version of the CTLformula (a ‘pair-program’) in such a way that the produced Kripke structure couldbe safely generalized into an N-process solution. This allowed them to producea synchronization skeleton for a mutex with 1000 or more processes, far largerthan other techniques. However, while this process scales very well, only certainCTL properties can be guaranteed to be preserved in the resulting Kripke structure,and in general the Kripke structure produced this way may be much larger thanthe minimal solution to the instance. In particular, EX and AX properties are notpreserved in this process [11].Additionally, the similar-process synthesis techniques of Attie and Emersonrely on a generic CTL synthesis method to synthesize these pair-programs. Assuch, improvements to the scalability or expressiveness of generic CTL synthesismethods can be directly applied to improving this pair-program synthesis tech-nique. Their use of the synthesis method from [56] yields an initially large Kripkestructure that they minimize in an intermediate step. We note that our approachis particularly suited for synthesizing such pair-programs, not merely for perfor-mance reasons, but also because it is able to synthesize minimal models directly.On the topic of finding minimal models, Bustan and Grumberg [49] introduceda technique for minimizing Kripke structures. However, the minimal models thatour technique produces can in general be smaller than what can be achieved bystarting with a large Kripke structure and subsequently minimizing it. This is be-cause minimization techniques which are applied on an existing Kripke structureafter its synthesis only yield a structure minimal with respect to equivalent struc-tures (for some definition of equivalence, e.g. strong or weak bisimulation). Thisdoes not necessarily result in a structure that is the overall minimal model of theoriginal CTL formula. For this reason, techniques supporting the direct synthesisof minimal models, such as ours, have an advantage over post-synthesis minimiza-tion techniques.125Chapter 8. Monotonic Theory of CTLIn 2005, Heymans et al. [121] introduced a novel, constraint-based approachto the general CTL synthesis problem. They created an extension of answer setprogramming (ASP) that they called ‘preferential ASP’ and used it to generate a2-process mutex with the added property of being ‘maximally parallel’, meaningthat each state has a (locally) maximal number of outgoing transitions (withoutviolating the CTL specification). They argued that this formalized a property thatwas implicit in the heuristics of the original 1981 CTL synthesis algorithm, andthat it could result in Kripke structures that were easier to implement as efficientconcurrent programs. As the formulation in their paper does not require additionalstructural constraints (though it can support them), it is a general CTL synthesismethod. Furthermore, being a constraint-based method, one can flexibly add struc-tural or other constraints to guide the synthesis. However, the scalability of theirmethod was poor.Subsequently, high performance ASP solvers [104] built on techniques fromBoolean satisfiability solvers were introduced, allowing ASP solvers to solve muchlarger and much more difficult ASP formulas. In 2012, De Angelis, Pettorossi, andProietti [68] showed that (unextended) ASP solvers could also be used to performefficient bounded CTL synthesis, allowing them to use the high performance ASPsolver Clasp [104]. Similar to [12], they introduced a formulation for doing CTLsynthesis via ASP in the case where the desired Kripke structure is composed ofmultiple similar processes. Using this approach, they synthesized 2-process and 3-process mutexes with properties at least as strong as the original CTL specificationfrom [12]. The work we introduce in this chapter is also a constraint-solver based,bounded CTL-synthesis technique. However, in Section 8.6, we will show thatour approach scales to larger and more complex specifications than previous work,while simultaneously avoiding the limitations that prevent those approaches fromfinding minimal models.Our focus in this section has been on this history of the development of CTLsynthesis techniques; for a gentle introduction to CTL semantics (and CTL modelchecking), we refer readers to [128].126Chapter 8. Monotonic Theory of CTL8.2 CTL Operators as Monotonic FunctionsA Kripke structure is composed of a finite state transition system (T ) over states S,and a set of Boolean state properties (P), such that each state property p ∈ P has atruth value p(s) in every state s ∈ S. One state of the Kripke structure is uniquelyidentified as the starting state — below, we will assume that the starting state isalways state s0 unless otherwise stated. A Kripke structure K = (T,P) is said to bea model of a CTL formula φ if φ evaluates to TRUE in the starting state of K.The grammar of a CTL formula φ in existential normal form (ENF) is typicallydefined recursively asφ ::= TRUE|a|¬φ |φ ∧φ |EXφ |EGφ |E(φUφ)with a being a state property.1However, we will consider a slightly different presentation of CTL that is moreconvenient to represent as a theory of CTL model checking. We represent Kripkestructures as a tuple (T,P), with T a transition system and P a set of state propertyvectors. As in the graph theory in Chapter 5, we represent the transition system Tas a set of potential transitions T ⊆ S× S, and introduce, for each transition t inT , a theory atom (t ∈ T ), such that the transition t is enabled in T if and only if(t ∈ T ) is assigned TRUE. We will also introduce, for each property vector p ∈ Pand each state s ∈ S, a theory atom (p(s)), TRUE iff property p holds in states. Together, these transition atoms and property atoms define a bounded Kripkestructure, K = (T,P). P is the set of all state properties in the Kripke structure;an individual state property p ∈ P can also be represented as a vector of BooleansA = [p(s1), p(s2), . . . p(sk)] for the k states of the system. This formulation of CTLmodel checking as a predicate over a symbolic Kripke structure with a fixed set ofstates and a variable set of transitions and property assignments is similar to theones introduced in [68, 121].1CTL formulas are also often expressed over a larger set of operators:AG,EG,AF,EF,AX ,EX ,EU,AU , along with the standard propositional connectives. How-ever, it is sufficient to consider the smaller set of existentially quantified CTL operators EX, EG andEU, along with the propositional operators (¬,∧), and TRUE, which are known to form an adequateset. Any CTL formula can be efficiently converted into a logically equivalent existential normalform (ENF) in terms of these operators, linear in the size of the original formula [147].127Chapter 8. Monotonic Theory of CTLWe introduce for each unary CTL operator op an evaluation function op(T,A),taking as arguments a set of edges specifying the transition system, T , and a 1-dimensional vector of Booleans, A, of size |S| which indicates the truth value ofproperty a in each state. Element s of A, written A[s], is interpreted as representingthe truth value of state property a in state s (we assume that the states are uniquelyaddressable in such a way that they can be used as indices into this property vector).For each of the binary operators EU and ∧, we introduce an evaluation functionop(T,A,B), which takes a single transition system as above, and two state propertyvectors, A and B, representing the truth value of the arguments of op in each state.Each evaluation function returns a vector of Booleans C, of size |S|, repre-senting a fresh state property storing, for each state s, the truth value of evaluat-ing CTL operator op from initial state s in the Kripke structure K = (T,{A}) (orK = (T,{A,B}) if op takes two arguments). This is a standard interpretation ofCTL (and how explicit-state CTL model checking is often implemented), and werefer to the literature for common ways to compute each operator (see, e.g., [128]).We support the following CTL functions (with T a transition system, andA,B,C vectors of Booleans representing state properties):• ¬(T,A) 7→C• ∧(T,A,B) 7→C• EX(T,A) 7→C• EG(T,A) 7→C• EU(T,A,B) 7→CNotice that there is no operator for specifying state properties; rather, a stateproperty is specified in the formula directly by passing a vector of state propertyatoms as an argument of an operator. For example, consider the following CTLformula in ENF form:EG(¬a∧EX(b))128Chapter 8. Monotonic Theory of CTLThis formula has two state properties (‘a’ and ‘b’). Given a transition system T ,this formula can be encoded into our theory of CTL model checking asEG(T,∧(T,¬(T,a),EX(T,b)))where a,b are vectors of Booleans representing the truth values of the state prop-erties a and b at each state in the Kripke structure.2Before describing our CTL theory solver, we will briefly establish that eachCTL operator as defined above is monotonic. First, we show that the operatorsEX(T,A), EG(T,A) or EU(T,A,B) are each monotonic with respect to the stateproperty vectors A and B. Let solves(φ(T,A)) be a predicate that denotes whetheror not the formula φ holds in state s of the Kripke structure determined by thevector of Booleans T (transitions) and A (state properties).Lemma 8.2.1. solves(EX(T,A)), solves(EG(T,A)), and solves(EU(T,A,B)) areeach positive monotonic in both the transition system T and state property A (andstate property B, for EU).Proof. Take any T , A,B that determine a structure K for which the predicate holds.Let K′ be a structure determined by some T ′, A′,B′ such that K′ has the same states,state properties and transitions as K, except for one transition that is enabled in K′but not in K, or one state property which holds in K′ but not in K. Formally, thereis exactly one argument in either T ′, A′, or B′ that is 0 in T (or A or B respectively)and 1 in T ′ (or A′ or B′ respectively). Then either (a) one of the states satisfies oneof the atomic propositions in K′, but not in K, or (b) there is a transition in K′, butnot in K.We assume solves(φ(T,A)) holds (for EU, we assume that solves(φ(T,A,B))holds). Then, there must exist a witnessing infinite sequence starting from s inK. If (b), the exact same sequence must exist in K′, since it has a superset ofthe transitions in K. Thus we can conclude solves(φ(T ′,A′)) holds (respectively,2Notice that the transition system T has to be passed as an argument to each function. We canask whether a formula EG(T1,EX(T2,a)), with T1 6= T2, is well-formed? In fact T1 does not need toequal T2 in order to have meaningful semantics, however, if the transition systems are not the same,then the semantics of the formula will not match the expected semantics of CTL. We will assumethat all formulas of interest have the same transition system passed as each argument of the formula.129Chapter 8. Monotonic Theory of CTLsolves(φ(T ′,A′,B′)) holds). If (a), then the sequence will only differ in at mostone state, where p holds instead of ¬p (or q instead of ¬q). We note that for eachof the three CTL operators, this sequence will be a witness for K′, if the originalsequence was a witness for K. Thus, solves(φ(T ′,A′)) holds as well (respectively,solves(φ(T ′,A′,B′)) holds).It is easy to see that ∧ and ∨ are positive monotonic in the same way, and ¬is negative monotonic. Excluding negation, then, all the CTL operators needed toexpress formulas in ENF have positive monotonic solves predicates, while negationalone has a negative monotonic solves predicate.Each CTL operator op(T,A) returns a vector of Booleans C, where Ci ←solvei(T,A) (or Ci ← solvei(T,A,B) for the binary operators). It directly followsthat for each op(T,A) 7→C, each entry of Ci is monotonic with respect to argumentsT,A (and also with respect to B, for binary operators).Finally, we also introduce a predicate Model(A), which takes a Boolean stateproperty vector A and returns TRUE iff the property A holds in the initial state. Forsimplicity, we will further assume (without loss of generality) that the initial stateis always state s0; predicate Model(A) then simply returns the Boolean at index s0of the vector A: Model(A) 7→ A[s0], and as a result is trivially positive monotonicwith respect to the property state vector A.8.3 Theory Propagation for CTL Model CheckingAs described in Section 8.2 each CTL operator is monotonic, as is the predicateModel(A). In Section 4.3, we described an approach to supporting theory propa-gation for compositions of positive and negative monotonic functions, relying on afunction, approx(φ ,M+A ,M−A ). In Algorithm 15 we describe our implementationof Algorithm 8 for the theory of CTL model checking, and in Algorithm 16, we de-scribe our implementation of approx for the theory of CTL model checking (withthe over- and under-approximative models represented as two Kripke structures,K+, K−).3 This approximation function in turn calls a function evaluate(op,T,A),which takes a CTL operator, transition system, and a state property vector and3Similar recursive algorithms for evaluating CTL formulas on partial Kripke structures, can befound in [43, 115, 142], applied to CTL model checking, rather than CTL synthesis.130Chapter 8. Monotonic Theory of CTLreturns a vector of Booleans containing, for each state s in T , the evaluation ofop starting from state s. Our implementation of evaluate is simply the standardimplementation of CTL model checking semantics (e.g., as described in [128]).While our theory propagation implementation for CTL model checking closelyfollows that of Section 4.3, our implementation of conflict analysis differs from(and improves on) the general case described in Section 4.3. We describe ourimplementation of conflict analysis in the next section.131Chapter 8. Monotonic Theory of CTLAlgorithm 15 Theory propagation for CTL model checking. T-Propagate takesa (partial) assignment,M. It returns a tuple (FALSE, conflict) if M is found tobe unsatisfiable, and returns a tuple (TRUE, M) otherwise. This implementationcreates an over-approximative Kripke structure K+ and an under-approximativeKripke structure K−. These Kripke structures are then used to safely over- andunder-approximate the evaluation of the CTL formula, in procedure APPROXCTL,and also during conflict analysis (procedure ANALYZECTL). The function NNFreturns a negation-normal-form formula, discussed in Section 8.4.function THEORYPROPAGATE(M)T−←{},T+←{}for each transition atom ti doif (ti /∈ T ) /∈M thenT+← T+∪{ti}if (ti ∈ T ) ∈M thenT−← T−∪{ti}P−←{},P+←{}for each state property atom p(s) doif ¬(p(s)) /∈M thenP+← P+∪{p(s)}if (p(s) ∈M thenP−← P−∪{p(s)}K+← (T+,P+)K−← (T−,P−)for each predicate atom Model(φ) doif ¬Model(φ) ∈M thenif APPROXCTL(Model(φ),K−,K+) 7→ TRUE thenreturn FALSE,ANALYZECTL(NNF(φ),s0))else if Model(φ) ∈M thenif APPROXCTL(Model(φ),K+,K−) 7→ FALSE thenreturn FALSE,ANALYZECTL(NNF(φ),s0))elseif APPROXCTL(Model(φ),K−,K+) 7→ TRUE thenM←M∪{Model(φ)}else if APPROXCTL(Model(φ),K+,K−) 7→ FALSE thenM←M∪{¬Model(φ)}return TRUE,M132Chapter 8. Monotonic Theory of CTLAlgorithm 16 APPROXCTL(φ ,K+,K−) takes a formula φ and two Kripke struc-tures, K+ and K−. If φ is the predicate Model, it returns a truth value; otherwiseit returns a vector of Booleans, representing the value of φ starting from each stateof the Kripke structure.Let (T+,P+) = K+Let (T−,P−) = K−if φ is atomic state vector p thenLookup the state property vector p in the vector of vectors P+return P+[p]else if φ is predicate Model(ψ) thenA← APPROXCTL(ψ,K+,K−)Lookup the value of starting state s0 in the vector of Booleans Areturn A[s0]else if φ is a unary operator op with argument ψ thenif op is ¬ then(op is negative monotonic)A := APPROXCTL(ψ , K−, K+)return evaluate(¬,T−,A)else if op ∈ {EX, EG} thenA := APPROXCTL(ψ , K+, K−)return evaluate(op,T+,A)else φ is binary op ∈ {EU,∧} with arguments ψ1, ψ2A1 := APPROXCTL(ψ1, K+, K−)A2 := APPROXCTL(ψ2, K+, K−)return evaluate(op,A1,A2,T+)133Chapter 8. Monotonic Theory of CTL8.4 Conflict Analysis for the Theory of CTLAbove we described our support for theory propagation for the theory of CTLmodel checking, relying on the function composition support described in Section4.3. Here, we describe our implementation of conflict analysis for the theory ofCTL model checking.Unlike our theory propagation implementation, which operates on formulas inexistential normal form, to perform conflict analysis we convert the CTL formulainto negation normal form (NNF), pushing any negation operators down to theinnermost terms of the formula. To obtain an adequate set, the formula may nowalso include universally quantified CTL operators and Weak Until. In Algorithm15 above, the function NNF(φ) takes a formula in existential normal form andreturns an equivalent formula in negation normal form. This translation to negationnormal form takes time linear in the size of the input formula; the translation canbe performed once, during preprocessing, and cached for future calls.Our procedure ANALYZECTL(φ ,s,K+,K−,M) operates recursively on theNNF formula, handling each CTL operator with a separate case, and returns a con-flict set of theory literals. In Algorithm 17 below, we show only the cases handlingoperators EX ,AX ,EF,AF ; the full algorithm including support for the remainingcases can be found in Appendix D.4.134Chapter 8. Monotonic Theory of CTLAlgorithm 17 ANALYZECTL(φ ,s,K+,K−,M) analyzes φ recursively, and re-turns a conflict set. φ is a CTL formula in negation normal form, s is a state, K+and K− are over-approximate and under-approximate Kripke structures, and Mis a conflicting assignment. Notice that for the existentially quantified operators,the over-approximation transition system K+ is analyzed, while for the universallyquantified operators, the under-approximation transition system K− is analyzed.function ANALYZECTL(φ ,s,K+,K−,M)Let (T+,P+) = K+Let (T−,P−) = K−c←{}if φ is EX(ψ) thenfor each transition t outgoing from s doif (t /∈ T ) ∈M thenc← c∪{(t ∈ T )}.for each transition t = (s,u) in T+ doif evaluate(ψ,u,K+) 7→ FALSE thenc← c∪ANALYZECTL(ψ,n,K+,K−,M).else if φ is AX(ψ) thenLet t = (s,u) be a transition in T−, with evaluate(ψ,u,K+) 7→ FALSE.(At least one such state must exist)c← c∪{(t /∈ T )}c← c∪ANALYZECTL(ψ,u,K+,K−,M).else if φ is EF(ψ) thenLet R be the set of all states reachable from s in T+.for each state r ∈ R dofor each transition t outgoing from r doif (t /∈ T ) ∈M thenc← c∪{(t ∈ T )}c← c∪ANALYZECTL(ψ,r,K+,K−,M).else if φ is AF(ψ) thenLet L be a set of states reachable from s in T−, such that L forms a lassofrom s, and such that for ∀u ∈ L,evaluate(ψ,r,K+) 7→ FALSE.for each transition t ∈ lasso doc← c∪{(t /∈ T )}for each state u ∈ L doc← c∪ANALYZECTL(ψ,u,K+,K−,M)else if φ has operator op ∈ {EG,AG,EW,AW,EU,AU,∨,∧,¬p, p} thenSee full description of ANALYZECTL in Appendix D.4.return c135Chapter 8. Monotonic Theory of CTL8.5 Implementation and OptimizationsIn Section 8.2, we showed how CTL model checking can be posed as a monotonictheory. In Sections 8.3 and 8.4, we then described our implementation of the theorypropagation and conflict analysis procedures of a lazy SMT theory solver, follow-ing the techniques for supporting compositions of positive and negative monotonicfunctions and predicates described in Section 4.3. We have also implemented someadditional optimizations which greatly improve the performance of our CTL the-ory solver. One basic optimization that we implement is pure literal filtering (see,e.g. [179]): For the case where Model(φ ) is assigned TRUE (resp. FALSE), weneed to check only whether Model(φ) is falsified (resp., made true) during theorypropagation. In all of the instances we will examine in this chapter, Model(φ)is asserted TRUE in the input formula, and so this optimization greatly simplifiestheory propagation. We describe several further improvements below.In Section 8.5.1 we describe symmetry breaking constraints, which can greatlyreduce the search space of the solver, and in Section 8.5.2 we show how severalcommon types of CTL constraints can be cheaply converted into CNF, reducingthe size of the formula the theory solver must handle. Finally, in Section 8.5.3, wediscuss how in the common case of a CTL formula describing multiple commu-nicating processes, we can (optionally) add support for additional structural con-straints, similarly to the approach described in [68]. These structural constraintsallow our solver even greater scalability, at the cost of adding more states into thesmallest solution that can be synthesized.48.5.1 Symmetry BreakingDue to the way we expose atomic propositions and transitions to the SAT solverwith theory atoms, the SAT solver may end up exploring large numbers of iso-morphic Kripke structures. We address this by enforcing extra symmetry-breakingconstraints which prevent the solver from considering (some) redundant configu-rations of the Kripke structure. Symmetry reduction is especially helpful to proveinstances UNSAT, which aids the search for suitable bounds.4Thus, if structural constraints are used, iteratively decreasing the bound may no longer yield aminimal structure.136Chapter 8. Monotonic Theory of CTLLet label(si) be the binary representation of the atomic propositions of state si,and let out(si) be the set of outgoing edges of state si. Let s0 be the initial state.The following constraint enforces an order on the allowable assignments of stateproperties and transitions in the Kripke structure:∀i, j :[i < j∧ i 6= 0∧ j 6= 0]→[(label(si)≤ label(s j)) ∧((label(si) = label(s j))→ (|out(si)| ≤ |out(s j)|))]8.5.2 PreprocessingGiven a CTL specification φ , we identify certain common sub-expressions whichcan be cheaply converted directly into CNF, which is efficiently handled by theSAT solver at the core of MONOSAT. We do so if φ matches∧i φi, as is commonlythe case when multiple properties are part of the specification. If φi is purely propo-sitional, or of the form AGp, with p purely propositional, we eliminate φi from theformula and convert φi into a logically equivalent CNF expression over the stateproperty assignment atoms of the theory.5 This requires a linear number of clausesin the number of states in K. We also convert formulas of the form AGψ , with ψcontaining only propositional logic and at most a single Next-operator (EX or AX).Both of these are frequent sub-expressions in the CTL formulas that we have seen.8.5.3 Wildcard Encoding for Concurrent ProgramsAs will be further explained later, the synthesis problem for synchronization skele-tons assumes a given number of processes, which each have a local transition sys-tem. The state transitions in the full Kripke structure then represent the possibleinterleavings of executing the local transition system of each process. This localtransition system is normally encoded into the CTL specification.Both [12] and [68] explored strategies to take advantage of the case where thelocal transition systems of these processes are made explicit. The authors of [68]found they could greatly improve the scalability of their answer-set-programming5Since AGp only specifies reachable states, the clause is for each state s, a disjunction of p beingsatisfied in s, or s having no enabled incoming transitions. This changes the semantics of CTL forunreachable states, but not for reachable states.137Chapter 8. Monotonic Theory of CTLbased CTL synthesis procedure by deriving additional ‘structural’ constraints forsuch concurrent processes. As our approach is also constraint-based, we can (op-tionally) support similar structural constraints. In experiments below, we show thateven though our approach already scales better than existing approaches withoutthese additional structural constraints, we also benefit from such constraints.Firstly, we can exclude any global states with state properties that are an illegalencoding of multiple processes. If the local state of each process is identified by aunique atomic proposition, then we can enforce that each global state must maketrue exactly one of the atomic propositions for each process. For every remainingcombination of state property assignments, excluding those determined to be illegalabove, we add a single state into the Kripke structure, with a pre-determined as-signment of atomic propositions, such that only the transitions between these statesare free for the SAT solver to assign. This is in contrast to performing synthesiswithout structural constraints, in which case all states are completely undetermined(but typically fewer are required).Secondly, since we are interested in interleavings of concurrent programs, oneach transition in the global Kripke structure, we enforce that only a single processmay change its local state, and it may change its local state only in a way that isconsistent with its local transition system.The above two constraints greatly reduce the space of transitions in the globalKripke structure that are left free for the SAT solver to assign (and completelyeliminate the space of atomic propositions to assign in each state). However theseconstraints make our procedure incomplete, since in general more than a singlestate with the same atomic propositions (but different behavior) need to be distin-guished. To allow multiple states with equivalent atomic propositions, we also adda small number of ‘wildcard’ states into the Kripke structure, whose state proper-ties and transitions (incoming and outgoing) are not set in advance. In the exampleswe consider in this chapter, we have found that a small number of such wildcardstates (between 3 and 20) are sufficient to allow for a Kripke structure that satisfiesthe CTL formula, while still greatly restricting the total space of Kripke structuresthat must be explored by the SAT solver.We disable symmetry breaking when using the wildcard encoding, as the wild-card encoding is incompatible with the constraint in Section 8.5.1.138Chapter 8. Monotonic Theory of CTL8.6 Experimental ResultsThere are few CTL synthesis implementations available for comparison. Indeed,the original CTL synthesis/model-checking paper [56] presents an implementationof CTL model checking, but the synthesis examples were simulated by hand. Theonly publicly available, unbounded CTL synthesis tool we could find is Prezza’sopen-source CTLSAT tool6, which is a modern implementation of the classictableau-based CTL synthesis algorithm [56].We also compare to De Angelis et al.’s encoding of bounded CTL synthesisinto ASP [68]. De Angelis et al. provide encodings7 specific to the n-process mu-tual exclusion example, which exploit structural assumptions about the synthesizedmodel (for example, that it is the composition of n identical processes). We labelthis encoding “ASP-structural” in the tables below. For ASP-structural, we haveonly the instances originally considered in [68].To handle the general version of CTL synthesis (without added structural in-formation), we also created ASP encodings using the methods from De Angelis etal.’s paper, but without problem-specific structural assumptions and optimizations.We label those results “ASP-generic”. For both encodings, we use the latest ver-sion (4.5.4) of Clingo [104], and for each instance we report the best performanceover the included Clasp configurations.8We compare these tools to two versions of MONOSAT: MONOSAT-structural,which uses the wildcard optimization presented in Section 8.5.3, and MONOSAT-generic, without the wildcard optimization.With the exception of CTLSAT, the tools we consider are bounded synthe-sis tools, which take as input both a CTL formula and a maximum number ofstates. For ASP-structural, the state bounds follow [68]. For the remaining tools,we selected the state bound manually, by repeatedly testing each tool with differentbounds, and reporting for each tool the smallest bound for which it found a satis-fying solution. In cases where a tool could not find any satisfying solution withinour time or memory bounds, we report out-of-time or out-of-memory.6https://github.com/nicolaprezza/CTLSAT7http://www.sci.unich.it/∼{}deangelis/papers/mutex FI.tar.gz8 These are: “auto”, “crafty”, “frumpy”, “handy”, “jumpy”, “trendy”, and “tweety”.139Chapter 8. Monotonic Theory of CTL8.6.1 The Original Clarke-Emerson MutexThe mutex problem assumes that there are n processes that run concurrently andon occasion access a single shared resource. Instead of synthesizing entire pro-grams, the original Clarke-Emerson example [56] considers an abstraction of theprograms called synchronization skeletons. In the instance of a mutex algorithm, itis assumed that each process is in one of three states: non-critical section (NCS),the try section (TRY) or the critical section (CS). A process starts in the non-critical section in which it remains until it requests to access the resource, andchanges to the try section. When it finally enters the critical section it has access tothe resource, and eventually loops back to the non-critical section. The synthesisproblem is to find a global Kripke structure for the composition of the n processes,such that the specifications are met. Our first set of benchmarks is based on theClarke and Emerson specification given in [56], that includes mutual exclusionand starvation freedom for all processes.# of ProcessesApproach 2 3 4 5 6CTLSAT TO TO TO TO TOASP-generic 3.6 (7*) 1263.7 (14) TO MEM MEMASP-structural 0.0 (12) 1.2 (36) - - -MONOSAT-gen 0.0 (7*) 1.4 (13*) 438.6 (23*) 1744.9 (42) TOMONOSAT-str 0.2 (7) 0.5 (13) 4.5 (23) 166.7 (41) 1190.5 (75)Table 8.1: Results on the original Clarke-Emerson mutual exclusion example.Table entries are in the format time(states), where states is the number ofstates in the synthesized model, and time is the run time in seconds. ForASP-structural, we only have the manually encoded instances providedby the authors. An asterisk indicates that the tool was able to prove min-imality, by proving the instance is UNSAT at the next lower bound. TOdenotes exceeding the 3hr timeout. MEM denotes exceeding 16GB ofRAM. All experiments were run on a 2.67GHz Intel Xeon x5650 proces-sor.Table 8.1 presents our results on the mutex formulation from [56]. Both ver-sions of MONOSAT scale to much larger instances than the other approaches, find-ing solutions for 5 and 6 processes, respectively. CTLSAT, implementing the clas-140Chapter 8. Monotonic Theory of CTLsical tableau approach, times out on all instances.9 Only the -generic versionscan guarantee minimal solutions, and MONOSAT-generic is able to prove minimalmodels for several cases.As expected, structural constraints greatly improve efficiency for both ASP-structural and MONOSAT-structural relative to their generic counterparts.8.6.2 Mutex with Additional PropertiesAs noted in [121], the original Clarke-Emerson specification permits Kripke struc-tures that are not maximally parallel, or even practically reasonable. For instance,our methods synthesize a structure in which one process being in NCS will blockanother process in TRY from getting the resource — the only transition such aglobal state has is to a state in which both processes are in the TRY section. Inaddition to the original formula, we present results for an augmented version inwhich we eliminate that solution10 by introducing the “Non-Blocking” property,which states that a process may always remain in the NCS:AG (NCSi→ EX NCSi) (NB)In addition, in the original paper there are structural properties implicit in thegiven local transition system, preventing jumping from NCS to CS, or from CS toTRY. We encode these properties into CTL as “No Jump” properties.AG (NCSi→ AX ¬CSi) ∧ AG (CSi→ AX ¬TRYi) (NJ)We also consider two properties from [68]: Bounded Overtaking (BO), whichguarantees that when a process is waiting for the critical section, each other processcan only access the critical section at most once before the first process enters the9Notably, CTLSAT times out even when synthesizing the original 2-process mutex from [56],which Clarke and Emerson originally synthesized by hand. This may be because in that work, thelocal transition system was specified implicitly in the algorithm, instead of in the CTL specificationas it is here.10 While the properties that we introduce in this chapter mitigate some of the effects of underspec-ification, we have observed that the formulas of many instances in our benchmarks are not strongenough to guarantee a sensible solution. We are mainly interested in establishing benchmarks forsynthesis performance, which is orthogonal to the task of finding suitable CTL specifications, whichresolve these problems.141Chapter 8. Monotonic Theory of CTL# of ProcessesApproach 2 3 4 5 6 7Property: ORIG ∧ BOASP-generic 3.4 (7*) 1442.0 (14) TO/MEM MEM MEM MEMASP-structural 0.0 (12) 2.3 (36) - - - -MONOSAT-gen 0.0 (7*) 11.1 (13*) 438.3 (23*) 1286.6 (42) TO TOMONOSAT-str 0.1 (7) 0.6 (13) 5.3 (23) 59.5 (41) 375.3 (75) 10739.5 (141)Property: ORIG ∧ BO ∧MRASP-generic 10.1 (9*) TO MEM MEM MEM MEMASP-structural 0.8 (10) 950.9 (27) - - - -MONOSAT-gen 0.0 (9*) 6.0 (25*) TO TO TO TOMONOSAT-str 0.1 (10) 8.7 (26) TO TO TO TOProperty: ORIG ∧ NB ∧ NJASP-generic 34.8 (9*) TO MEM MEM MEM MEMASP-structural 0.1 (10) 7326.1 (27) - - - -MONOSAT-gen 0.0 (9*) 1275.7 (22*) TO TO TO TOMONOSAT-str 0.2 (10) 1.6 (26) 5314.7 (51) TO TO TOProperty: ORIG ∧ NB ∧ NJ ∧ BOASP-generic 15.4 (9*) TO MEM MEM MEM MEMASP-structural 0.1 (10) TO - - - -MONOSAT-gen 0.0 (9*) 127.7 (22*) TO TO TO TOMONOSAT-str 0.1 (10) 1.3 (24) TO TO TO TOProperty: ORIG ∧ NB ∧ NJ ∧ BO ∧MRASP-generic 10.7 (9*) TO MEM MEM MEM MEMASP-structural 0.1 (10) 1917.6 (27) - - - -MONOSAT-gen 0.0 (9*) 4.4 (25*) TO TO TO TOMONOSAT-str 0.1 (10) 2.7 (26) TO TO TO TOTable 8.2: Results on the mutual exclusion example with additional proper-ties (described in Section 8.6.2). As with Table 8.1, entries are in theformat time(states). ORIG denotes the original mutual exclusion proper-ties from Section 8.6.1. As before, although problem-specific structuralconstraints improve efficiency, MONOSAT-generic is comparably fast toASP-structural on small instances, and scales to larger numbers of pro-cesses. MONOSAT-structural performs even better.142Chapter 8. Monotonic Theory of CTLcritical section, and Maximal Reactivity (MR), which guarantees that if exactly oneprocess is waiting for the critical section, then that process can enter the criticalsection in the next step.In Table 8.2, we repeat our experimental procedure from Section 8.6.1, addingvarious combinations of additional properties. This provides a richer set of bench-marks, most of which are harder than the original. As before, the -structural con-straints greatly improve efficiency, but nevertheless, MONOSAT-generic outper-forms ASP-structural. MONOSAT-generic is able to prove minimality on severalbenchmarks, and on one benchmark, MONOSAT-structural scales to 7 processes.8.6.3 Readers-WritersTo provide even more benchmarks, we present instances of the related Readers-Writers problem [63]. Whereas the Mutex problem assumes that all processesrequire exclusive access to a resource, the Readers-Writers problem permits somesimultaneous access. Two types of processes are distinguished: writers, which re-quire exclusive access, and readers, which can share their access with other readers.This is a typical scenario for concurrent access to shared memory, in which writepermissions and read permissions are to be distinguished. The local states of eachprocess are as in the Mutex instances.We use Attie’s [11] CTL specification. We note however that this specificationallows for models which are not maximally parallel, and in particular disallowsconcurrent access by two readers. In addition to this original formula, we alsoconsider one augmented with the Multiple Readers Eventually Critical (MREC)property. This ensures that there is a way for all readers, if they are in TRY, tosimultaneously enter the critical section, if no writer requests the resource.AG (∧wiNCSwi → (∧riTRYri → EF∧riCSri)) (RW-MREC)This property turns out not to be strong enough to enforce that concurrent ac-cess for readers must always be possible. We introduce the following property,which we call Multiple Readers Critical. It states that if a reader is in TRY, and allother readers are in CS, it is possible to enter the CS in a next state – as long as allwriters are in NCS, since they have priority access over readers.143Chapter 8. Monotonic Theory of CTL# of Processes (# of readers, # of writers)Approach 2 (1, 1) 3 (2, 1) 4 (2, 2) 5 (3, 2) 6 (3, 3) 7 (4, 3)Property: RWCTLSAT TO TO TO TO TO TOASP-generic 0.6 (5*) 9.5 (9*) TO MEM MEM MEMMONOSAT-gen 0.0 (5*) 0.0 (9*) 2.8 (19*) 30.0 (35*) 5312.7 (74) TOMONOSAT-str 0.1 (5) 0.5 (9) 0.7 (19) 2.9 (35) 98.8 (74) 384.4 (142)Property: RW ∧ NB ∧ NJASP-generic 6.8 (8*) 2865.5 (16) MEM MEM MEM MEMMONOSAT-gen 0.0 (8*) 1.4 (16*) 110.4 (27*) 843.8 (46*) TO TOMONOSAT-str 0.1 (9) 0.2 (16) 3.4 (27) 35.9 (54) TOProperty: RW ∧ NB ∧ NJ ∧ RW-MRECASP-generic 2.4 (8*) 120.6 (22) MEM MEM MEM MEMMONOSAT-gen 0.0 (8*) 238.4 (22*) TO TO TO TOMONOSAT-str 0.1 (9) 0.25 (23) 5.3 (52) 159.1 (127) TO TOProperty: RW ∧ NB ∧ NJ ∧ RW-MRCASP-generic 2.4 (8*) TO MEM MEM MEM MEMMONOSAT-gen 0.0 (8*) 1114.1 (22) 18.1 (27*) 251.6 (46*) TO TOMONOSAT-str 0.1 (9) 0.2 (23) 2.5 (28) 28.0 (47) TO TOTable 8.3: Results on the readers-writers instances. Property (RW) is At-tie’s specification [11]. Data is presented as in Table 8.1, in the formattime(states).AG (∧wiNCSwi → (TRYri∧r j 6=riCSr j → EX∧riCSri)) (RW-MRC)Using this property, we are able to synthesize a structure for two readers anda single writer, in which both readers can enter the critical section concurrently,independently of who enters it first, without blocking each other.We ran benchmarks on problem instances of various numbers of readers andwriters, and various combinations of the CTL properties. Since ASP-structuralhas identical process constraints, which make it unsuitable to solve an asymmet-ric problem such as Readers-Writers, we exclude it from these experiments. Aswith the Mutex problem, as CTLSAT is unable to solve even the simplest probleminstances, we do not include benchmarks for the more complex instances.144Chapter 8. Monotonic Theory of CTLOur experiments on each variation of the Readers-Writer problem are presentedin Table 8.3. We observe that in general, Readers-Writers instances are easierto solve than Mutex instances with the same number of processes. At the sametime, the additional properties introduced by us restrict the problem further, andmake the instances harder to solve than the original Readers-Writers formulation.Taken together with the results from Tables 8.1 and 8.2, this comparison furtherstrengthens our argument that MONOSAT-generic scales better than ASP-generic.The results also confirm that the structural MONOSAT solver making use of thewildcard encoding performs much better than MONOSAT-generic.These experiments demonstrate that MONOSAT greatly outperforms existingtools for CTL synthesis. Further, MONOSAT has the ability to solve CTL formulasunder additional constraints (e.g., about the structure of the desired solution), andcan do so without sacrificing generality (by e.g., assuming identical processes).In many cases, we are also able to compute a provably minimal satisfying Kripkestructure.145Chapter 9Conclusions and Future WorkIn this thesis, we have introduced a general framework for building lazy SMTsolvers for finite, monotonic theories. Using this SAT Modulo Monotonic Theoriesframework, we have built a lazy SMT solver (MONOSAT) supporting importantproperties from graph theory, geometry, and automata theory, including many thatpreviously had only poor support in SAT, SMT, or similar constraint solvers.Furthermore, we have provided experimental evidence that SMT solvers builtusing our framework are efficient in practice. In fact, our solvers have signif-icantly advanced the state of the art across several important problem domains,demonstrating real-world applications to circuit layout, data center management,and protocol synthesis.For near-term future work, the monotonic theories already implemented inMONOSAT include many of the most commonly used properties of graph the-ory. Many important, real-world problems could benefit from our graph theorysolver, and we have only just begin to explore them. To name just a few: productline configuration and optimization (e.g., [6, 149, 150]), many problems arising insoftware-defined networks and data centers (e.g., SDN verification [36], configu-ration [120, 155], and repair [206]), road/traffic design [139], and many problemsarising in circuit routing (in addition to escape-routing, which we considered inChapter 6.2).In the longer term, in addition to the monotonic properties we describe in thisthesis, there are many other important properties that can be modeled as finite146Chapter 9. Conclusions and Future Workmonotonic theories, and which are likely amenable to our approach. For example,many important graph properties beyond the ones we already support are mono-tonic, and could be easily supported by our approach (e.g., Hamiltonian cycle de-tection, minimum Steiner tree weight, and many variations of network flow prob-lems, are all monotonic in the edges of a graph). Alternatively, the geometry the-ory we describe in Chapter 7 can likely be generalized into a more comprehensivetheory of constructive solid geometry, with applications to computer-aided design(CAD) tools.A particularly promising monotonic theory, which we have not yet discussed inthis thesis, is a theory of non-deterministic finite state machine string acceptance.Just as our CTL model checking theory is used to synthesize Kripke structures forwhich certain CTL properties do or do not hold, an FSM string acceptance theorycan be used to synthesize finite state machines (of bounded size) that do or do notaccept certain strings. This is a classic NP-hard machine learning problem (see,e.g., [8, 196]), with potential applications to program synthesis.In fact, many properties of non-deterministic state machines, including not onlyfinite state machines but also push-down automata, Lindenmayer-systems [170],and even non-deterministic Turing machines, are monotonic with respect to thetransition system of the state machine.SAT and SMT solvers have proven themselves to be powerful constraintsolvers, with great potential. The techniques in this thesis provide an easy andeffective way to extend their reach to many new domains.147Bibliography[1] IBM Platform Resouce Scheduler,http://www.ibm.com/support/knowledgecenter/ss8mu9 2.4.0/prs kc/prs kc administering.html,(accessed 01-09-2016). URL http://www.ibm.com/support/knowledgecenter/SS8MU9 2.4.0/prs kc/prs kc administering.html. →pages 94[2] OpenStack Nova scheduler, http://docs.openstack.org/juno/config-reference/content/section compute-scheduler.html, (accessed 01-09-2016).URL http://docs.openstack.org/juno/config-reference/content/section compute-scheduler.html. → pages[3] VMware Distributed Resource Management: Design, Implementation, andLessons Learned, labs.vmware.com/vmtj/vmware-distributed-resource-management-design-implementation-and-lessons-learned, (accessed05-10-2016). URL labs.vmware.com/vmtj/vmware-distributed-resource-management-design-implementation-and-lessons-learned.→ pages 94[4] F. A. Aloul, A. Ramani, I. L. Markov, and K. A. Sakallah. Solving difficultSAT instances in the presence of symmetry. In Proceedings of the 39thannual Design Automation Conference, pages 731–736. ACM, 2002. →pages 69[5] F. A. Aloul, B. Al-Rawi, and M. Aboelaze. Routing and wavelengthassignment in optical networks using Boolean satisfiability. In Proceedingsof the Consumer Communications and Networking Conference, pages185–189. IEEE, 2008. → pages 62, 69[6] N. Andersen, K. Czarnecki, S. She, and A. Wasowski. Efficient synthesis offeature models. In Proceedings of the 16th International Software ProductLine Conference-Volume 1, pages 106–115. ACM, 2012. → pages 62, 146148Bibliography[7] A. M. Andrew. Another efficient algorithm for convex hulls in twodimensions. Information Processing Letters, 9(5):216–219, 1979. → pages110[8] D. Angluin. Learning regular sets from queries and counterexamples.Information and Computation, 75(2):87–106, 1987. → pages 147[9] A. Armando, C. Castellini, and E. Giunchiglia. SAT-based procedures fortemporal reasoning. In Recent Advances in AI Planning, pages 97–108.Springer, 2000. → pages 17[10] A. Armando, C. Castellini, E. Giunchiglia, and M. Maratea. A SAT-baseddecision procedure for the Boolean combination of difference constraints.In Proceedings of the International Conference on Theory and Applicationsof Satisfiability Testing, pages 16–29. Springer, 2005. → pages 20, 178[11] P. C. Attie. Synthesis of large concurrent programs via pairwisecomposition. In Proceedings of the International Conference onConcurrency Theory, pages 130–145. Springer, 1999. → pages 125, 143,144[12] P. C. Attie and E. A. Emerson. Synthesis of concurrent systems with manysimilar processes. Transactions on Programming Languages and Systems(TOPLAS), 20(1):51–115, 1998. → pages 124, 125, 126, 137[13] P. C. Attie and E. A. Emerson. Synthesis of concurrent programs for anatomic read/write model of computation. Transactions on ProgrammingLanguages and Systems (TOPLAS), 23(2):187–242, 2001. → pages 123[14] G. Audemard and L. Simon. Predicting learnt clauses quality in modernSAT solvers. In Proceedings of the International Joint Conference onArtificial Intelligence, volume 9, pages 399–404, 2009. → pages 12, 64, 67[15] G. Audemard and L. Simon. Refining restarts strategies for SAT andUNSAT. In Principles and Practice of Constraint Programming, pages118–126. Springer, 2012. → pages 12[16] G. Audemard, P. Bertoli, A. Cimatti, A. Korniłowicz, and R. Sebastiani. ASAT based approach for solving formulas over Boolean and linearmathematical propositions. In Proceedings of the International Conferenceon Automated Deduction, pages 195–210. Springer, 2002. → pages 17, 20149Bibliography[17] G. Audemard, J.-M. Lagniez, B. Mazure, and L. Saı¨s. Boosting localsearch thanks to CDCL. In International Conference on Logic forProgramming Artificial Intelligence and Reasoning, pages 474–488.Springer, 2010. → pages 7[18] Autodesk. AutoCAD Version 1.0. → pages 1[19] L. Bachmair, A. Tiwari, and L. Vigneron. Abstract congruence closure.Journal of Automated Reasoning, 31(2):129–168, 2003. → pages 62[20] G. J. Badros, A. Borning, and P. J. Stuckey. The Cassowary lineararithmetic constraint solving algorithm. ACM Transactions onComputer-Human Interaction (TOCHI), 8(4):267–306, 2001. → pages 1[21] O. Bailleux and Y. Boufkhad. Efficient CNF encoding of Booleancardinality constraints. In Principles and Practice of ConstraintProgramming. Springer, 2003. → pages 98[22] H. Ballani, P. Costa, T. Karagiannis, and A. Rowstron. Towards predictabledatacenter networks. In SIGCOMM Computer Communication Review,2011. → pages 94[23] K. Bansal, A. Reynolds, T. King, C. Barrett, and T. Wies. Deciding localtheory extensions via e-matching. In International Conference onComputer Aided Verification, pages 87–105. Springer, 2015. → pages 23[24] R. Bar-Yehuda and S. Even. A local-ratio theorem for approximating theweighted vertex cover problem. North-Holland Mathematics Studies, 1985.→ pages 98[25] C. Baral. Knowledge representation, reasoning and declarative problemsolving. Cambridge University Press, 2003. → pages 63, 77[26] M. F. Bari, R. Boutaba, R. Esteves, L. Z. Granville, M. Podlesny, M. G.Rabbani, Q. Zhang, and M. F. Zhani. Data center network virtualization: Asurvey. Communications Surveys & Tutorials, 15(2):909–928, 2013. →pages 94[27] C. Barrett, D. Dill, and J. Levitt. Validity checking for combinations oftheories with equality. In Proceedings of the International Conference onFormal Methods in Computer-Aided Design, pages 187–201. Springer,1996. → pages 17150Bibliography[28] C. Barrett, R. Nieuwenhuis, A. Oliveras, and C. Tinelli. Splitting ondemand in SAT modulo theories. In Logic for Programming, ArtificialIntelligence, and Reasoning, pages 512–526. Springer, 2006. → pages 35[29] C. Barrett, I. Shikanian, and C. Tinelli. An abstract decision procedure fora theory of inductive data types. Journal on Satisfiability, BooleanModeling and Computation, 3:21–46, 2007. → pages 1[30] C. Barrett, C. L. Conway, M. Deters, L. Hadarean, D. Jovanovic´, T. King,A. Reynolds, and C. Tinelli. CVC4. In Proceedings of the InternationalConference on Computer-Aided Verification, 2011. → pages 173[31] C. W. Barrett, D. L. Dill, and A. Stump. Checking satisfiability offirst-order formulas by incremental translation to SAT. In Proceedings ofthe International Conference on Computer-Aided Verification, pages236–249. Springer, 2002. → pages 17[32] R. J. Bayardo Jr and R. Schrag. Using CSP look-back techniques to solvereal-world SAT instances. In Proceedings of the AAAI Conference onArtificial Intelligence, pages 203–208, 1997. → pages 11, 12[33] S. Bayless, N. Bayless, H. Hoos, and A. Hu. SAT Modulo MonotonicTheories. In Proceedings of the 29th AAAI Conference on ArtificialIntelligence, 2015. → pages 29, 68[34] A. Biere. Lingeling, Plingeling and Treengeling. Proceedings of SATCompetition, 2013. → pages 64, 67[35] A. Biere, A. Cimatti, E. Clarke, and Y. Zhu. Symbolic model checkingwithout BDDs. Springer, 1999. → pages 1, 60[36] N. Bjørner and K. Jayaraman. Checking cloud contracts in MicrosoftAzure. In International Conference on Distributed Computing and InternetTechnology, pages 21–32. Springer, 2015. → pages 146[37] G. Boenn, M. Brain, M. De Vos, et al. Automatic composition of melodicand harmonic music by answer set programming. In Logic Programming,pages 160–174. Springer, 2008. → pages 77[38] J. F. Botero, X. Hesselbach, M. Duelli, D. Schlosser, A. Fischer, andH. De Meer. Energy efficient virtual network embedding. IEEECommunications Letters, 16(5):756–759, 2012. → pages 69151Bibliography[39] M. Bozzano, R. Bruttomesso, A. Cimatti, T. Junttila, P. Van Rossum,S. Schulz, and R. Sebastiani. MathSAT: Tight integration of SAT andmathematical decision procedures. Journal of Automated Reasoning, 35(1-3):265–293, 2005. → pages 20[40] A. R. Bradley. SAT-based model checking without unrolling. InVerification, Model Checking, and Abstract Interpretation, pages 70–87.Springer, 2011. → pages 1[41] A. R. Bradley and Z. Manna. Property-directed incremental invariantgeneration. Formal Aspects of Computing, 20(4-5):379–405, 2008. →pages 26, 60[42] R. Brummayer and A. Biere. Boolector: An efficient SMT solver forbit-vectors and arrays. In Proceedings of the International Conference onTools and Algorithms for the Construction and Analysis of Systems, pages174–177. Springer, 2009. → pages 18, 173[43] G. Bruns and P. Godefroid. Model checking partial state spaces with3-valued temporal logics. In Proceedings of the International Conferenceon Computer-Aided Verification. Springer, 1999. → pages 23, 130[44] R. Bruttomesso, A. Cimatti, A. Franze´n, A. Griggio, and R. Sebastiani. TheMathSAT 4 SMT solver. In Proceedings of the International Conference onComputer-Aided Verification, pages 299–303. Springer, 2008. → pages 1[45] R. E. Bryant. Graph-based algorithms for Boolean function manipulation.Computers, IEEE Transactions on, 100(8):677–691, 1986. → pages 7[46] R. E. Bryant, S. K. Lahiri, and S. A. Seshia. Modeling and verifyingsystems using a logic of counter arithmetic with lambda expressions anduninterpreted functions. In Proceedings of the International Conference onComputer-Aided Verification, pages 78–92. Springer, 2002. → pages 17[47] J. Burch, E. Clarke, K. McMillan, D. Dill, and L. Hwang. Symbolic modelchecking: 1020 states and beyond. Information and Computation, 98(2):142–170, 1992. → pages 60[48] L. S. Buriol, M. G. Resende, and M. Thorup. Speeding up dynamicshortest-path algorithms. INFORMS Journal on Computing, 20(2):191–204, 2008. → pages 61, 187[49] D. Bustan and O. Grumberg. Simulation-based minimization. Transactionson Computational Logic, 2003. → pages 125152Bibliography[50] C. Cadar, D. Dunbar, and D. R. Engler. Klee: Unassisted and automaticgeneration of high-coverage tests for complex systems programs. InProceedings of the USENIX Symposium on Operating Systems Design andImplementation, volume 8, pages 209–224, 2008. → pages 1[51] C. Carathe´odory. U¨ber den variabilita¨tsbereich der fourierschen konstantenvon positiven harmonischen funktionen. Rendiconti del CircoloMatematico di Palermo (1884-1940), 32(1):193–217, 1911. → pages 113,194[52] W.-T. Chan, F. Y. Chin, and H.-F. Ting. Escaping a grid by edge-disjointpaths. Algorithmica, 2003. → pages 86[53] J. Chen. A new SAT encoding of the at-most-one constraint. ConstraintModelling and Reformulation, 2010. → pages 98[54] N. Chowdhury, M. R. Rahman, and R. Boutaba. Virtual networkembedding with coordinated node and link mapping. In Proceedings of theInternational Conference on Computer Communications, pages 783–791.IEEE, 2009. → pages 95, 96[55] V. Chvatal. A combinatorial theorem in plane geometry. Journal ofCombinatorial Theory, Series B, 18(1):39–41, 1975. → pages 119[56] E. Clarke and E. Emerson. Design and synthesis of synchronizationskeletons using branching time temporal logic. Logics of Programs, 1982.→ pages 122, 123, 124, 125, 139, 140, 141[57] E. Clarke, D. Kroening, and F. Lerda. A tool for checking ANSI-Cprograms. In Proceedings of the International Conference on Tools andAlgorithms for the Construction and Analysis of Systems, pages 168–176.Springer, 2004. → pages 1[58] E. M. Clarke and E. A. Emerson. Design and synthesis of synchronizationskeletons using branching time temporal logic. In Workshop on Logic ofPrograms, pages 52–71. Springer, 1981. → pages 60[59] E. Coban, E. Erdem, and F. Ture. Comparing ASP, CP, ILP on twochallenging applications: Wire routing and haplotype inference.Proceedings of the International Workshop on Logic and Search, 2008. →pages 86[60] T. S. B. committee. Sort benchmark. http://sortbenchmark.org/, (accessed26-01-2016). → pages 104153Bibliography[61] S. A. Cook. The complexity of theorem-proving procedures. InProceedings of the third annual ACM symposium on Theory of Computing,pages 151–158. ACM, 1971. → pages 3[62] T. H. Cormen, C. E. Leiserson, R. L. Rivest, and C. Stein. Introduction toAlgorithms, volume 6. MIT press Cambridge, 2001. → pages 71[63] P. J. Courtois, F. Heymans, and D. L. Parnas. Concurrent control withreaders and writers. Communications of the ACM, 1971. → pages 143[64] I. Cplex. Cplex 11.0 users manual. ILOG SA, Gentilly, France, page 32,2007. → pages 69[65] J. Cussens. Bayesian network learning by compiling to weightedMAX-SAT. In Proceedings of the Conference on Uncertainty in ArtificialIntelligence, 2008. → pages 66[66] M. Davis and H. Putnam. A computing procedure for quantification theory.Journal of the ACM (JACM), 7(3):201–215, 1960. → pages 6[67] M. Davis, G. Logemann, and D. Loveland. A machine program fortheorem-proving. Communications of the ACM, 5(7):394–397, 1962. →pages 6[68] E. De Angelis, A. Pettorossi, and M. Proietti. Synthesizing concurrentprograms using answer set programming. Fundamenta Informaticae, 120(3-4):205–229, 2012. → pages 124, 126, 127, 136, 137, 139, 141[69] M. De Berg, M. Van Kreveld, M. Overmars, and O. C. Schwarzkopf.Computational Geometry. Springer, 2000. → pages 111[70] L. De Moura and N. Bjorner. Z3: An efficient SMT solver. In Proceedingsof the International Conference on Tools and Algorithms for theConstruction and Analysis of Systems. 2008. → pages 1, 21, 30, 64, 100,170, 173[71] L. de Moura and N. Bjørner. Satisfiability modulo theories: An appetizer.In Formal Methods: Foundations and Applications, pages 23–36. Springer,2009. → pages 17[72] L. De Moura, H. Rueß, and M. Sorea. Lazy theorem proving for boundedmodel checking over infinite domains. In Proceedings of the InternationalConference on Automated Deduction, pages 438–455. Springer, 2002. →pages 17154Bibliography[73] L. De Moura, H. Rueß, and M. Sorea. Lemmas on demand for satisfiabilitysolvers. In Proceedings of the International Conference on Theory andApplications of Satisfiability Testing, 2002. → pages 17[74] R. Dechter. Enhancement schemes for constraint processing:Backjumping, learning, and cutset decomposition. Artificial Intelligence,41(3):273–312, 1990. → pages 11[75] C. Demetrescu and G. F. Italiano. Fully dynamic all pairs shortest pathswith real edge weights. In Proceedings of the 42nd IEEE Symposium onFoundations of Computer Science, pages 260–267. IEEE, 2001. → pages66[76] G. Dequen and O. Dubois. Kcnfs: An efficient solver for random k-SATformulae. In Proceedings of the International Conference on Theory andApplications of Satisfiability Testing, pages 486–501. Springer, 2004. →pages 6[77] D. Detlefs, G. Nelson, and J. B. Saxe. Simplify: a theorem prover forprogram checking. Journal of the ACM (JACM), 52(3):365–473, 2005. →pages 20[78] B. L. Dietrich and A. J. Hoffman. On greedy algorithms, partially orderedsets, and submodular functions. IBM Journal of Research andDevelopment, 47(1):25, 2003. → pages 49[79] E. W. Dijkstra. A note on two problems in connexion with graphs.Numerische mathematik, 1(1):269–271, 1959. → pages 61, 187[80] B. Dutertre and L. De Moura. A fast linear-arithmetic solver for DPLL (T).In Proceedings of the International Conference on Computer-AidedVerification, pages 81–94. Springer, 2006. → pages 1, 33[81] B. Dutertre and L. De Moura. The Yices SMT solver. Tool paper athttp://yices. csl. sri. com/tool-paper. pdf, 2:2, 2006. → pages 1[82] D. East and M. Truszczynski. More on wire routing with ASP. InWorkshop on ASP, 2001. → pages 86[83] N. Ee´n and A. Biere. Effective preprocessing in SAT through variable andclause elimination. In Proceedings of the International Conference onTheory and Applications of Satisfiability Testing, pages 61–75. Springer,2005. → pages 12, 169155Bibliography[84] N. Ee´n and N. So¨rensson. An extensible SAT-solver. In Proceedings of theInternational Conference on Theory and Applications of SatisfiabilityTesting, pages 333–336. Springer, 2004. → pages 11, 64, 67, 77, 169[85] N. Ee´n and N. So¨rensson. Translating pseudo-Boolean constraints intoSAT. Journal on Satisfiability, Boolean Modeling and Computation, 2(1-4):1–26, 2006. → pages 98[86] P. Elias, A. Feinstein, and C. Shannon. A note on the maximum flowthrough a network. Transactions on Information Theory, 2(4):117–119,1956. → pages 72[87] E. A. Emerson and J. Y. Halpern. Decision procedures and expressivenessin the temporal logic of branching time. In Symposium on Theory ofComputing, pages 169–180. ACM, 1982. → pages 123[88] E. Erdem and M. D. Wong. Rectilinear Steiner tree construction usinganswer set programming. In Proceedings of the International Conferenceon Logic Programming, 2004. → pages 86[89] E. Erdem, V. Lifschitz, and M. D. Wong. Wire routing and satisfiabilityplanning. In Computational Logic. 2000. → pages 86[90] A. Erez and A. Nadel. Finding bounded path in graph using SMT forautomatic clock routing. In International Conference on Computer AidedVerification, pages 20–36. Springer, 2015. → pages 56, 62, 67, 81, 86[91] C. Ericson. Real-time collision detection. Elsevier Amsterdam/Boston,2005. → pages 110, 114[92] S. Even, A. Itai, and A. Shamir. On the complexity of time table andmulti-commodity flow problems. In Proceedings of the 16th AnnualSymposium on Foundations of Computer Science, 1975. → pages 95[93] J.-W. Fang, I.-J. Lin, P.-H. Yuh, Y.-W. Chang, and J.-H. Wang. A routingalgorithm for flip-chip design. In Proceedings of the InternationalConference on ComputerAided Design, 2005. → pages 69, 86[94] J.-W. Fang, I.-J. Lin, Y.-W. Chang, and J.-H. Wang. A network-flow-basedRDL routing algorithms for flip-chip design. Transactions onComputer-Aided Design of Integrated Circuits and Systems, 2007. →pages 86156Bibliography[95] W. R. Franklin. Pnpoly-point inclusion in polygon test. Web site:http://www. ecse. rpi. edu/Homepages/wrf/Research/Short Notes/pnpoly.html, 2006. → pages 113, 194[96] M. Fra¨nzle and C. Herde. Hysat: An efficient proof engine for boundedmodel checking of hybrid systems. Formal Methods in System Design, 30(3):179–198, 2007. → pages 23[97] J. W. Freeman. Improvements to propositional satisfiability searchalgorithms. PhD thesis, Citeseer, 1995. → pages 6[98] A. Fro¨hlich, A. Biere, C. M. Wintersteiger, and Y. Hamadi. Stochastic localsearch for satisfiability modulo theories. In Proceedings of the AAAIConference on Artificial Intelligence, pages 1136–1143, 2015. → pages 18[99] H. N. Gabow and R. E. Tarjan. A linear-time algorithm for a special case ofdisjoint set union. Journal of computer and system sciences, 30(2):209–221, 1985. → pages 193[100] V. Ganesh and D. L. Dill. A decision procedure for bit-vectors and arrays.In Proceedings of the International Conference on Computer-AidedVerification, pages 519–531. Springer, 2007. → pages 1, 17, 173[101] H. Ganzinger, G. Hagen, R. Nieuwenhuis, A. Oliveras, and C. Tinelli.DPLL (T): Fast decision procedures. In Proceedings of the InternationalConference on Computer-Aided Verification, pages 175–188. Springer,2004. → pages 17, 18, 19, 22[102] M. N. V. P. Gao. Efficient pseudo-Boolean satisfiability encodings forrouting and wavelength assignment in optical networks. In Proceedings ofthe Ninth Symposium on Abstraction, Reformulation and Approximation,2011. → pages 62, 69[103] C. Ge, F. Ma, J. Huang, and J. Zhang. SMT solving for the theory ofordering constraints. In International Workshop on Languages andCompilers for Parallel Computing, pages 287–302. Springer, 2015. →pages 66[104] M. Gebser, B. Kaufmann, A. Neumann, and T. Schaub. clasp: Aconflict-driven answer set solver. In Logic Programming andNonmonotonic Reasoning, pages 260–265. Springer, 2007. → pages 63,77, 126, 139157Bibliography[105] M. Gebser, T. Janhunen, and J. Rintanen. Answer set programming as SATmodulo acyclicity. In Proceedings of the Twenty-first European Conferenceon Artificial Intelligence (ECAI14), 2014. → pages 66[106] M. Gebser, T. Janhunen, and J. Rintanen. SAT modulo graphs: Acyclicity.In Logics in Artificial Intelligence, pages 137–151. Springer, 2014. →pages 56, 63, 66[107] F. Giunchiglia and R. Sebastiani. Building decision procedures for modallogics from propositional decision proceduresthe case study of modal K. InInternational Conference on Automated Deduction, pages 583–597.Springer, 1996. → pages 17[108] F. Giunchiglia and R. Sebastiani. A SAT-based decision procedure forALC. 1996. → pages 20[109] E. Goldberg and Y. Novikov. Berkmin: A fast and robust SAT-solver.Discrete Applied Mathematics, 155(12):1549–1561, 2007. → pages 9[110] R. Goldman. Intersection of two lines in three-space. In Graphics Gems,page 304. Academic Press Professional, 1990. → pages 115, 195[111] D. Goldwasser, O. Strichman, and S. Fine. A theory-based decisionheuristic for DPLL (T). In Proceedings of the International Conference onFormal Methods in Computer-Aided Design, page 13. IEEE Press, 2008.→ pages 22[112] A. Griggio, Q.-S. Phan, R. Sebastiani, and S. Tomasi. Stochastic localsearch for SMT: combining theory solvers with walkSAT. In Frontiers ofCombining Systems, pages 163–178. Springer, 2011. → pages 18[113] C. Guo, G. Lu, H. J. Wang, S. Yang, C. Kong, P. Sun, W. Wu, andY. Zhang. Secondnet: a data center network virtualization architecture withbandwidth guarantees. In Proceedings of the 6th International COnference(CONEXT), page 15. ACM, 2010. → pages 94, 99, 100, 101, 102[114] A. Gupta, J. Kleinberg, A. Kumar, R. Rastogi, and B. Yener. Provisioning avirtual private network: a network design problem for multicommodityflow. In Proceedings of the Thirty-Third Annual ACM Symposium onTheory of Computing, pages 389–398, 2001. → pages 95[115] A. Gupta, A. E. Casavant, P. Ashar, A. Mukaiyama, K. Wakabayashi, andX. Liu. Property-specific testbench generation for guided simulation. In158BibliographyProceedings of the 2002 Asia and South Pacific Design AutomationConference, page 524. IEEE Computer Society, 2002. → pages 23, 130[116] M. R. Henzinger and V. King. Fully dynamic biconnectivity and transitiveclosure. In Proceedings of the 36th Annual Symposium on Foundations ofComputer Science, pages 664–672. IEEE, 1995. → pages 66[117] M. R. Henzinger, P. Klein, S. Rao, and S. Subramanian. Fastershortest-path algorithms for planar graphs. Journal of Computer andSystem Sciences, 55(1):3–23, 1997. → pages 66[118] M. Heule, M. Dufour, J. Van Zwieten, and H. Van Maaren. March eq:implementing additional reasoning into an efficient look-ahead SAT solver.In Proceedings of the International Conference on Theory and Applicationsof Satisfiability Testing, pages 345–359. Springer, 2005. → pages 6[119] M. J. Heule, O. Kullmann, S. Wieringa, and A. Biere. Cube and conquer:Guiding CDCL SAT solvers by lookaheads. In Proceedings of the HaifaVerification Conference, pages 50–65. Springer, 2011. → pages 7[120] J. A. Hewson, P. Anderson, and A. D. Gordon. A declarative approach toautomated configuration. In Large Installation System AdministrationConference, volume 12, pages 51–66, 2012. → pages 146[121] S. Heymans, D. Van Nieuwenborgh, and D. Vermeir. Synthesis fromtemporal specifications using preferred answer set programming. InTheoretical Computer Science, pages 280–294. Springer, 2005. → pages124, 126, 127, 141[122] T. Hickey, Q. Ju, and M. H. Van Emden. Interval arithmetic: Fromprinciples to implementation. Journal of the ACM (JACM), 48(5):1038–1068, 2001. → pages 23[123] Y.-K. Ho, H.-C. Lee, and Y.-W. Chang. Escape routing forstaggered-pin-array PCBs. Transactions on Computer-Aided Design ofIntegrated Circuits and Systems, 2011. → pages 69[124] F. Hoffmann. On the rectilinear art gallery problem. In InternationalColloquium on Automata, Languages, and Programming, pages 717–728.Springer, 1990. → pages 119[125] I. D. Horswill and L. Foged. Fast procedural level population withplayability constraints. In Proceedings of the AAAI Conference on ArtificialIntelligence and Interactive Digital Entertainment, 2012. → pages 63159Bibliography[126] I. Houidi, W. Louati, W. B. Ameur, and D. Zeghlache. Virtual networkprovisioning across multiple substrate networks. Computer Networks, 55(4):1011–1023, 2011. → pages 69[127] C.-H. Hsu, H.-Y. Chen, and Y.-W. Chang. Multi-layer global routingconsidering via and wire capacities. In Proceedings of the InternationalConference on ComputerAided Design, 2008. → pages 85[128] M. Huth and M. Ryan. Logic in Computer Science: Modelling andReasoning about Systems. 1999. → pages 126, 128, 131[129] D. Jackson. Automating first-order relational logic. In ACM SIGSOFTSoftware Engineering Notes, volume 25, pages 130–139. ACM, 2000. →pages 62[130] D. Jackson. Alloy: a lightweight object modelling notation. ACMTransactions on Software Engineering and Methodology (TOSEM), 11(2):256–290, 2002. → pages 62[131] D. Jackson, I. Schechter, and I. Shlyakhter. Alcoa: the alloy constraintanalyzer. In Proceedings of the International Conference on SoftwareEngineering, pages 730–733. IEEE, 2000. → pages 62[132] S. Jacobs and R. Bloem. Parameterized synthesis. In Proceedings of theInternational Conference on Tools and Algorithms for the Construction andAnalysis of Systems, pages 362–376. Springer, 2012. → pages 124[133] T. Janhunen, S. Tasharrofi, and E. Ternovska. SAT-TO-SAT: Declarativeextension of SAT solvers with new propagators. In Proceedings of theAAAI Conference on Artificial Intelligence, 2016. → pages 23, 56, 62[134] M. Ja¨rvisalo, D. Le Berre, O. Roussel, and L. Simon. The internationalSAT solver competitions. AI Magazine, 33(1):89–92, 2012. → pages 6[135] R. G. Jeroslow and J. Wang. Solving propositional satisfiability problems.Annals of mathematics and Artificial Intelligence, 1(1-4):167–187, 1990.→ pages 9[136] S. Jha, S. Gulwani, S. A. Seshia, and A. Tiwari. Oracle-guidedcomponent-based program synthesis. In Proceedings of the InternationalConference on Software Engineering, volume 1, pages 215–224. IEEE,2010. → pages 1160Bibliography[137] R. M. Karp, F. T. Leighton, R. L. Rivest, C. D. Thompson, U. V. Vazirani,and V. V. Vazirani. Global wire routing in two-dimensional arrays.Algorithmica, 1987. → pages 85[138] H. A. Kautz, B. Selman, et al. Planning as satisfiability. In Proceedings ofthe European Conference on Artificial Intelligence, volume 92, pages359–363, 1992. → pages 1[139] B. Kim, A. Jarandikar, J. Shum, S. Shiraishi, and M. Yamaura. TheSMT-based automatic road network generation in vehicle simulationenvironment. In Proceedings of the 13th International Conference onEmbedded Software, page 18. ACM, 2016. → pages 146[140] P. Kohli and P. H. Torr. Efficiently solving dynamic Markov random fieldsusing graph cuts. In Proceedings of the International Conference onComputer Vision, volume 2, pages 922–929. IEEE, 2005. → pages 61, 70,188, 190[141] D.-T. Lee and A. Lin. Computational complexity of art gallery problems.Information Theory, IEEE Transactions on, 32(2):276–282, 1986. → pages119[142] W. Lee, A. Pardo, J.-Y. Jang, G. Hachtel, and F. Somenzi. Tearing basedautomatic abstraction for CTL model checking. In Proceedings of the 1996IEEE/ACM international conference on Computer-aided design, pages76–81. IEEE Computer Society, 1997. → pages 23, 130[143] L. Luo and M. D. Wong. Ordered escape routing based on Booleansatisfiability. In Proceedings of the Asia and South Pacific DesignAutomation Conference, 2008. → pages 86[144] M. Mahfoudh, P. Niebert, E. Asarin, and O. Maler. A satisfiability checkerfor difference logic. Proceedings of the International Conference onTheory and Applications of Satisfiability Testing, 2:222–230, 2002. →pages 33, 178[145] J. Marques-Silva, M. Janota, and A. Belov. Minimal sets over monotonepredicates in Boolean formulae. In Proceedings of the InternationalConference on Computer-Aided Verification, pages 592–607. Springer,2013. → pages 26[146] J. P. Marques-Silva and K. A. Sakallah. GRASP: A search algorithm forpropositional satisfiability. IEEE Transactions on Computers, 48(5):506–521, 1999. → pages 4, 6, 8, 11161Bibliography[147] A. Martin. Adequate sets of temporal connectives in CTL. ElectronicNotes in Theoretical Computer Science, 2002. EXPRESS’01, 8thInternational Workshop on Expressiveness in Concurrency. → pages 127[148] K. L. McMillan. Interpolation and SAT-based model checking. InProceedings of the International Conference on Computer-AidedVerification, pages 1–13. Springer, 2003. → pages 1[149] M. Mendonc¸a. Efficient reasoning techniques for large scale featuremodels. PhD thesis, University of Waterloo, 2009. → pages 62, 146[150] M. Mendonca, A. Wasowski, and K. Czarnecki. SAT-based analysis offeature models is easy. In Proceedings of the 13th International SoftwareProduct Line Conference, pages 231–240. Carnegie Mellon University,2009. → pages 62, 146[151] M. W. Moskewicz, C. F. Madigan, Y. Zhao, L. Zhang, and S. Malik. Chaff:Engineering an efficient SAT solver. In Proceedings of the 38th annualDesign Automation Conference, pages 530–535. ACM, 2001. → pages 4,6, 8[152] A. Nadel. Routing under constraints. 2016. → pages 67, 86[153] G.-J. Nam, K. A. Sakallah, and R. A. Rutenbar. A new FPGA detailedrouting approach via search-based Boolean satisfiability. Transactions onComputer-Aided Design of Integrated Circuits and Systems, 21(6):674–684, 2002. → pages 62[154] G.-J. Nam, F. Aloul, K. A. Sakallah, and R. A. Rutenbar. A comparativestudy of two Boolean formulations of FPGA detailed routing constraints.Transactions on Computers, 53(6):688–696, 2004. → pages 62[155] S. Narain, G. Levin, S. Malik, and V. Kaul. Declarative infrastructureconfiguration synthesis and debugging. Journal of Network and SystemsManagement, 16(3):235–258, 2008. → pages 146[156] G. Nelson and D. C. Oppen. Simplification by cooperating decisionprocedures. Transactions on Programming Languages and Systems(TOPLAS), 1(2):245–257, 1979. → pages 52[157] G. Nelson and D. C. Oppen. Fast decision procedures based on congruenceclosure. Journal of the ACM (JACM), 27(2):356–364, 1980. → pages 62162Bibliography[158] A. Niemetz, M. Preiner, and A. Biere. Precise and complete propagationbased local search for satisfiability modulo theories. In InternationalConference on Computer Aided Verification, pages 199–217. Springer,2016. → pages 18[159] R. Nieuwenhuis and A. Oliveras. Proof-producing congruence closure. InInternational Conference on Rewriting Techniques and Applications, pages453–468. Springer, 2005. → pages 62[160] R. Nieuwenhuis and A. Oliveras. Fast congruence closure and extensions.Information and Computation, 205(4):557–580, 2007. → pages 62[161] R. Nieuwenhuis, A. Oliveras, and C. Tinelli. Abstract DPLL and abstractDPLL modulo theories. In Logic for Programming, Artificial Intelligence,and Reasoning, pages 36–50. Springer, 2005. → pages 17, 18, 22[162] G. Optimization et al. Gurobi optimizer reference manual. URL:http://www.gurobi.com, 2:1–3, 2012. → pages 69[163] V. Paruthi and A. Kuehlmann. Equivalence checking combining a structuralSAT-solver, BDDs, and simulation. In Proceedings of the InternationalConference on Computer Design, pages 459–464. IEEE, 2000. → pages 6[164] D. J. Pearce and P. H. Kelly. A dynamic topological sort algorithm fordirected acyclic graphs. Journal of Experimental Algorithmics (JEA), 11:1–7, 2007. → pages 65, 188[165] L. Perron. Operations research and constraint programming at Google. InProceedings of the International Conference on Principles and Practice ofConstraint Programming. 2011. → pages 89[166] K. Pipatsrisawat and A. Darwiche. A lightweight component cachingscheme for satisfiability solvers. In Proceedings of the InternationalConference on Theory and Applications of Satisfiability Testing, pages294–299. Springer, 2007. → pages 9[167] D. A. Plaisted and S. Greenbaum. A structure-preserving clause formtranslation. Journal of Symbolic Computation, 2(3):293–304, 1986. →pages 6[168] M. R. Prasad, A. Biere, and A. Gupta. A survey of recent advances inSAT-based formal verification. International Journal on Software Tools forTechnology Transfer, 7(2):156–173, 2005. → pages 7163Bibliography[169] P. Prosser. Hybrid algorithms for the constraint satisfaction problem.Computational intelligence, 9(3):268–299, 1993. → pages 12[170] P. Prusinkiewicz and A. Lindenmayer. The algorithmic beauty of plants.Springer Science & Business Media, 2012. → pages 147[171] G. Ramalingam and T. Reps. An incremental algorithm for a generalizationof the shortest-path problem. Journal of Algorithms, 21(2):267–305, 1996.→ pages 61, 187, 190[172] D. Ranjan, D. Tang, and S. Malik. A comparative study of 2QBFalgorithms. In The Seventh International Conference on Theory andApplications of Satisfiability Testing (SAT 2004), pages 292–305, 2004. →pages 23[173] J. C. Reynolds. Separation logic: A logic for shared mutable datastructures. In Proceedings of the 17th Annual Symposium on Logic inComputer Science, pages 55–74. IEEE, 2002. → pages 1[174] J. Rintanen. Madagascar: Efficient planning with SAT. The 2011International Planning Competition, page 61, 2011. → pages 1[175] J. Rintanen, K. Heljanko, and I. Niemela¨. Parallel encodings of classicalplanning as satisfiability. In European Workshop on Logics in ArtificialIntelligence, pages 307–319. Springer, 2004. → pages 66[176] M. Rost, C. Fuerst, and S. Schmid. Beyond the stars: Revisiting virtualcluster embeddings. In SIGCOMM Computer Communication Review,2015. → pages 94[177] N. Ryzhenko and S. Burns. Standard cell routing via Boolean satisfiability.In Proceedings of the 49th Annual Design Automation Conference, pages603–612. ACM, 2012. → pages 62[178] S. Schewe and B. Finkbeiner. Bounded synthesis. In AutomatedTechnology for Verification and Analysis, pages 474–488. Springer, 2007.→ pages 124[179] R. Sebastiani. Lazy satisfiability modulo theories. Journal on Satisfiability,Boolean Modeling and Computation, 3:141–224, 2007. → pages 18, 21,30, 136, 169[180] B. Selman, H. Kautz, B. Cohen, et al. Local search strategies forsatisfiability testing. Cliques, coloring, and satisfiability: Second DIMACSimplementation challenge, 26:521–532, 1993. → pages 4, 6164Bibliography[181] A. Smith. Personal communication, 2014. → pages 73[182] A. M. Smith and M. Mateas. Variations forever: Flexibly generatingrulesets from a sculptable design space of mini-games. In Symposium onComputational Intelligence and Games (CIG), pages 273–280. IEEE,2010. → pages 63, 77[183] A. M. Smith and M. Mateas. Answer set programming for proceduralcontent generation: A design space approach. TransactionsonComputational Intelligence and AI in Games, 3(3):187–200, 2011. →pages 63, 77[184] A. Solar-Lezama. The sketching approach to program synthesis. InProgramming Languages and Systems, pages 4–13. Springer, 2009. →pages 1[185] A. Solar-Lezama, L. Tancau, R. Bodik, S. Seshia, and V. Saraswat.Combinatorial sketching for finite programs. ACM SIGOPS OperatingSystems Review, 40(5):404–415, 2006. → pages 23[186] N. So¨rensson and A. Biere. Minimizing learned clauses. In Proceedings ofthe International Conference on Theory and Applications of SatisfiabilityTesting, pages 237–243. Springer, 2009. → pages 11[187] P. M. Spira and A. Pan. On finding and updating spanning trees and shortestpaths. SIAM Journal on Computing, 4(3):375–380, 1975. → pages 192[188] G. M. Stalmarck. System for determining propositional logic theorems byapplying values and rules to triplets that are generated from Booleanformula, 1994. US Patent 5,276,897. → pages 7[189] W. Stein et al. Sage: Open source mathematical software. 2008. → pages22[190] O. Strichman, S. A. Seshia, and R. E. Bryant. Deciding separationformulas with SAT. In International Conference on Computer AidedVerification, pages 209–222. Springer, 2002. → pages 1, 178[191] A. Stump, C. W. Barrett, D. L. Dill, and J. R. Levitt. A decision procedurefor an extensional theory of arrays. In Logic in Computer Science,volume 1, pages 29–37. DTIC Document, 2001. → pages 1[192] A. Stump, C. W. Barrett, and D. L. Dill. CVC: A cooperating validitychecker. In International Conference on Computer Aided Verification,pages 500–504. Springer, 2002. → pages 1, 17165Bibliography[193] W. Szeto, Y. Iraqi, and R. Boutaba. A multi-commodity flow basedapproach to virtual network resource allocation. In GlobalTelecommunications Conference and Exhibition (GLOBECOM), 2003. →pages 95[194] R. Tamassia and I. G. Tollis. Dynamic reachability in planar digraphs withone source and one sink. Theoretical Computer Science, 119(2):331–343,1993. → pages 66[195] R. Tarjan. Depth-first search and linear graph algorithms. SIAM Journal onComputing, 1(2):146–160, 1972. → pages 189[196] M. Tomita. Learning of construction of finite automata from examplesusing hill-climbing. 1982. → pages 147[197] G. S. Tseitin. On the complexity of derivation in propositional calculus. InAutomation of Reasoning, pages 466–483. Springer, 1983. → pages 5[198] H. Visser and P. Roling. Optimal airport surface traffic planning usingmixed integer linear programming. In Proceedings of the AIAA’s 3rdAnnual Aviation Technology, Integration, and Operations (ATIO) TechnicalForum, 2003. → pages 69[199] T. Walsh. SAT v CSP. In International Conference on Principles andPractice of Constraint Programming, pages 441–456. Springer, 2000. →pages 7[200] C. Wang, F. Ivancˇic´, M. Ganai, and A. Gupta. Deciding separation logicformulae by SAT and incremental negative cycle elimination. InInternational Conference on Logic for Programming Artificial Intelligenceand Reasoning, pages 322–336. Springer, 2005. → pages 33, 178[201] D. Wang, P. Zhang, C.-K. Cheng, and A. Sen. A performance-driven I/Opin routing algorithm. In Proceedings of the Asia and South Pacific DesignAutomation Conference, 1999. → pages 86[202] R. Wang, R. Shi, and C.-K. Cheng. Layer minimization of escape routingin area array packaging. In Proceedings of the International Conference onComputerAided Design, 2006. → pages 86, 91[203] R. G. Wood and R. A. Rutenbar. FPGA routing and routability estimationvia Boolean satisfiability. IEEE Transactions on Very Large ScaleIntegration (VLSI) Systems, 6(2):222–231, 1998. → pages 62166Bibliography[204] C.-A. Wu, T.-H. Lin, C.-C. Lee, and C.-Y. R. Huang. QuteSAT: a robustcircuit-based SAT solver for complex circuit structure. In Proceedings ofthe conference on Design, automation and test in Europe, pages1313–1318. EDA Consortium, 2007. → pages 6[205] Y. Wu, L. Chiaraviglio, M. Mellia, and F. Neri. Power-aware routing andwavelength assignment in optical networks. Proceedings of teh EuropeanConference on Optical Communication (ECOC), 2009. → pages 69[206] Y. Wu, A. Chen, A. Haeberlen, W. Zhou, and B. T. Loo. Automatednetwork repair with meta provenance. In Proceedings of the 14th ACMWorkshop on Hot Topics in Networks, page 26. ACM, 2015. → pages 146[207] T. Yan and M. D. Wong. A correct network flow model for escape routing.In Proceedings of the Design Automation Conference, 2009. → pages 86,87, 91[208] T. Yan and M. D. Wong. Recent research development in PCB layout. InProceedings of the International Conference on Computer-Aided Design,2010. → pages 86[209] M. Yu, Y. Yi, J. Rexford, and M. Chiang. Rethinking virtual networkembedding: substrate support for path splitting and migration. ACMSIGCOMM Computer Communication Review, 38(2):17–29, 2008. →pages 95, 96[210] M.-F. Yu and W. W.-M. Dai. Pin assignment and routing on a single-layerpin grid array. In Proceedings of the Asia and South Pacific DesignAutomation Conference, 1995. → pages 86[211] M.-f. Yu, J. Darnauer, and W. W.-M. Dai. Planar interchangeable2-terminal routing. Technical Report, UCSC, 1995. → pages 85[212] Y. Yuan, A. Wang, R. Alur, and B. T. Loo. On the feasibility of automationfor bandwidth allocation problems in data centers. In Proceedings of theInternational Conference on Formal Methods in Computer-Aided Design,2013. → pages 69, 94, 99, 100, 101, 102, 103[213] Z. Yuanlin and R. H. Yap. Arc consistency on n-ary monotonic and linearconstraints. In International Conference on Principles and Practice ofConstraint Programming, pages 470–483. Springer, 2000. → pages 23167Bibliography[214] H. Zang, C. Ou, and B. Mukherjee. Path-protection routing and wavelengthassignment (RWA) in WDM mesh networks under duct-layer constraints.IEEE/ACM Transactions on Networking (TON), 11(2):248–258, 2003. →pages 69[215] ZeroStack. Private cloud provider. https://www.zerostack.com/, (accessed26-01-2016). → pages 104[216] H. Zhang and M. E. Stickel. An efficient algorithm for unit propagation. InIn Proceedings of the Fourth International Symposium on ArtificialIntelligence and Mathematics (AI-MATH96), Fort Lauderdale (FloridaUSA. Citeseer, 1996. → pages 8[217] L. Zhang and S. Malik. The quest for efficient Boolean satisfiabilitysolvers. In Proceedings of the International Conference onComputer-Aided Verification, pages 17–36. Springer, 2002. → pages 12[218] L. Zhang, C. F. Madigan, M. H. Moskewicz, and S. Malik. Efficientconflict driven learning in a Boolean satisfiability solver. In Proceedings ofthe 2001 IEEE/ACM international conference on Computer-aided design,pages 279–285. IEEE Press, 2001. → pages 11[219] A. Zook and M. O. Riedl. Automatic game design via mechanicgeneration. In Proceedings of the AAAI Conference on ArtificialIntelligence, pages 530–537, 2014. → pages 63[220] E. Zulkoski, V. Ganesh, and K. Czarnecki. Mathcheck: A math assistantvia a combination of computer algebra systems and SAT solvers. InProceedings of the International Conference on Automated Deduction,pages 607–622. Springer, 2015. → pages 22, 23, 56, 62168Appendix AThe MONOSAT SolverHere we provide a brief, technical overview of our implementation of the SMTsolver MONOSAT, used throughout this thesis for our experiments. While each ofthe theory solvers is described in detail in this thesis (first as an abstract frameworkin Chapter 4, and then in specifics in Chapters 5, 7, and 8), here we discuss theimplementation of the SMT solver itself, and its integration with the theory solvers.Except for the unusual design of the theory solvers, MONOSAT is a typicalexample of a lazy SMT solver, as described, for example, in [179]. The SAT solveris based on MINISAT 2.2 [84], including the integrated SatELite [83] preprocessor.Unlike in MINISAT, some variables are marked as being theory atoms, and areassociated with theory solvers. Theory atoms are prevented from being eliminatedduring preprocessing.As described in Chapter 4, each theory solver implements both a T-Propagateand a T-Analyze method. However, we also implement a few additional methods ineach theory solver:1. T-Enqueue(l), called each time a theory literal l is assigned in the SAT solver.2. T-Backtrack(l), called each time a theory literal l is unassigned in the SATsolver.3. T-Decide(M), called each time the SAT solver makes a decision.Throughout Chapter 4, each algorithm (e.g., Algorithm 3) is described in a169Appendix A. The MONOSAT Solverstateless manner. For example, each time T-Propagate is called, the Update-Bounds section of Algorithm 3 recomputes the over- and under-approximate as-signments from scratch. In practice, such an implementation would be very inef-ficient. In MONOSAT, we make a significant effort to avoid redundant work, inseveral ways. First, as described throughout the thesis, where available, we usedynamic algorithms to actually evaluate predicates during theory propagation andanalysis.Secondly, differing from the description in Chapter 4, our actual implemen-tation of the UpdateBounds method described in Chapter 4 is stateful, ratherthan stateless, and is implemented in the T-Enqueue and T-Backtrack methods,rather than in T-Propagate. In fact, when T-Propagate is called, the under- andover-approximate assignments have already been computed, and T-Propagate onlyneeds to execute the PropagateBounds section. For example, Algorithm 18 de-scribes T-Enqueue and T-Backtrack, as implemented in our graph theory solver.Each time the SAT solver assigns a theory literal (to either polarity), it callsT-Enqueue on the corresponding theory solver; each time it unassigned a theoryliteral (when backtracking), it calls T-Backtrack. Each of our theory solvers imple-ments the UpdateBounds functionality in this manner.A second important implementation detail is the way that justification clausesare produced. In some SMT solvers (such as Z3 [70]), each time a theory solverimplies a theory literal l during theory propagation, the theory solver immediatelygenerates a ‘justification set’ for l. (Recall from Section 2.4 that a justification set isa collection of mutually unsatisfiable atoms.) The justification set is then negatedand turned into a learned clause which would have been sufficient to imply l byunit propagation, under the current assignment. Adding this learned clause to thesolver, even though l has already been implied by the theory solver, allows the SATsolver to derive learned clauses correctly if the assignment of l leads to a conflict.In our implementation, we delay creating justification sets for theory impliedliterals, creating them only lazily when, during conflict analysis, the SAT solverneeds to access the ‘reason’ cause for l. At that time, we create the justificationclause in the theory solver by backtracking the theory solver to just after l wouldbe assigned, and executing T-Analyze in the theory solver to create a justificationset for l. As the vast majority of implied theory literals are never actually directly170Appendix A. The MONOSAT SolverAlgorithm 18 Implementations of T-Enqueue and T-Backtrack in MONOSAT forthe theory of graphs, which replace the UpdateBounds section described in Chap-ter 4, with a stateful implementation, in which E− and E+ are maintained betweencalls to the theory solver (E− is initialized to the empty set, while E+ is initializedto E).function T-Enqueue(l)l is a theory literalif l is an edge literal (e ∈ E) thenE−← E−∪{ei}else if l is a negated edge literal (e /∈ E) thenE+← E+ \{ei}function T-Backtrack(l)l is a theory literalif l is an edge literal (e ∈ E) thenE−← E− \{ei}else if l is a negated edge literal (e /∈ E) thenE+← E+∪{ei}involved in a conflict analysis, justification sets are only rarely actually constructedby MONOSAT. This is important, as in the theories we consider, T-Analyze is oftenvery expensive (for example, for the maximum flow and reachability predicates,T-Analyze performs an expensive minimum cut analysis).Finally, each theory solver is given an opportunity to make decisions in the SATsolver (with a call to the function T-Decide), superseding the SAT solver’s defaultVSIDS heuristic. Currently, only the reach and maximum flow predicates actuallyimplement T-Decide. Those implementations are described Chapter 5. While inmany cases these theory decisions lead to major improvements in performance,we have also found that in some common cases these theory decisions can causepathologically bad behaviour in the solver, and in particular are poorly suited formaze generation tasks (such as the ones described in Section 6.1). In this thesis,theory solver decision heuristics are enabled only in the experiments of Sections6.2, 6.3, for the maximum flow predicate.MONOSAT is written in C++ (with a user-facing API provided in Python 2/3).The various theory solvers include code from several open source libraries, whichare documented in the source code. In the course of our experiments in this thesis,171Appendix A. The MONOSAT Solvervarious, improving versions of MONOSAT were tested. In Section 6.1, MONO-SAT was compiled with G++ 4.8.1; in the remaining experiments, it was compiledwith g++ 6.0. Experiments were conducted using the Python API of MONOSAT,with Python 3.3 (Section 6.1) and Python 3.4 (all other experiments).MONOSAT is freely available under an open-source license, and can be foundat www.cs.ubc.ca/labs/isd/projects/monosat.172Appendix BMonotonic Theory of BitvectorsWe describe a simple monotonic theory of bitvectors. This theory solver is de-signed to be efficiently combinable with the theory of graphs (Chapter 5), by deriv-ing tight bounds on each bitvector variable as assignments are made in the solver,and passing those bounds to other theory solvers in the form of comparison atomson shared bitvector variables (as described in Section 4.5).The theory of bitvectors we introduce here is a theory of fixed-width, non-wrapping bitvectors, supporting only comparison predicates and addition. In con-trast, the standard theory of bitvectors widely considered in the literature (e.g.,[30, 42, 70, 100]), which is essentially a theory of finite modular arithmetic, sup-ports many inherently non-monotonic operators (such as bitwise xor).In the standard theory of bitvectors, the formulax+2 = ywith x,y both 3-bit unsigned bitvectors, both {x = 5,y= 7}, and {x = 7,y= 1} aresatisfying assignments, as 7+ 2 ≡ 1 (mod 8). In contrast, in our non-wrappingtheory of bitvectors, {x = 5,y = 7} is satisfying, while {x = 7,y = 1} is not.Formally, the theory of non-wrapping bitvectors for bit-width n is a subset ofthe theory of integer arithmetic, in which every variable and constant x is con-strained to the range 0 ≤ x < 2n, for fixed n. As all variables are restricted to afinite range of values, this fragment could be easily bit-blasted into propositional173Appendix B. Monotonic Theory of Bitvectorslogic, however we instead will handle it as a lazy SMT theory, using the SMMTframework. While the performance of this theory solver on its own is in fact notcompetitive with bit-blasting bitvector theory solvers, handling this theory lazilyhas the advantage that it allows for comparison atoms to be created on the fly with-out needing to bit-blast comparator circuits for each one. This is important, as thisbitvector theory will be combined with other monotonic theory solvers by derivingand passing large numbers of comparison atoms for each bitvector. In practice, avery large number of such comparison atoms will be created by the solver, and wefound it to be prohibitively expensive to construct bit-blasted comparator circuitsfor each of these comparison atoms, thus motivating the creation of a new theorysolver.In addition to comparison predicates, this theory of bitvectors supports onefunction: addition. Formally, the function add(a,b) 7→ c takes two bitvector ar-guments (a,b) of the same width and outputs a third bitvector (c) of the samebit-width. In our non-wrapping bitvector theory, if a and b are the n-width bitvec-tor arguments of an addition function, it is enforced that a+ b < 2n (by treatinga+b≥ 2n as a conflict in the theory).In Section 4.3, we described two approaches to supporting function composi-tion in monotonic theory solvers:1. Flatten any function composition by introducing fresh variables, treating thefunctions as predicates relating those variables, and then apply Algorithm 3.2. Directly support propagation with function composition, using Algorithm 8.While we take the latter approach in Chapter 8, we take the former approachhere (for historical reasons, as the bitvector theory solver was developed before webegan work on the CTL theory solver).Our implementation of theory propagation for the (non-wrapping) theory ofbitvectors is described in below. For each n-bit bitvector b, we introduce atoms(b0 ∈ b),(b1 ∈ b) . . .(bn−1 ∈ b) to expose the assignment of each bit to the SATsolver. For notational convenience, we represent the negation of (bi ∈ b) as (bi /∈ b).174Appendix B. Monotonic Theory of BitvectorsAlgorithm 19 Theory propagation for the theory of non-wrapping bitvectors.function T-Propagate(M)M is a partial assignment; theory propagate returns a tuple (FALSE, con-flictSet) ifM is found to be UNSAT, and otherwise returns (TRUE,M), possiblyderiving additional atoms to add toM. In practice, we maintain bounds b−,b+from previous calls, only updating them if an atom involving b has been addedto or removed fromM.assignmentChanged← FALSEfor each n-bit bitvector variable b dob−← 0,b+← 0for i in 0..n−1 doif (bi /∈ b) /∈M thenb+← b++2iif (bi ∈ b) ∈M thenb−← b−+2ifor each comparison to constant (b≤ σi) doif (b≤ σi) ∈M thenb+← min(b+,σi)else if (b > σi) ∈M thenb−← max(b−,σi+1)elseif b+ ≤ σi thenassignmentChanged← TRUEM←M∪{(b≤ σi)}else if b− > σi thenassignmentChanged← TRUEM←M∪{(b > σi)}for each comparison atom between bitvectors (a≤ b) doif (a≤ b) ∈M thena+← min(a+,b+)b−← max(b−,a−)else if (a > b) ∈M then175Appendix B. Monotonic Theory of Bitvectorsa−← max(a−,b−+1)b+← min(b+,a++1)elseif a+ ≤ b− thenassignmentChanged← TRUEM←M∪{(a≤ b)}else if a− > b+ thenassignmentChanged← TRUEM←M∪{(b > a)}for each addition function add(a+b) 7→ c doc+← min(c+,a++b+)c−← max(c−,a++b+)a+← min(a+,c+−b−)a−← max(a−,c−−b+)b+← min(b+,c+−a−)b−← max(b−,c−−a+)for each bitvector variable b of width n dob−← REFINE LBOUND(b−,n,M)b+← REFINE UBOUND(b+,n,M)if b− > b+ thenreturn FALSE,T-Analyze(M)elseif (b≤ b+) /∈M thenassignmentChanged← TRUEM←M∪{(b≤ b+)}if (b≥ b−) /∈M thenassignmentChanged← TRUEM←M∪{(b≥ b−)}if assignmentChanged thenreturn T-Propagate(M)elsereturn TRUE,M176Appendix B. Monotonic Theory of BitvectorsWe describe our theory propagation implementation in Algorithm 19. Thereare two improvements that we make in this implementation, as compared to Algo-rithm 3. The first is to add calls to REFINE LBOUND and REFINE UBOUND; thesefunctions attempt to derive tighter bounds on a bitvector by excluding assignmentsthat are incompatible with the bit-level assignment of the variable. For example, ifwe have a bound b≥ 2, but we also have (b1 /∈ b) ∈M, then we know that b can-not equal exactly 2, and so we can refine the bound on b to b≥ 3. These functionsperform a search attempting to tighten the lower bound using reasoning as above,requiring linear time in the bit-width n. We omit the pseudo-code from this thesis,as the implementation is a bit involved; it can be found in “BvTheorySolver.h” inthe (open-source) implementation of MONOSAT. We also omit the description ofthe implementation of T-Analyze, which is a straightforward adaptation of Algo-rithm 6.The second change we make relative to Algorithm 3 is, in the case whereany literals have been added to M, to call THEORYPROPAGATE again, with thenewly refined M as input. This causes THEORYPROPAGATE to repeat until theM reaches a fixed point (termination is guaranteed, as literals are never removedfromM during THEORYPROPAGATE, and there is a finite set of literals that canbe added to M). Calling THEORYPROPAGATE when the M changes allows re-finements to the bounds on one bitvector to propagate through addition functionsor comparison predicates to refine the bounds on other bitvectors, regardless of theorder in which we examine the bitvectors and predicates in a given call to THEO-RYPROPAGATE.The non-wrapping bitvector theory solver we describe here is sufficient for ourpurposes in this thesis, supporting the weighted graph applications described inSection 6.3, which makes fairly limited use of the bitvector theory solver. How-ever, the performance of this solver is not competitive with existing bitvector orarithmetic theory solvers (in particular, the repeated calls to establish a fixed pointcan be quite expensive).An alternative approach, which we have not yet explored, would be to imple-ment an efficient difference logic theory solver. Difference logic is sufficientlyexpressive for some of our applications, and should be combinable with our mono-tonic theories in the same manner as the theory of bitvectors, as difference logic177Appendix B. Monotonic Theory of Bitvectorssupports comparisons to constants. A fast difference logic solver would almostcertainly outperform the above bitvector theory (considered in isolation from othermonotonic theory solvers). However, while existing difference logic theory solvers(e.g., as described in [10, 144, 190, 200]) are designed to quickly determine thesatisfiability of a set of difference logic constraints; they are not (to the best ofour knowledge) designed to efficiently derive not just satisfying, but tight, upperand lower bounds on each variable, from partial assignments during theory prop-agation. In contrast, the bitvector theory solver we described in this section isdesigned to find and maintain tight upper and lower bounds on each variable, frompartial assignments. This is critically important, as passing tight bounds to othermonotonic theory solvers (such as the weighted graph theory of Chapter 5) canallow those other theory solvers to prune unsatisfying assignments early.178Appendix CSupporting MaterialsC.1 Proof of Lemma 4.1.1This lemma can be proven for the case of totally ordered σ , by considering thefollowing exhaustive, mutually exclusive cases:Proof. 1. T |= ¬MA. In this case, MA is by itself already an inconsistentassignment to the comparison atoms. This can happen if some variable x isassigned to be both larger and smaller than some constant, or is assigned tobe larger than σ> or smaller than σ⊥. In this case, as MA (and hence anysuperset ofMA) is theory unsatisfiable,MA =⇒Tp holds trivially for any p,and hence both the left and right hand sides of 4.1 and 4.2 hold.2. MA T , T ∪MA |= ¬(x ≥ σ j). In this case, MA ∪{x ≥ σi} =⇒Tp holdstrivially for any p, and so 4.1 holds.3. MA T , T ∪MA |= ¬(x ≥ σi). In this case, MA ∪ {x ≥ σi} =⇒Tp holdstrivially for any p. AsMA is by itself theory satisfiable, andMA only con-tains assignments of variables to constants, there must exist some σk < σi,such that atom (x ≤ σk) ∈ MA. As σ j ≥ σi, we also have σ j > σk, andT ∪MA |=¬(x≥ σ j). Therefore,MA∪{x≥ σi}=⇒Tp holds trivially for anyp, and so 4.1 holds.179Appendix C. Supporting Materials4. MA ∪{x ≥ σi,x ≥ σ j} T . In this case, if MA ∪{x ≥ σi} =⇒Tp, then byDefinition 1,MA∪{x≥ σ j}=⇒Tp, as p is positive monotonic. Otherwise, ifMA∪{x≥ σi}=⇒T¬p, then the antecedent of 4.1 is false, and so 4.1 holds.The cases for T ∪MA |=¬(x≤ σ j),T ∪MA |=¬(x≤ σi),MA∪{x≤ σi,x≤σ j} T are symmetric, but for 4.2.180Appendix C. Supporting MaterialsC.2 Proof of Correctness for Algorithm 7In Section 4.3, we introduced algorithm APPROX, for finding safe upper and lowerapproximations to compositions of positive and negative monotonic functions. Werepeat it here for reference:Algorithm 7 Approximate evaluation of composed monotonic functionsfunction APPROX(φ ,M+A ,M−A )φ is a formula,M+A ,M+A are assignments.if φ is a variable or constant term thenreturnM+A [φ ]else φ is a function term or predicate atom f (t0, t1, . . . , tn)for 0≤ i≤ n doif f is positive monotonic in ti thenxi = APPROX(ti,M+A ,M−A )else argumentsM+A ,M−A are swappedxi = APPROX(ti,M−A ,M+A )return evaluate( f (x0,x1,x2, . . . ,xn))In the same section, we claimed that the following lemma holds:Lemma C.2.1. Given any term φ composed of mixed positive and negative mono-tonic functions (or predicates) over variables vars(φ), and given any three com-plete, T -satisfying models for φ ,M+A ,M∗A, andM−A :∀x ∈ vars(φ),M+A [x]≥M∗A[x]≥M−A [x] =⇒ APPROX(φ ,M+A ,M−A )≥M∗A[φ ]—and—∀x ∈ vars(φ),M+A [x]≤M∗A[x]≤M−A [x] =⇒ APPROX(φ ,M+A ,M−A )≤M∗A[φ ]The proof is given by structural induction over APPROX:Proof.181Appendix C. Supporting MaterialsInductive hypothesis:∀x ∈ vars(φ),M+A [x]≥M∗A[x]≥M−A [x] =⇒ APPROX(φ ,M+A ,M−A )≥M∗A[φ ]—and—∀x ∈ vars(φ),M+A [x]≤M∗A[x]≤M−A [x] =⇒ APPROX(φ ,M+A ,M−A )≤M∗A[φ ]Base case (φ is x, a variable or constant term):APPROX returnsM+A [x]; therefore,M+A [x]≥M∗A[x] implies APPROX(x,M+A ,M−A )≥M∗A[x].M+A [x]≤M∗A[x] implies APPROX(x,M+A ,M−A )≤M∗A[x].Inductive step (φ is a function or predicate term f (t0, t1, . . . tn)):APPROX returns evaluate( f (x0,x1, . . . ,xn)). For each xi, there are two cases:1. f is positive monotonic in argument i, xi← APPROX(ti,M+A ,M−A )In this case we have (by the inductive hypothesis):∀x ∈ vars(φ),M+A [x] ≥ M∗A[x] ≥ M−A [x] implies xi =APPROX(ti,M+A ,M−A )≥M∗A[ti], and we also have∀x ∈ vars(φ),M+A [x] ≤ M∗A[x] ≤ M−A [x] implies xi =APPROX(ti,M+A ,M−A )≤M∗A[ti].2. f is negative monotonic in argument i, xi← APPROX(ti,M−A ,M+A )In this case we have (by the inductive hypothesis):∀x ∈ vars(φ),M+A [x] ≥ M∗A[x] ≥ M−A [x] implies xi =APPROX(ti,M−A ,M+A )≤M∗A[ti], and we also have∀x ∈ vars(φ),M+A [x] ≤ M∗A[x] ≤ M−A [x] implies xi =APPROX(ti,M−A ,M+A )≥M∗A[ti].In other words, in the case thatM+A [x]≥M−A [x], then for each positive mono-tonic argument i, we have xi ≥M∗A[ti], and for each negative monotonic argumenti, we have xi ≤M∗A[ti]. Additionally, in the case that M+A [x] ≤M−A [x], we havethe opposite: then for each positive monotonic argument i, we have xi ≤M∗A[ti],and for each negative monotonic argument i, we have xi ≥M∗A[ti].182Appendix C. Supporting MaterialsTherefore, we also have:∀x ∈ vars(φ),M+A [x] ≥ M∗A[x] ≥ M−A [x] implies f (x0,x1, . . . ,xn) ≥f (M∗A[t0],M∗A[t1], . . . ,M∗A[tn]) implies APPROX( f (t0, t1, . . . , tn),M+A ,M−A ) ≥M∗A[ f (t0, t1, . . . , tn)],and we also have∀x ∈ vars(φ),M+A [x] ≤ M∗A[x] ≤ M−A [x] implies f (x0,x1, . . . ,xn) ≤f (M∗A[t0],M∗A[t1], . . . ,M∗A[tn]) implies APPROX( f (t0, t1, . . . , tn),M+A ,M−A ) ≤M∗A[ f (t0, t1, . . . , tn)].This completes the proof.183Appendix C. Supporting MaterialsC.3 Encoding of Art Gallery SynthesisIn Table 7.1, we described a set of comparisons between MONOSAT and the SMTsolver Z3 on Art Gallery Synthesis instances. While the translation of this probleminto the geometry theory supported by MONOSAT is straightforward, the encodinginto Z3’s theory of linear real arithmetic is more involved, and we describe it here.Each synthesis instance consists of a set of points P, at pre-determined 2Dcoordinates, within a bounded rectangle. From these points, we must find N non-overlapping convex polygons, with vertices selected from those points, such thatthe area of each polygon is greater than some specified constant, and all vertices ofall the polygons (as well as the 4 corners of the bounding rectangle) can be ‘seen’by a fixed set of at most M cameras. A camera is modeled as a point, and is definedas being able to see a vertex if a line can reach from that camera to the vertexwithout intersecting any of the convex polygons. Additionally, to avoid degeneratesolutions, we enforce that the convex polygons cannot meet at only a single point:they must either share entire edges, or must not meet at all.Encoding this synthesis task into linear real arithmetic (with propositional con-straints) requires modeling notions of point containment, line intersection, area,and convex hulls, all in 2 dimensions. The primary intuition behind the encodingwe chose is to ensure that at each step, all operations can be applied in arbitrary pre-cision rational arithmetic. In particular, we will rely on computing cross productsto test line intersection, using the following function:crossDi f (O,A,B) = (A[0]−O[0])∗ (B[1]−O[1])− (A[1]−O[1])∗ (B[0]−O[0])This function takes non-symbolic, arbitrary precision arithmetic 2D points asarguments, and return non-symbolic arbitrary precision results.Our encoding can be broken down into two parts. First, we create a setof N symbolic polygons. Each symbolic polygon consists of a set of Booleansenabled(p), controlling for each point p ∈ P whether p is enabled in that sym-bolic polygon. Additionally, for each pair of points (p1, p2), the symbolic polygondefines a formula onHull(p1, p2) which evaluates to TRUE iff (p1, p2) is one ofthe edges that make up the convex hull of the enabled points. Given points p1, p2,onHull(p1, p2) evaluates to true iff:184Appendix C. Supporting Materialsenabled(p1)∧enabled(p2)∧∀p3∈P\{p1,p2}(crossDi f (p2, p1, p3)≥ 0 =⇒ enabled(p3))As there are a bounded number of points, the universal quantifier is unrolledinto a conjunction of |P| − 2 individual conditionals; as onHull is computed foreach ordered pair p1, p2, there are a cubic number of total constraints required tocompute onHull for each polygon.Given two symbolic polygons S1,S2 with onHullS1, onHullS2 computed asabove, our approach to enforcing that they do not overlap is to assert that thereexists an edge normal to one of the edges on the hull of one of the two polygons,such that each enabled point of S1 and S2 do not overlap when projected onto thatnormal. The normal edge is selected non-deterministically by the solver (using |P|2constraints), and the projections are computed and compared using dot products,using |P|3 constraints.The separation constraint described above requires |P|3 constraints for a givenpair of symbolic polygons S1,S2, and is applied pairwise between each unique pairof symbolic polygons.The next step of the encoding is to ensure that it is possible to place M cam-eras in the room while covering all of the vertices. Before describing how this iscomputed, we first introduce three formulas for testing point and line intersection.containsPoint(S, p) takes a symbolic polygon S and a point p, and returns true if pis contained within the convex hull of S.We define containsPoint(S, p) as follows:containsPoint(S, p) =∧∀(e1,e2)∈edges(S)(¬onHullS(e1,e2)∨ crossDi f (e2,e1, p)< 0)lineIntersectsPolygon(S,(p1, p2)) takes a symbolic polygon S, and a line de-fined as a tuple of points (p1, p2), and evaluates TRUE if the line (p1, p2) intersectsone of the edges of the hull of S (not including vertices). If the line does not in-tersect S, or only intersects a vertex of S, then lineIntersectsPolygon evaluates toFALSE.185Appendix C. Supporting MaterialsWe define lineIntersectsPolygon(S,(p1, p2)) as follows:lineIntersectsPolygon(S,(p1, p2)) = containsPoint(S, p1)∨ containsPoint(S, p2)∨∀(e1,e2)∈edges(S)(onHullS(e1,e2)∧ lineIntersectsLine((p1, p2),(e1,e2)))Finally, lineIntersectsLine((p1, p2),(e1,e2)) returns true if the two (non sym-bolic) lines defined by endpoints p1, p2 and e1,e2 intersect each other. This can betested for non-symbolic rational constants using cross products.Given a fixed camera position pc, and a fixed vertex to observe pv,isVisible(pc, pv) is evaluated as follows:isVisible(pc, pv) =∧∀S∈Polygons¬lineIntersectsPolygon(S,(pc, pv))We then use a cardinality constraint to force the solver to pick C positions fromthe point set P to be cameras, and use the line intersection formula described aboveto force that every vertex that is on hull of a symbolic polygon must be visible fromat least one vertex that is a camera (meaning that no symbolic polygon intersectsthe line between the camera and that vertex).Each instance also specifies a minimum area A for each symbolic polygon (toensure we do not get degenerate solutions consisting of very thin polygon seg-ments). We compute the area of the convex hull of each symbolic polygon S as:area(S) =12 ∑∀(e1,e2)∈edges(S)I f (onHullS(e1,e2),e1[0]∗ e2[1]− e1[1]∗ e2[0])To complete the encoding, we assert:∀S,area(S)≥ A186Appendix DMonotonic PredicatesThis is a listing of the monotonic predicates considered in this thesis, arranged bytopic.D.1 Graph PredicatesD.1.1 ReachabilityMonotonic Predicate: reachs,t(E) , true iff t can be reached from u in the graphformed of the edges enabled in E.Implementation of evaluate(reachs,t(E),G): We use the dynamic graph reacha-bility/shortest paths algorithm of Ramalingam and Reps [171] to test whethert can be reached from s in G. Ramalingam-Reps is a dynamic variant of Dijk-stra’s Algorithm [79]); our implementation follows the one described in [48].If there are multiple predicate atoms reachs,t(E) sharing the same source s,then Ramalingam-Reps only needs to be updated once for the whole set ofatoms.Implementation of analyze(reachs,tE,G−) Node s reaches t in G−, butreachs,t(E) is assigned FALSE in M. Let e0,e1, . . . be a u− v path in G−;return the conflict set {(e0 ∈ E),(e1 ∈ E), . . . ,¬reachs,t(E)}.Implementation of analyze(¬reachs,t(E),G+) Node t cannot be reached from s187Appendix D. Monotonic Predicatesin G+, but reachs,t(E) is assigned TRUE. Let e0,e1, . . . be a cut of disablededges (ei /∈ E) separating u from v in G+. Return the conflict set {(e0 /∈E),(e1 /∈ E), . . . ,reachs,t(E)}.We find a minimum separating cut by creating a graph containing all edgesof E (including both the edges of G+ and the edges that are disabled in G+in the current assignment), in which the capacity of each disabled edge ofE is 1, and the capacity of all other edges is infinity (forcing the minimumcut to include only edges that correspond to disabled edge atoms). Anystandard maximum s-t flow algorithm can then be used to find a minimumcut separating u from v. In our implementation, we use the dynamic Kohli-Torr [140] minimum cut algorithm for this purpose.Decision Heuristic: (Optional) If reachs,t(E) is assigned TRUE in M, but theredoes not yet exist a u− v path in G−, then find a u− v path in G+ and pickthe first unassigned edge in that path to be assigned true as the next decision.In practice, such a path has typically already been discovered, during theevaluation of reachs,t on G+ during theory propagation.D.1.2 AcyclicityMonotonic Predicate: acyclic(E), true iff there are no (directed) cycles in thegraph formed by the enabled edges in E.Implementation of evaluate(acyclic(E),G): Apply the PK dynamic topologicalsort algorithm (as described in [164]). The PK algorithm is a fully dynamicgraph algorithm that maintains a topological sort of a directed graph as edgesare added to or removed from the graph; as a side effect, it also detectsdirected cycles (in which case no topological sort exists). Return FALSE ifthe PK algorithm successfully produces a topological sort, and return TRUEif it fails (indicating the presence of a directed cycle).Implementation of analyze(acyclic(E),G−) There is a cycle in E, but acyclic(E)is assigned TRUE. Let e0,e1, . . . be the edges that make up a directed cyclein E; return the conflict set {(e0 ∈ E),(e1 ∈ E), . . . ,acyclic(E)}.188Appendix D. Monotonic PredicatesImplementation of analyze(¬acyclic(E),G+) There is no cycle in E, butacyclic(E) is assigned FALSE. Let e0,e1, . . . be the set of all edges not inE+; return the conflict set {(e0 /∈ E),(e1 /∈ E), . . . ,¬acyclic(E)}. (Note thatthis is the default monotonic conflict set.)D.1.3 Connected ComponentsHere we consider a predicate which constraints the number of (simple) connectedcomponents in the graph to be less than or greater than some constant. There areboth directed and undirected variations of this predicate; we describe the directedversion, which counts the number of strongly connected components.Monotonic Predicate: connectedComponentsE,m, true iff the number of con-nected components in the graph formed of edges E is≤m, with m a bitvector.Implementation of evaluate(connectedComponents(E,m),G): Use Tarjan’sSCC algorithm [195] to count the number of distint strongly connectedcomponents in G; return TRUE iff the count is ≤ m.Implementation of analyze(connectedComponents(E,m),G−): The connectedcomponent count is less or equal to m, but connectedComponents(E,m)is FALSE in M. Construct a spanning tree for each component ofG− (using depth-first search). Let edges e1,e2, . . . be the edges inthese spanning trees; return the conflict set {(e1 ∈ E),(e2 ∈ E), . . . ,(m >m−),¬connectedComponents(E,m)}.Implementation of analyze(¬connectedComponents(E,m),G+): The con-nected component count is greater than m, but connectedComponents(E,m)is TRUE in M Collect all disabled edges ei = (u,v) where uand v belong to different components in G+; the conflict set is{(e1 /∈ E),(e2 /∈ E), . . . ,(m < m+),connectedComponents(E,m)}.Above we describe the implementation for directed graphs; for undirectedgraphs, Tarjan’s SCC algorithm does not apply. For undirected graphs, we countthe number of connected components using disjoint-sets/union-find.189Appendix D. Monotonic PredicatesD.1.4 Shortest PathMonotonic Predicate: shortestPaths,t(E, l,c0,c1, . . .), true iff the length of theshortest s− t path in the graph formed of the edges enabled in E with edgeweights c0,c1, . . . is less than or equal to bit-vector l.Implementation of evaluate(shortestPaths,t(E, l,c0,c1, . . .),G): We use the dy-namic graph reachability/shortest paths algorithm of Ramalingam andReps [171] to test whether the shortest path to t from s is less than l. If thereare multiple predicate atoms shortestPaths,t(E) sharing the same source s,then Ramalingam-Reps only needs to be updated once for the whole set ofatoms.Implementation of analyze(shortestPaths,t ,E, l,c0,c1, . . .): There exists an s-tpath in G− of length ≤ l, but shortestPaths,t is assigned FALSE inM. Let e0,e1, . . . be a shortest s − t path in G− with weightsw0,w1, . . .; return the conflict set {(e0 ∈ E),(w0 ≥ w−0 ),(e1 ∈ E),(w1 ≥w−1 ), . . . ,¬shortestPaths,t}.Implementation of analyze(¬shortestPaths,t ,E, l,c0,c1, . . .): Walk back from uin G+ along enabled and unassigned edges with edge weights w0,w1, . . .;collect all incident disabled edges e0,e1, . . .; return the conflict set {(e0 /∈E),(e1 /∈ E), . . . ,(w0 ≤ w+0 ),(w1 ≤ w+1 ) . . . ,shortestPathu,v,G≤C(edges)}.D.1.5 Maximum FlowMonotonic Predicate: maxFlows,t(E,m,c0,c1, . . .), true iff the maximum s-t flowin G with edge capacities c0,c1, . . . is ≥ m.Implementation of evaluate(maxFlows,t ,G,m,c0,c1, . . .): We apply the dynamicminimum-cut/maximum s-t algorithm by Kohli and Torr [140] to computethe maximum flow of G, with edge capacities set by ci. Return TRUE iff thatflow is greater or equal to m.Implementation of analyze(maxFlows,t ,G−,m+,c−0 ,c−1 , . . .): The maximum s-tflow in G− is f , with f ≥ m+). In the computed flow, each edge ei is either190Appendix D. Monotonic Predicatesdisabled in G−, or it has been allocated a (possibly zero-valued) flow fi, withfi ≤ c−i . Let ea,eb, . . . be the edges enabled in G− with non-zero allocatedflows fa, fb, . . .. Either one of those edges must be disabled in the graph, orone of the capacities of those edges must be decreased, or the flow will beat least f . Return the conflict set {(ea ∈ E),(eb ∈ E), . . . ,(ca ≥ fa),(cb ≥fb) . . . ,(m≤ f ),maxFlows,t(E,m,c0,c1, . . .)}.Implementation of analyze(¬maxFlows,t ,G+,m−,c+0 ,c+1 , . . .): The maximum s-t flow in G+ is f , with f < m−. In the flow, each edge that is enabled in G+has been allocated a (possibly zero-valued) flow fi, with fi ≤ c+i .Create a graph Gcut . For each edge ei = (u,v) in G+, with fi < c+i , add aforward edge (u,v) to Gcut with infinite capacity, and also a backward edge(v,u) with capacity fi. For each edge ei = (u,v) in G+ with fi = c+i , add aforward edge (u,v) to Gcut with capacity 1, and also a backward edge (v,u)with capacity fi. For each edge ei = (u,v) that is disabled in G+, add onlythe forward edge (u,v) to Gcut , with capacity 1.Compute the minimum s-t cut of Gcut . Some of the edges along this cut mayhave been edges disabled in G+, while some may have been edges enabledin G+ with fully utilized edge capacity. Let ea,eb, . . . be the edges of theminimum cut of Gcut that were disabled in G+. Let cc,cd , . . . be the capacitiesof edges in the minimum cut for which the edge was included in G+, withfully utilized capacity. Return the conflict set {(ea /∈ E),(eb /∈ E), . . . ,(cc ≤fc),(cd ≤ fd), . . . ,(m > f ),¬maxFlows,t(E,m,c0,c1, . . .)}.In practice, we maintain a graph Gcut for each maximum flow predicate atom,updating its edges only lazily when needed for conflict analysis.Decision Heuristic: (Optional) If maxFlows,t is assigned TRUE inM, but theredoes not yet exist a sufficient flow in G−, then find a maximum flow in G+,and pick the first unassigned edge with non-zero flow to be assigned TRUEas the next decision. If no such edge exists, then pick the first unassignededge capacity and assign its capacity to its flow G+, as the next decision. Inpractice, such a flow is typically already discovered, during the evaluation ofmaxFlows,t on G+ during theory propagation.191Appendix D. Monotonic PredicatesD.1.6 Minimum Spanning Tree WeightHere we consider constraints on the weight of the minimum spanning tree in aweighted (non-negative) undirected graph. For the purposes of this solver, we willdefine unconnected graphs to have infinite weight.Monotonic Predicate: minSpanningTree(E,m,c0,c1, . . .), evaluates to TRUE iffthe minimum weight spanning tree in the graph formed of edges E withedge weights given by c0,c1, . . ., has weight ≤ m.Implementation of evaluate(minSpanningTree(E,m,c0,c1, . . .),G: We use Spiraand Pan’s [187] fully dynamic algorithm for minimum spanning trees to de-temrine the minimum weight spanning tree of G; return TRUE iff that weightis ≤ m.Implementation of analyze(minSpanningTree,G−,m−,c+0 ,c+1 , . . .): Theminimum weight spanning tree of G− is less or equal to m−, butminSpanningTree is FALSE in M. Let e0,e1, . . ., with edge weightsw0,w1, . . ., be the edges of that spanning tree. Return the conflict set{(e0 ∈ E),(w0≤w+0 ),(e1 ∈ E),(w1≤w+1 ),(m≥m−),¬minSpanningTree}.Implementation of analyze(¬minSpanningTree,G+,m−,c−0 ,c−1 , . . .): The mini-mum weight spanning tree of G+ is > m+, but minSpanningTree is TRUE inM. There are two cases to consider:1. If G+ is disconnected, then we consider its weight to be infinite. Inthis case, we find a cut {e1,e2, . . .} of disabled edges separating anyone component from the remaining components. Given a componentin G+, a valid separating cut consists of all disabled edges (u,v) suchthat u is in the component and v is not. We can either return thefirst such cut we find, or the smallest one from among all the com-ponents. For disconnected G+, return the conflict set {(e1 /∈ E),(e2 /∈E), . . . ,minSpanningTree(E,m)}.2. If G+ is not disconnected, then we search for a minimal set of edgesrequired to decrease the weight of the minimum spanning tree. Todo so, we visit each disabled edge (u,v), and then examine the cycle192Appendix D. Monotonic Predicatesthat would be created if we were to add (u,v) into the minimum span-ning tree. (Because the minimum spanning tree reaches all nodes, andreaches them exactly once, adding a new edge between any two nodeswill create a unique cycle in the tree.) If any edge in that cycle hashigher weight than the disabled edge, then if that disabled edge wereto be enabled in the graph, we would be able to create a smaller span-ning tree by replacing that larger edge in the cycle with the disablededge. Let e0,e1, . . . be the set of such disabled edges that can be usedto create lower weight spanning trees. Additionally, let w0,w1, . . . bethe weights of the edges in the minimum spanning tree of G+. Returnthe conflict set is {(e0 /∈ E),(e1 /∈ E), . . . ,(w0 ≥ w−0 ),(w1 ≥ w−1 ),(m≤m+),minSpanningTree(E,m)}.In practice, we can visit each such cycle in linear time by using Tar-jan’s off-line lowest common ancestors algorithm [99] in the minimumspanning tree found in G+. Visiting each edge in the cycle to check ifit is larger than each disabled edge takes linear time in the number ofedges in the tree for each cycle. Since the number of edges in the tree is1 less than the number of nodes, the total runtime is then O(|V |2 · |D|),where D is the set of disabled edges in M.D.2 Geometric PredicatesD.2.1 Area of Convex HullsMonotonic Predicate: hullArea≥x(S), hullArea>x(S), with S a finite set S ⊆ S′,true iff the convex hull of the points in S has an area ≥ (resp. >) than x(where x is a constant). This predicate is positive monotonic with respect tothe set S.Algorithm: Initially, compute area(bound−),area(bound+). If area(bound−) >x, compute area(H−); if area(bound+) < x, compute area(H+). The areascan be computed explicitly, using arbitrary precision rational arithmetic.Conflict set for hullArea≥x(S): The area of H− is greater or equal to x. Let193Appendix D. Monotonic Predicatesp0, p1, . . . be the points of S− that form the vertices of the under-approximatehull H−, with area(H−) ≥ x. Then at least one of the points must bedisabled for the area of the hull to decrease below x. The conflict set is{(p0 ∈ S),(p1 ∈ S), . . . ,¬hullArea≥x(S)}.Conflict set for ¬hullArea≥x(S): The area of H+ is less than x; then at least onepoint pi /∈ S+ that is not contained in H+ must be added to the point set toincrease the area of H+. Let points p0, p1, . . . be the points of S′ not containedin H+. The conflict set is {(p0 /∈ S+),(p1 /∈ S+), . . . ,hullArea≥x(S)}, wherep0, p1, . . . are the (possibly empty) set of points pi /∈ S+.D.2.2 Point Containment for Convex HullsMonotonic Predicate: hullContainsq(S), true iff the convex hull of the 2D pointsin S contains the (fixed point) q.Algorithm: First, check whether q is contained in the under-approximative bound-ing box bound−. If it is, then check if q is contained in H−. We use thePNPOLY [95] point inclusion test to perform this check, using arbitraryprecision rational arithmetic, which takes time linear in the number of ver-tices of H−. In the same way, only if q is contained in bound+, check if q iscontained in H+ using PNPOLY.Conflict set for hullContainsq(S): Convex hull H− contains p. Let p0, p1, p2 bethree points from H− that form a triangle containing p (such a triangle mustexist, as H− contains p, and we can triangulate H−, so one of those trian-gles must contain p; this follows from Carathodory’s theorem for convexhulls [51]). So long as those three points are enabled in S, H− must containthem, and as they contain p, p must be contained in hunder. The conflict setis {(p0 ∈ S),(p1 ∈ S),(p2 ∈ S),¬hullContainsq(S)}.Conflict set for ¬hullContainsq(S): p is outside of H+. If H+ is empty, then theconflict set is the default monotonic conflict set. Otherwise, by the sepa-rating axis theorem, there exists a separating axis between H+ and p. Letp0, p1, . . . be the (disabled) points of S whose projection onto that axis is194Appendix D. Monotonic Predicates≥ the projection of p onto that axis1. At least one of those points mustbe enabled in S in order for H+ to grow to contain p. The conflict set is{(p0 /∈ S),(p1 /∈ S), . . . ,hullContainsq(S)}.D.2.3 Line-Segment Intersection for Convex HullsMonotonic Predicate: hullIntersectsr(S), true iff the convex hull of the 2D pointsin S intersects the (fixed) line-segment r.Algorithm: First, check whether line-segment r intersects bound−. If it does,check if r intersects H−. If H− is empty, is a point, or is itself a line-segment,this is trivial (and can be checked precisely in arbitrary precision arithmeticby computing cross products following [110]). Otherwise, we check if ei-ther end-point of r is contained in H−, using PNPOLY as above for pointcontainment. If neither end-point is contained, we check whether the line-segment intersects H−, by testing each edge of H− for intersection with r(as before, by computing cross-products). If r does not intersect the under-approximation, repeat the above on bound+ and H+.Conflict set for hullIntersectsr(S): Convex hull H− intersects line-segment r. Ifeither end-point or r was contained in H−, then proceed as for the point con-tainment predicate. Otherwise, the line-segment r intersects with at least oneedge (pi, p j) of H−. Conflict set is {(pi ∈ S),(p j ∈ S),¬hullIntersectsr(S)}.Conflict set for ¬hullIntersectsr(S): r is outside of H+. If H+ is empty, then theconflict set is the naı¨ve monotonic conflict set. Otherwise, by the separatingaxis theorem, there exists a separating axis between H+ and line-segment r.Let p0, p1, . . . be the (disabled) points of S whose projection onto that axis is≥ the projection of the nearest endpoint of r onto that axis. At least one ofthose points must be enabled in S in order for H+ to grow to contain p. Theconflict set is {(p0 /∈ S),(p1 /∈ S), . . . ,hullIntersectsr(S)}.1We will make use of the separating axis theorem several times in this section. The standard pre-sentation of the separating axis theorem involves normalizing the separating axis (which we wouldnot be able to do using rational arithmetic). This normalization is required if one wishes to computethe minimum distance between the projected point sets; however, if we are only interested in com-paring distances (and testing collisions), we can skip the normalization step, allowing the separatingaxis to be found and applied using only precise rational arithmetic.195Appendix D. Monotonic PredicatesD.2.4 Intersection of Convex HullsMonotonic Predicate: hullsIntersect(S0,S1), true iff the convex hull of the pointsin S0 intersects the convex hull of the points of S1.Algorithm: If the bounding box for convex hull H0 intersects the bounding boxfor H1, then there are two possible cases to check for:1. 1: A vertex of one hull is contained in the other, or2. 2: an edge of H0 intersects an edge of H1.If neither of the above cases holds, then H0 and H1 do not intersect.Each of the above cases can be tested in quadratic time (in the number ofvertices of H0 and H1), using arbitrary precision arithmetic. In our imple-mentation, we use PNPOLY, as described above, to test vertex containment,and test for pair-wise edge intersection using cross products.Conflict set for hullsIntersect(S0,S1): H−0 intersects H−1 . There are two (not mu-tually exclusive) cases to consider:1. A vertex of one hull is contained within the other hull. Let the point bepa; let the hull containing it be H−b . Then (as argued above) there mustexist three vertices pb1, pb2, pb3 of H−b that form a triangle containingpa. So long as those three points and pa are enabled, the two hullswill overlap. Conflict set is {(pa ∈ Sa),(pb1 ∈ Sb),(pb2 ∈ Sb),(pb3 ∈Sb),¬hullsIntersect(S0,S1)}.2. An edge of H−0 intersects an edge of H−1 . Let p1a, p1b be points of H−0 ,and p2a, p2b points of H−1 , such that line segments p1a, p1b and p2a, p2bintersect. So long as these points are enabled, the hulls of the twopoint sets must overlap. Conflict set is {(p0a ∈ S0),(p0b ∈ S0),(p1a ∈S1),(p1b ∈ S1),¬hullsIntersect(S0,S1)}.Conflict set for ¬hullsIntersect(S0,S1): H+0 do not intersect H+1 . In this case,there must exist a separating axis between H+0 and H+1 . (Such an axis can bediscovered as a side effect of computing the cross products of each edge instep 2 above.) Project all disabled points of S0 and S1 onto that axis. Assume196Appendix D. Monotonic Predicates(without loss of generality) that the maximum projected point of H+1 is lessthan the minimum projected point of H+1 . Let p0a, p0b, . . . be the disabledpoints of S0 whose projections are on the far side of the maximum projectedpoint of H+1 . Let p1a, p1b, . . . be the disabled points of S1 whose projectionsare near side of the minimum projected point of H+1 . At least one of thesedisabled points must be enabled, or this axis will continue to separate thetwo hulls. The conflict set is {(p0a /∈ S0),(p0b /∈ S0), . . . ,(p1a /∈ S1),(p1b /∈S1), . . . ,hullsIntersect(S0,S1)}.D.3 Pseudo-Boolean ConstraintsMonotonic Predicate: p = ∑n−1i=0 cibi ≥ cn, with each ci a constant, and bi aBoolean argument.Implementation of evaluate(p = ∑n−1i=0 cibi ≥ cn,b0,b1, . . .): Compute the sum of∑n−1i=0 cibi over the supplied arguments bi, and return TRUE iff the sum ≥ cn.In practice, we maintain a running sum between calls to evaluate, and onlyupdate it as the theory literals bi are assigned or unassigned.Implementation of analyze(p = ∑n−1i=0 cibi ≥ cn): Return the default monotonicconflict set (as described in Section 4.2).Implementation of analyze(¬maxFlows,t ,G+,m−,c+0 ,c+1 , . . .): Return the de-fault monotonic conflict set (as described in Section 4.2).197Appendix D. Monotonic PredicatesD.4 CTL Model CheckingIn Chapter 8, we describe a predciate for CTL model checking, along withassociated functions. We describe theory propagation in that chapter, however theconflict analysis procedure for the theory of CTL model checking handles eachCTL operator separately, and as a result was too long to include in the main bodyof Chapter 8. The complete procedure follows:function ANALYZECTL(φ ,s,K+,K−,M)Let (T+,P+) = K+Let (T−,P−) = K−c←{}if φ is EX(ψ) thenfor each transition t outgoing from s doif (t /∈ T ) ∈M thenc← c∪{(t ∈ T )}.for each transition t = (s,u) in T+ doif evaluate(ψ,u,K+) 7→ FALSE thenc← c∪ANALYZECTL(ψ,n,K+,K−,M).else if φ is AX(ψ) thenLet t = (s,u) be a transition in T−, with evaluate(ψ,u,K+) 7→ FALSE.(At least one such state must exist)c← c∪{(t /∈ T )}c← c∪ANALYZECTL(ψ,u,K+,K−,M).else if φ is EF(ψ) thenLet R be the set of all states reachable from s in T+.for each state r ∈ R dofor each transition t outgoing from r doif (t /∈ T ) ∈M thenc← c∪{(t ∈ T )}c← c∪ANALYZECTL(ψ,r,K+,K−,M).198Appendix D. Monotonic Predicateselse if φ is AF(ψ) thenLet L be a set of states reachable from s in T−, such that L forms a lassofrom s, and such that for ∀u ∈ L,evaluate(ψ,r,K+) 7→ FALSE.for each transition t ∈ lasso doc← c∪{(t /∈ T )}for each state u ∈ L doc← c∪ANALYZECTL(ψ,u,K+,K−,M)else if φ is EG(ψ) thenLet R be the set of all states reachable from s in T+, while only takingtransitions to states v for which evaluate(ψ,v,K+) 7→ TRUE.for each state r ∈ R dofor each transition t = (r,v) outgoing from r doif (t /∈ T ) ∈M thenc← c∪{(t ∈ T )}if evaluate(ψ,v,K+) 7→ FALSE thenc← c∪ANALYZECTL(ψ,v,K+,K−,M)}else if φ is AG(ψ) thenLet P be a path of transitions in T− from s to some state r for whichevaluate(ψ,r,K+) 7→ FALSE.for each transition t ∈ P doc← c∪{(t /∈ T )}c← c∪ANALYZECTL(ψ,r,K+,K−,M)}else if φ is EW (ψ1,ψ2) thenLet R be the set of all states reachable from s in T+, while only takingtransitions to states v for which evaluate(ψ1,v,K+) 7→ TRUE.for each state r ∈ R doc← c∪ANALYZECTL(ψ2,r,K+,K−,M)}for each transition t = (r,v) outgoing from r doif (t /∈ T ) ∈M thenc← c∪{(t ∈ T )}if v /∈ R and evaluate(ψ1,v,K+) 7→ FALSE thenc← c∪ANALYZECTL(∨(ψ1,ψ2),v,K+,K−,M)}199Appendix D. Monotonic Predicateselse if φ is AW (ψ1,ψ2) thenLet R be a set of states, for which there is path P of transitions in T−from s to some state v, such that each of the following conditions hold:1) ∀r ∈ R,evaluate(ψ2,r,K+) 7→ FALSE2) ∀r ∈ R/{v},evaluate(ψ1,r,K+) 7→ TRUE3) evaluate(ψ1,v,K+) 7→ FALSE.for each transition t ∈ P doc← c∪{(t /∈ T )}c← c∪ANALYZECTL(ψ1,v,K+,K−,M)}for each state r ∈ R/{v} doc← c∪ANALYZECTL(ψ2,r,K+,K−,M)}else if φ is EU(ψ1,ψ2) thenLet R be the set of all states reachable from s in T+, while only takingtransitions to states v for which evaluate(ψ1,v,K+) 7→ TRUE.for each state r ∈ R doc← c∪ANALYZECTL(ψ2,r,K+,K−,M)}for each transition t = (r,v) outgoing from r doif v /∈ R and (t /∈ T ) ∈M thenc← c∪{(t ∈ T )}if evaluate(ψ1,v,K+) 7→ FALSE thenc← c∪ANALYZECTL(∨(ψ1,ψ2),v,K+,K−,M)}else if φ is AU(ψ1,ψ2) thenif there exists a set of states R, for which there is path P of transitions inT− from s to some state v, such that each of the following conditions hold:1) ∀r ∈ R,evaluate(ψ2,r,K+) 7→ FALSE2) ∀r ∈ R/{v},evaluate(ψ1,r,K+) 7→ TRUE3) evaluate(ψ1,v,K+) 7→ FALSE.then return ANALYZECTL(AW (ψ1,ψ2),s,K+,K−,M)}elsereturn ANALYZECTL(AF(ψ2),s,K+,K−,M)}200Appendix D. Monotonic Predicateselse if φ is ∨(ψ1,ψ2) thenc← c∪ANALYZECTL(ψ1,s,K+,K−,M)}c← c∪ANALYZECTL(ψ2,s,K+,K−,M)}else if φ is ∧(ψ1,ψ2) thenif evaluate(ψ1,s,K+) 7→ FALSE and evaluate(ψ2,s,K+) 7→ FALSE thenLearn the smaller of the two conflict sets for ψ1,ψ2c1← ANALYZECTL(ψ1,s,K+,K−,M)}c2← ANALYZECTL(ψ2,s,K+,K−,M)}if |c1| ≤ |c2 thenc← c∪ c1elsec← c∪ c2else if evaluate(ψ1,s,K+) 7→ FALSE thenc← c∪ANALYZECTL(ψ1,s,K+,K−,M)}elsec← c∪ANALYZECTL(ψ2,s,K+,K−,M)}else if φ is the negation of a property state ¬p thenc← c∪{¬(p(s))}else if φ is property state p thenc← c∪{(p(s))}return c201
- Library Home /
- Search Collections /
- Open Collections /
- Browse Collections /
- UBC Theses and Dissertations /
- SAT modulo monotonic theories
Open Collections
UBC Theses and Dissertations
Featured Collection
UBC Theses and Dissertations
SAT modulo monotonic theories Bayless, Sam 2017
pdf
Page Metadata
Item Metadata
Title | SAT modulo monotonic theories |
Creator |
Bayless, Sam |
Publisher | University of British Columbia |
Date Issued | 2017 |
Description | Satisfiability Modulo Theories (SMT) solvers are a class of efficient constraint solvers which form integral parts of many algorithms. Over the years, dozens of different Satisfiability Modulo Theories solvers have been developed, supporting dozens of different logics. However, there are still many important applications for which specialized SMT solvers have not yet been developed. We develop a framework for easily building efficient SMT solvers for previously unsupported logics. Our techniques apply to a wide class of logics which we call monotonic theories, which include many important elements of graph theory and automata theory. Using this SAT Modulo Monotonic Theories framework, we created a new SMT solver, MonoSAT. We demonstrate that MonoSAT improves the state of the art across a wide body of applications, ranging from circuit layout and data center management to protocol synthesis - and even to video game design. |
Genre |
Thesis/Dissertation |
Type |
Text |
Language | eng |
Date Available | 2017-03-30 |
Provider | Vancouver : University of British Columbia Library |
Rights | Attribution-NonCommercial-NoDerivatives 4.0 International |
DOI | 10.14288/1.0343418 |
URI | http://hdl.handle.net/2429/61062 |
Degree |
Doctor of Philosophy - PhD |
Program |
Computer Science |
Affiliation |
Science, Faculty of Computer Science, Department of |
Degree Grantor | University of British Columbia |
Graduation Date | 2017-05 |
Campus |
UBCV |
Scholarly Level | Graduate |
Rights URI | http://creativecommons.org/licenses/by-nc-nd/4.0/ |
Aggregated Source Repository | DSpace |
Download
- Media
- 24-ubc_2017_may_bayless_samuel.pdf [ 1.08MB ]
- Metadata
- JSON: 24-1.0343418.json
- JSON-LD: 24-1.0343418-ld.json
- RDF/XML (Pretty): 24-1.0343418-rdf.xml
- RDF/JSON: 24-1.0343418-rdf.json
- Turtle: 24-1.0343418-turtle.txt
- N-Triples: 24-1.0343418-rdf-ntriples.txt
- Original Record: 24-1.0343418-source.json
- Full Text
- 24-1.0343418-fulltext.txt
- Citation
- 24-1.0343418.ris
Full Text
Cite
Citation Scheme:
Usage Statistics
Share
Embed
Customize your widget with the following options, then copy and paste the code below into the HTML
of your page to embed this item in your website.
<div id="ubcOpenCollectionsWidgetDisplay">
<script id="ubcOpenCollectionsWidget"
src="{[{embed.src}]}"
data-item="{[{embed.item}]}"
data-collection="{[{embed.collection}]}"
data-metadata="{[{embed.showMetadata}]}"
data-width="{[{embed.width}]}"
async >
</script>
</div>
Our image viewer uses the IIIF 2.0 standard.
To load this item in other compatible viewers, use this url:
http://iiif.library.ubc.ca/presentation/dsp.24.1-0343418/manifest