基于演绎长度的学习子句删除策略

学习子句删除策略是CDCL-SAT求解器中的一个重要内容,可以避免内存爆炸和加速单元传播。评估学习子句有用性的标准不同导致所删除的学习子句是不同的,极大地影响求解效率。基于CDCL算法的求解过程可被形式化为增加管理学习子句策略的归结演绎过程,基于此,提出一种基于演绎长度的学习子句评估方法,并与现有的基于文字块距离的评估方法结合,根据排序子句的基准不同,形成两种不同的结合算法。采用国际SAT竞赛的基准实例,与目前主流的求解器进行了实验对比分析。结果表明,所提的结合算法能更好地评估学习子句的有用性,较基于文字块距离策略的求解个数提高了4.1%,说明所提策略具有一定的优势。     

基于学习子句删除策略的SAT求解器分支策略

对于SAT求解器,目前流行的分支变量决策策略大多是基于冲突的变量活跃度评估算法,选择具有最大活性的未赋值变量作为决策变量,优先解决最近的冲突.但是,它们都忽略了包含决策变量的子句数目对布尔约束传播(BCP)的影响.针对此问题,提出了 一种基于学习子句删除策略的分支变量决策策略(VDALCD),在删除学习子句的同时减小被删除子句中变量的活跃度.基于VDALCD策略分别对Glucose4.1,MapleLCMDistChronoBT-DL-v2.1进行改进,形成了求解器Glucose4.1_VDALCD和Maple-DL_VDALCD.以2018年、2019年SAT国际竞赛题为基准测试例,将改进版本与原版本求解器进行比较.实验结果表明,在2018年的例子测试中,Gluose4.1_VDALCD比Gluose4.1多求出26个例子,增加了 15.5％.在2019年的例子测试中,Maple-DL_VDALCD 比 MapleLCMDistChronoBT-DL-v2.1 多求出 17个例子,增加了 7.6％.     

基于趋势强度的SAT问题学习子句评估算法

针对命题逻辑公式求解过程中难以有效评估学习子句是否有利于后续搜索的问题,提出了一种基于学习子句趋势强度的评估算法。该算法首先通过分析学习子句在生存期内参与冲突分析的时间分布特征,将随机、离散的时间分布转换为连续的累积趋势强度;然后在删除周期达到时,通过设定趋势强度阈值删除在后续搜索过程中“不大可能”被使用的子句,保留“可能”被使用的子句;最后采用2015年、2016年SAT问题国际竞赛实例,将该算法与经典的活跃度评估算法和文字块距离(LBD)评估算法进行对比。实验结果表明,趋势强度评估算法在效率上明显优于活跃度评估算法,且求解的实例更多,同时与LBD算法基本持平。     

基于回跳层数的SAT求解器学习子句删除策略

目前学习子句删除策略广泛采用的是基于LBD的评估方式,LBD评估方式在每次执行删除时都会删除前一半LBD值大的学习子句,这种方式对LBD值大的学习子句的删除过于激进。针对此问题,提出了一种利用冲突回跳层数（back-jump levels）的评估方式来保留LBD值较大的相关学习子句。以CDCL（conflict driven clause learning）完备算法为框架,在子句删除环节形成了BJL删除算法。通过测试2017年SAT国际竞赛例,对新改进的版本与原版求解器进行了对比实验。实验表明,所提策略可显著提高求解器的求解性能和求解效率。     

基于频次的SAT问题学习子句混合评估算法

为了有效管理学习子句,避免学习子句规模呈几何级增长,减少冗余学习子句对系统内存占用,从而提高布尔可满足性问题SAT求解器的求解效率,需要对学习子句进行评估,然后删减学习子句。传统的评估方式是基于学习子句的长度,保留较短的子句。当前主流的做法一个是变量衰减和VSIDS的子句评估方式,另外一个是基于文字块距离LBD的评估方式,也有将二者结合使用作为子句评估的依据。通过对学习子句参与冲突分析次数与问题求解的关系进行分析,将学习子句使用频率与LBD评估算法混合使用,既反映了学习子句在冲突分析中的作用,也充分利用了文字与决策层之间的信息。以Syrup求解器（GLUCOSE 4.1并行版本）为基准,在评估算法与并行子句共享策略方面做改进测试,通过实验对比发现,混合评估算法比LBD评估算法有优势,求解问题个数明显增多。     

基于子句的动态检查强制文字的SAT求解器

检查强制文字是一种重要的预处理方法。结合学习子句,提出一种在求解过程中使用的策略—基于子句的动态检查强制文字(CNL),并且设计了一种易实现低成本的数据结构。分别实现了两个不同版本的求解器:Glucose_PRE和Glucose_CNL,前者在求解初始时将检查强制文字作为预处理,后者实现了基于子句的动态检查强制文字策略。实验测试结果表明,与Glucose_PRE和Glucose3.0求解器相比,求解器Glucose_CNL在求解2015年和2016年SAT竞赛的应用类型的实例时,求解实例个数更多,耗时更少,说明所提策略和所设计的数据结构均可提高求解器的求解性能。     

基于子句权重学习的求解SAT问题的遗传算法

该文提出了一种求解SAT问题的改进遗传算法（SAT—WAGA）．SAT-WAGA算法有多个改进性特点：将SAT问题的结构信息量化为子句权重,增加了学习算子和判定早熟参数,学习算子能根据求解过程中的动态信息对子句权重进行调整,以便防止遗传进程的早熟,同时,算法还采用了最优染色体保存策略,防止进化过程的发散．该文最后描述了实现包括SAT—WAGA等多个算法的实验系统,对选择最佳早熟判定参数值给出了一些有效的建议．实验结果表明：与一般遗传算法相比,SAT—WAGA算法在求解速度、成功率和求解问题的规模等方面都有明显的改善．     

基于Partial MAX-SAT求解法的RBAC授权查询方法

为保证系统的安全性并体现授权的有效性,结合部分最大可满足性问题(Partial MAX-SAT)的研究,提出一种基于Partial MAX-SAT求解法的授权查询方法。使用转换规则将静态授权逻辑和动态互斥角色约束转化为严格子句,采用子句更新算法将满足不同匹配的请求权限转化为松弛子句,并利用子句编码及递归算法寻求真值指派,以满足所有严格子句和尽可能多的松弛子句。实验结果表明,该方法搜索的角色组合能够保证系统的安全性,并满足最小权限分配要求,且最大、精确匹配请求的查询效率优于MAX-SAT求解法。     

基于扩展规则的启发式#SAT求解算法

#SAT在人工智能领域取得了广泛应用,很多现实问题可以规约成#SAT进行求解,得到命题理论的模型个数.通过对基于扩展规则的#SAT求解器的深入研究,发现选择规约子句的顺序对极大项空间的大小有着较大的影响,因此提出两种加速#SAT求解的启发式策略：MW和LC&MW.MW每次选择具有最大权值的子句作为规约子句;LC&MW每次选择最长子句作为规约子句,若最长子句存在多个,则在多个最长子句中选择具有最大权值的子句作为规约子句.利用MW策略设计了算法CER_MW,利用LC&MW策略设计了算法CER_LC&MW.实验结果表明,CER_MW和CER_LC&MW相对于先前的#SAT求解算法在求解效率和求解能力上都有显著的提高.在求解效率方面,CER_MW和CER_LC&MW的求解速度是其他算法的1.4倍~100倍.在求解能力方面,CER_MW和CER_LC&MW在限定时间内可解的测试用例更多.     

基于动态奖惩的分支策略的SAT完备算法

针对学习子句数量有限或相似度高导致历史信息有限、搜索树不平衡的问题,提出了基于动态奖惩的分支策略。首先,对每次单子句传播的变元进行惩罚,依据变元是否产生冲突和产生冲突的间隔,确立不同的惩罚函数;其次,在学习阶段,利用学习子句确定对构造冲突有益的变元,非线性增加它们的活跃度;最后,选择活跃度最大的变元作为新分支变元。在glucose3.0算法基础上,完成了改进的动态奖惩算法——AP7。实验结果表明,相比glucose3.0算法,AP7算法的剪枝率提高了14.2%~29.3%,少数算例剪枝率的提高可达51%,且改进后的AP7算法相比glucose3.0算法,运行时间缩短了7%以上。所提分支策略可以有效降低搜索树规模,使搜索树更加平衡,减少计算时间。     

Link inheritance in abstract clause graphs

Clause graphs, as they were defined in the 1970s, are graphs representing first order formulas in conjunctive normal form: the nodes are labelled with literals and the edges (links) connect complementary unifiable literals, i.e. they provide an explicit representation of the resolution possibilities. This report describes a generalization of this concept, called abstract clause graphs. The nodes of abstract clause graphs are still labelled with literals, the links however connect literals that are unifiable relative to a given relation between literals. This relation is not explicitely defined, only certain abstract properties are required. For instance the existence of a special purpose unification algorithm is assumed, which computes substitutions, the application of which makes the relation hold for two literals.When instances of already existing literals are added to the graph (e.g. due to resolution or factoring), the links to the new literals are derived from the links of their ancestors. An inheritance mechanism for such links is presented which operates only on the attached substitutions and does not have to unify the literals. The definition of abstract clause graphs and the theory about link inheritance is general enough to provide a framework so that as new ideas are proposed for graph based theorem provers, the methodology for both implementing links and proving their properties will be readily available.This research was supported by the Sonderforschungsbereich 314, Künstliche Intelligenz, of the German Research Agency.     

Branching Strategy Selection Approach Based on Vivification Ratio

The two most effective branching strategies LRB and VSIDS perform differently on different types of instances. Generally, LRB is more effective on crafted instances, while VSIDS is more effective on application ones. However, distinguishing the types of instances is difficult. To overcome this drawback, we propose a branching strategy selection approach based on the vivification ratio. This approach uses the LRB branching strategy more to solve the instances with a very low vivification ratio. We tested the instances from the main track of SAT competitions in recent years. The results show that the proposed approach is robust and it significantly increases the number of solved instances. It is worth mentioning that, with the help of our approach, the solver Maple_CM can solve additional 16 instances for the benchmark from the 2020 SAT competition.     

Solving the resolution-free SAT problem by submodel propagation in linear time

We present a method, called Unicorn-SAT, based on submodel propagation, which solves the resolution-free SAT problem in linear time. A formula is resolution-free if there are no two clauses which differ only in one variable, i.e., each clause is blocked for each literal in it. A resolution-free formula is satisfiable or it contains the empty clause. For such a restricted formula we can find a model in linear time by submodel propagation. Submodel propagation is hyper-unit propagation by a submodel generated from a minimal clause. Hyper-unit propagation is unit propagation simultaneously by literals, as unit clauses, of a partial assignment. We obtain a submodel, i.e., a part of the model, by negation of a neighbor-resolution-mate of a minimal clause, which is a clause with the smallest number of literals in the formula. We obtain a neighbor-resolution-mate of a clause by negating one literal in it. By submodel propagation we obtain a formula which has fewer variables and clauses and remains resolution-free. Therefore, we can obtain a model by joining the submodels while we perform submodel propagation recursively until the formula becomes empty.Sponsored by Upper Austrian Government (Ph.D. scholarship) and SFB/FWF project P1302.     

Solving the resolution-free SAT problem by submodel propagation in linear time

We present a method, called Unicorn-SAT, based on submodel propagation, which solves the resolution-free SAT problem in linear time. A formula is resolution-free if there are no two clauses which differ only in one variable, i.e., each clause is blocked for each literal in it. A resolution-free formula is satisfiable or it contains the empty clause. For such a restricted formula we can find a model in linear time by submodel propagation. Submodel propagation is hyper-unit propagation by a submodel generated from a minimal clause. Hyper-unit propagation is unit propagation simultaneously by literals, as unit clauses, of a partial assignment. We obtain a submodel, i.e., a part of the model, by negation of a neighbor-resolution-mate of a minimal clause, which is a clause with the smallest number of literals in the formula. We obtain a neighbor-resolution-mate of a clause by negating one literal in it. By submodel propagation we obtain a formula which has fewer variables and clauses and remains resolution-free. Therefore, we can obtain a model by joining the submodels while we perform submodel propagation recursively until the formula becomes empty.     

Inapproximability Results for Set Splitting and Satisfiability Problems with No Mixed Clauses

We prove hardness results for approximating set splitting problems andalso instances of satisfiability problems which have no “mixed”clauses, i.e., every clause has either all its literals unnegated orall of them negated. Results of Håstad imply tighthardness results for set splitting when all sets have size exactly $k\ge 4$ elements and also for non-mixed satisfiability problems withexactly $k$ literals in each clause for $k \ge 4$. We consider thecase $k=3$. For the MAX E3-SET SPLITTING, problem in which all sets have sizeexactly 3, we prove an NP-hardness result for approximating withinany factor better than ${\frac{19}{20}}$. This result holds even for satisfiableinstances of MAX E3-SET SPLITTING, and is based on a PCP construction due toHåstad. For “non-mixed MAX 3SAT,” we give aPCP construction which is a slight variant of the one given by Håstad for MAX 3SAT, and use it to prove the NP-hardness ofapproximating within a factor better than ${\frac{11}{12}}$.     

基于目标演绎距离的一阶逻辑子句集预处理方法

一阶逻辑定理证明是人工智能的核心基础,研究一阶逻辑自动定理证明器的相关理论和高效的算法实现具有重要的学术意义。当前一阶逻辑自动定理证明器首先通过子句集预处理约简子句集规模,然后通过演绎方法对定理进行判定。现有的应用于证明器中的子句集预处理方法普遍只从与目标子句项符号相关性角度出发,不能很好地从文字的互补对关系中体现子句间的演绎。为了在子句集预处理时从演绎的角度刻画子句间的关系,定义了目标演绎距离的概念并给出了计算方法,提出了一种基于目标演绎距离的一阶逻辑子句集预处理方法。首先对原始子句集进行包含冗余子句约简并应用纯文字删除规则,然后根据目标子句计算剩余子句集中的文字目标演绎距离、子句目标演绎距离,并最终通过设定子句演绎距离阈值来实现对子句集的进一步预处理。将该预处理方法应用于顶尖证明器Vampire,以2017年国际一阶逻辑自动定理证明器标准一阶逻辑问题组竞赛例为测试对象,在标准的300 s内,加入提出的子句集预处理方法的Vampire4.1相比原始的Vampire4.1多证明4个定理,能证明10个Vampire4.1未证明的定理,占其未证明定理总数的13.5%;在证明的定理中,提出的子句集预处理方法能对77.2%的子句集进行约简,最大子句集约简规模达到51.7%。实验结果表明,提出的一阶逻辑子句集预处理方法是一种有效的方法,能有效地约简一阶逻辑子句集的规模,提高一阶逻辑自动定理证明器的证明能力。