• <tr id="yyy80"></tr>
  • <sup id="yyy80"></sup>
  • <tfoot id="yyy80"><noscript id="yyy80"></noscript></tfoot>
  • 99热精品在线国产_美女午夜性视频免费_国产精品国产高清国产av_av欧美777_自拍偷自拍亚洲精品老妇_亚洲熟女精品中文字幕_www日本黄色视频网_国产精品野战在线观看 ?

    Uncertain Knowledge Reasoning Based on the Fuzzy Multi Entity Bayesian Networks

    2019-11-07 03:12:48DunLiHongWuJinzhuGaoZhuoyunLiuLunLiandZhiyunZheng
    Computers Materials&Continua 2019年10期

    Dun Li,Hong Wu,Jinzhu Gao,Zhuoyun Liu,Lun Li and Zhiyun Zheng,

    Abstract:With the rapid development of the semantic web and the ever-growing size of uncertain data,representing and reasoning uncertain information has become a great challenge for the semantic web application developers.In this paper,we present a novel reasoning framework based on the representation of fuzzy PR-OWL.Firstly,the paper gives an overview of the previous research work on uncertainty knowledge representation and reasoning,incorporates Ontology into the fuzzy Multi Entity Bayesian Networks theory,and introduces fuzzy PR-OWL,an Ontology language based on OWL2.Fuzzy PROWL describes fuzzy semantics and uncertain relations and gives grammatical definition and semantic interpretation.Secondly,the paper explains the integration of the Fuzzy Probability theory and the Belief Propagation algorithm.The influencing factors of fuzzy rules are added to the belief that is propagated between the nodes to create a reasoning framework based on fuzzy PR-OWL.After that,the reasoning process,including the SSFBN structure algorithm,data fuzzification,reasoning of fuzzy rules,and fuzzy belief propagation,is scheduled.Finally,compared with the classical algorithm from the aspect of accuracy and time complexity,our uncertain data representation and reasoning method has higher accuracy without significantly increasing time complexity,which proves the feasibility and validity of our solution to represent and reason uncertain information.

    Keywords:Ontology language,uncertainty representation,uncertainty reasoning,fuzzy multi entity Bayesian networks,belief propagation algorithm,fuzzy PR-OWL.

    1 Introduction

    Data acquisition technology,storage technology and computer performance have gained rapid development in the information era,and the global data size is doubled every 20 months.Moreover,data not only are big in volume,but also involve increasingly diversified types as well as complicated and changeable data relations.In order to make the computer automatically process and integrate valuable data in network,the semantic web is proposed.

    The increasing popularity of semantic web services makes it a challenging task to represent and reason uncertain data.However,the existing theory and practices do not provide enough feasible solutions to deal with the uncertainty of semantic web.To some extent,semantic markup languages,such as OWL(Ontology Web Language)or RDF(Resource Described Frame),can be used to signify the quality and quantity information of uncertainty,but they still have limitations.Previously proposed solutions include Probabilistic and Dempster-Shafer Models and Fuzzy and Possibilistic Models.BN and MEBN are two representative models based on probability and the corresponding ontology languages are OntoBayes based on BN[Yang(2007);Yang and Calmet(2006)],BayesOWL[Ding,Peng and Pan(2006)],and PR-OWL2 based on MEBN[Costa Laskey and Laskey(2008);Carvalho,Laskey and Costa(2013)],major ontology languages based on fuzzy model such as Fuzzy OWL[Calegari and Ciucci(2007)],etc.The existing uncertainty ontology languages cannot represent and reason probability and fuzzy information at the same time.

    Fuzzy Multi Entity Bayesian Networks(MEBN)[Golestan,Karray and Kamel(2013,2014)]was proposed to represent the fuzzy semantic and uncertainty relation between knowledge entities.This paper aims to extend the expression and reasoning ability of ontology for fuzzy probability knowledge and propose a method to denote and reason uncertainty.

    Based on the ontology language,Fuzzy PR-OWL of Multi Entity Bayesian Networks,in this paper,we propose a novel reasoning method of fuzzy probability according to Fuzzy MEBN.Our main contributions are as follows:

    (1)The fuzzy probability ontology of Fuzzy PR-OWL is created;

    (2)Through integrating the fuzzy probability theory and Bayesian network,the reasoning framework of Fuzzy PR-OWL is created and the reasoning method of fuzzy Belief Propagation are proposed;

    (3)Experimental results prove the feasibility and validity of our proposed reasoning method.

    2 Relevant theories

    2.1 Fuzzy set and fuzzy probability system

    Fuzzy set is a kind of description about fuzzy concept,which originates from the uncertainty and inaccuracy essence of abstract thought and concept,rather than the randomness of set elements.The logical value of fuzzy logic can be expressed with a real value in[0,1],which is the multivalued logic of fuzzy set.The fuzzy set theory transfers logical reasoning from two-valued logic into multivalued logic for predictive reasoning.

    Zadeh[Zadeh(1968)]defined the probability of fuzzy event by extending the classical theory of probability.Suppose thatxis the random variable in the sample spaceXandAis the subset that defines discrete events inX.The non-conditional probability ofAcan be calculated by Eq.(1):

    whereXA(x)is the binary characteristic function,that is,and.

    In fuzzy events,the characteristic function is the membership degree function.By replacingXA(x)in Eq.(3)with the membership degree functionμA(x):x∈[0,1],we can calculate the probability of fuzzy events.Berg et al.[Berg,Kaymak and Rui(2013)]calculated the probability of fuzzy eventAusing Eq.(2):

    And the probability of fuzzy discrete eventAcan be calculated by Eq.(3):

    wheref(x)is the probability density function of variablexandμ(x)is its membership degree function.Eq.(3)assumes that the conditional probability density function is correlated withμ(x).

    Suppose thatAc(c=1,…,C)is an event in sample spaceX,and each discrete sample belongs to one or zero event.If a samplexhas a group of membership degreesμA1,μA2,…,μAc,xis considered a fuzzy sample.To ensure the validity of probability theory,a good sample space is expected to meet the following condition:

    2.2 BN and MEBN

    Bayesian Networks(BN)[Heckerman(1997)]can deal with uncertain and probabilistic events according to the causal relationship of events or other relationships,and has the ability to process incomplete datasets,but BN has limitations when representing entity relationships.Fig.1(a)shows relevant knowledge of tracheitis represented by BN.Tracheitis can be triggered by either smoking or getting cold can in turn be triggered by weather factors.The causal relationship can be clearly seen in the Fig.1,but BN cannot show the influence of harmful gases produced by others' smoking behavior on the patients.Multi Entity Bayesian Networks(MEBN)[Chantas,Kitsikidis and Nikolopoulos(2014)]adopts the idea of “processing by division” by representing knowledge blocks with multientity fragments and then using multi-entity rules to link these fragments together.Fig.1(b)shows relevant information of tracheitis represented by MEBN.Inherent nodes,input nodes,and context nodes are indicated by ovals,trapezoids,and pentagons,respectively.otherandpersonare both examples ofPersontype.other=peopleAround(person)means thatotheris the neighbor toperson.In this way,MEBN can show the relationship between the entities as well as the influence of others' smoking on the patient's bronchitis by adding the nodeSmoke(other).

    However,human experience or knowledge is fuzzy by nature and therefore cannot be handled by MEBN.As shown in the above example,severeness of cold may affect the probability of suffering tracheitis.But MEBN can only use 0 or 1 to represent the possibility of suffering cold.In addition,when an inherent node has a nominal value,MEBN would assign equal probability to all nominal scales.For example,suppose that the weather today has two possible nominal scales,clear or cloudy,MEBN would assign the probability value 0.5 to both scales.It fails when the weather is partially cloudy.

    Figure 1:BN and MEBN Representation graph of bronchitis etiology

    2.3 Fuzzy MEBN and fuzzy PR-OWL

    2.3.1 Fuzzy MEBN

    Fuzzy Multi Entity Bayesian Networks(Fuzzy MEBN)redefines the grammar specification by combining the first-order logic and the fuzzy MEBN theory,allowing the context constraints of MEBN to express the fuzziness of incomplete semantic information[Zheng,Liu and Li(2016)].Hence,the expression and reasoning ability of inexact knowledge is improved.Therefore,fuzzy MEBN can represent fuzzy information that cannot be expressed with MEBN in Subsection 1.2,such asMild Cold{true0.3 1,false0.7 0},in which subscripts denote severeness of cold.Partiallycloudcan then be represented by{clear0.6 0.5,cloudy0.4 0.5},in which subscripts denote the membership degree.

    Fuzzy MEBN conducts modeling for specific domain ontology with the predefined attribute entity,semantics and their relationships.The attribute entity is represented by constant elements of a specific domain,and the variables are associated.The entities and relationships between entities are represented by logical and random variables of specific domain.As for the biggest difference between fuzzy MEBN and MEBN in entity and random variable symbol,the membership degree of a real number between 0 and 1 is added to the constant element symbol and entity identifier as subscript,such asVehicle0.85and!V 4280.75.The logic or value taking of random variable is the logical value in logical value chainL=<l1,l2,l3,...,ln >predefined by the language or within the scope of[0,1],rather than simpleTorF.

    The basic model of fuzzy MEBN is similar to that of MEBN.In fuzzy MEBN theory,the probability distribution is denoted by Fuzzy MEBN Fragments(FMFrag).The inherent random variable or input random variable value is set as the condition,and oneFMFragwill define one probability distribution or several fuzzy rules for inherent random variable.The formalized definition of FuzzyMFrag(FMFrag)isF=(C,I,R,G,D,S),where Cis the infinite set of context random variables,Irepresents the infinite set of input random variables,Rindicates the infinite set of inherent random variables,Grepresents the directed loop-free fragment graph,comprised of variable nodes inI∪R,Dmeans the local distribution of every inherent random variable inR,andSdenotes the set of fuzzy if-then rule used by fuzzy reasoning system.Context variables contain the semantic structure of knowledge that adopts fuzzy first-order logic.As the bridge with other inherent variables ofFMFlags,input variables will send relevant information to the currentFMFrag.Inherent variables is random variables that set the values of context and input variables as conditions.The setsC,IandRhave no intersection,and the random variable inIis the root node of fragment graphG.The context assignment item inCis used to strengthen the constraint of local distribution.

    InFMFrags,the context constraint is endowed with a true value,representing the degree of satisfying constraint.The consistency constraint degree ofFMFragsis determined by the fuzzy explanation of the item defined byFMFragand built-inFMFrags.When a common random variable is associated with an internal node,the aggregation function and the merging rule would be used.

    2.3.2 Fuzzy PR-OWL

    Fuzzy PR-OWL ontology adopted OWL2 DL language to define ontology class and attribute.In other words,OWL2 language is used to describe elements like classes and subclasses obtained from fuzzy MEBN concept abstraction.For instance,declaration node class(Node)can be expressed as function Declaration(Class(:Node)),and it represents the entity and random variable of fuzzy MEBN.Fig.2 shows Fuzzy PR-OWL ontology language(relation between classes and sub-classes)created with Protégé.Fuzzy PR-OWL ontology is based on fuzzy MEBN,and relevant classes and attributes of concepts like fuzzy random variable,fuzzy state,membership degree,and fuzzy rule are added based on PR-OWL2,to enhance the expressive force of uncertainty information.

    Figure 2:Fuzzy PR-OWL ontology elements

    The basic model of Fuzzy PR-OWL is presented in Fig.3,in which each piece of probability ontology has at least oneFMTheorytype,namely a label.It connects one group ofFMFragsto form an effectiveFMTheory.In Fuzzy PR-OWL grammar,the object attributehasFMFragis used to represent the connection.FMFragis composed of multiple nodes,and each node type is a random variable.The biggest difference between Fuzzy PROWL and PR-OWL2 is,when the fuzzy state of random variable node is defined,the set of fuzzy rule reflecting the state of domain expert is used to define the membership degree of fuzzy state,and the object attributehasFRSconnects each node to one or several fuzzy rule sets.The non-conditional or conditional probability distribution is represented with the classProbabilityDistribution,and is connected to their own nodes through the object attributehasProbDist.The object attributehasFMExpressionconnects every node to a logic expression or simple expression based on fuzzy first-order logic.

    Figure 3:Basic model of Fuzzy PR-OWL

    3 Reasoning based on fuzzy PR-OWL ontology

    3.1 Reasoning framework

    This section presents the fuzzy probability reasoning method based on Fuzzy PR-OWL ontology of fuzzy MEBN.Data fuzzification is first conducted and Situation-Specific Fuzzy Bayesian Network(SSFBN)of fuzzy MEBN is constructed.Then,fuzzy belief propagation based on fuzzy rule reasoning is conducted on the constructed network.The reasoning framework of Fuzzy PR-OWL ontology is mainly composed of four basic modules(as shown in Fig.4),including ontology analysis and memory module,data manipulation module,reasoning module,and representation module.

    Figure 4:Reasoning framework diagram of Fuzzy PR-OWL

    (1)Ontology analysis and TDB memory module

    Ontology analysis and TDB memory module is realized based on Jena API.It reads,analyzes the ontology constructed by Protégé,and stores it in the RDF database of Jena[Noy,Fergerson and Musen(2000)].The program first realizes the mapping between ontology URI and the corresponding ontology module file system address through URI address mapping file,and then reads various ontology module files via Java program or configuration files of Jena.In addition,data persistence is conducted in the RDF dataset.A built-in storage mode TDB officially recommended by Jena is used by memory module[Owens,Seaborne and Gibbins(2008)].Characterized by fast reading speed and simple operation,this module can store billions of records at most,and support hundreds of parallel queries.Hence,it is quite suitable for the demand of ontology reasoning.

    (2)Data manipulation module

    Data manipulation module implements the adding,deletion,modification and querying operations by utilizing the query operation interface of TDB.The module maps the loaded TDB model to corresponding Java ontology or examples.Moreover,by integrating adding,deletion,modification and querying operations through the ontology statement of the TDB dataset,Java ontology implements RDF database operations through TDB operation interface.

    (3)Reasoning module

    Reasoning module contains an implementation of the Belief Propagation reasoning algorithm[Shafer and Shenoy(1990)]based on fuzzy probability.Both FuzzyOWL reasoning engine and API of UnBbayes[Carvalho,Laskey and Costa(2010)]framework is used,and reasoning and querying can be conducted on specific nodes.Specifically speaking,it can handle knowledge base loading,fuzzy processing of data,SSFBN construction,and fuzzy probability reasoning functions.

    (4)Representation module

    Representation module handles GUI.It can manipulate the database through the graphical user interface,and display data variation in the database and reasoning results produced by the reasoning module.

    3.2 Belief propagation algorithm based on fuzzy probability

    In order to construct a Situation-Specific Bayesian Network(SSBN)based on knowledge and produce the minimum Bayesian network responding to query,common MEBN first sets prior distribution of target random variable and evidence random variable(also known as discovery random variable)as the condition.Then,posteriori distribution of target variable is reasoned through running the classic Bayesian network algorithm.

    3.2.1 Data fuzzification

    The algorithm proposed in this paper requires preprocessing the fuzzy concept or natural language variable gained from expert knowledge before establishing SSBN,so as to obtain the membership degree of fuzzy state of the variable.Some symbols and connotations involved in the algorithm need to be defined first.

    Definition 1:the set of entity identifiers.In FMFragF,the local node(or random variable)ψ(θ)has a set Nψcomposed offuzzy states,and the corresponding fuzzy set isA.Then:

    (1)f(μA,α)=(μA)α,in which μAis the membership degree function of setA,that is μA:U →[0,1];α∈[0,1]is the fuzzy factor.The functionfis a general form of the fuzzy theory.

    (2)The fuzzy state of ψ can be denoted by,in whichε,and αirepresents the membership degree of stateis the membership degree vector of ψ.

    (3)The probability distribution vector corresponding to CPT of ψ is denoted by πψ=

    In FMFragF,the local node ψ has multiple fuzzy states.For instance,if “height” is set as a random variable,there are three states(“short,medium,high”).“The person is relatively high” can be denoted by the fuzzy state <short0,medium0.2,high0.8>.

    Random variables are discrete variables,and the membership degree of fuzzy state of these random variables should be obtained through fuzzification and discretization.The method to calculate the membership degree of fuzzy state of random variables mainly depends on the type of expert data(as shown in Fig.5):

    Figure 5:Schematic diagram of data fuzzification

    (1)Discrete value input

    As the specific data x∈U is known,the membership degree can be calculated through the membership degree function μAi(x)(retain one digit after the decimal place).For instance,the height 1.8 can be substituted into membership degree functionμshort(x),μmedium(x),μhigh(x)defined in the domain of discourse[1.0,3.0],to determine the membership degree.

    (2)Language value input

    As the language's descriptor is known,the membership degree interval of corresponding extent description language variable can be determined using Tab.1.Created from researches on degree adverbs of modern Chinese language and the division of Vasilios et al.for degree in English[Carvalho,Laskey and Costa(2010);Moura and Roisenberg(2015)],Tab.1 gives the corresponding relation between language variable and membership degree suitable for Chinese context.In order to reduce the complexity,the average value was taken and the middle value in the interval was selected.Except the extreme case where 0 and 1 are taken in extremely low and high situations,the membership degree αAj:j≠iof other states ψ(θ)can be obtained using one of the following methods:

    1)If Aj:j≠ihas language description only and no membership degree function,αAjcan be obtained through the corresponding relation between language variable and membership degree;

    2)If Aj:j≠ihas membership degree function,x can be calculated through inverse function,and the membership degree αAjcan be obtained by substitutingxintoand retaining one digit after the decimal place.

    Table 1:Corresponding table of extent description language and fuzzy membership degree

    To further explain 2),two examples are given in the following.Suppose the random variable “weather” has two states{sunny,cloudy},and the current description is comparatively cloudy.According to Tab.1,the membership degree should be(0.2,0.8).If the sum of membership degrees of states is not equal to 1,standardization should be done to make the sum equal to 1.By taking “height” as an example,if we just know that someone is “very high”,the fuzzy membership degree of state “high” can be 0.9,and the rough height value x can be calculated through the inverse function.The membership degrees of“short” and “medium” can then be calculated throughμshort(x)andμmedium(x),that is,if αA?ig?=0.9,then αAs?ort=0,αmedium=0.1.

    3.2.2 Fuzzy IF-THEN rule

    In formal knowledge representation of reasoning system,IF-THEN rule is one of the basic approaches to represent the knowledge.This section will further expound the reasoning process of IF-THEN rule with the fuzzification.

    The general format of fuzzy IF-THEN rule is “IFxisAi,THENyisBj”,which decides the causal relationship of language variable[Zadeh(1968)].In this rule,xis the input variable,yindicates the output variable,andAiandBjrepresent the natural language values corresponding to the variables.They are represented with the fuzzy set determined by the model input.

    In IF-THEN rule of fuzzy MEBN,yandxrepresent the random variableψ(θ)and the language variable of its parent node πo(ψ(θ))respectively.Suppose thatand Boare the natural language values of variables;andrepresent the domain of discourseof rule o and the fuzzy set in Yorespectively.The IF-THEN rule can be simply represented by the following if-then rule statements:

    where O is the number of rules;N is the number of input variables;and Boindicate the language values of input random variables x1,…,xNand output random variableyof IF-THEN rule o;the fuzzy sets can be obtained through the method of the previous section,and denoted byrespectively,in whichand βjare the membership degrees of fuzzy states in fuzzy setsandseparately,is the number of states of input variable xι,andis the number of states of output variabley.

    Table 2:IF-THEN reasoning

    As shown in Tab.2,under the condition thatis given,according to Eq.(2),conclusioncan be calculated by Eq.(5):

    Each fuzzy MEBN node contains fuzzy and probabilistic information at the same time.Besides the IF-THEN rule,the Fuzzy PR-OWL ontology can also show the conditional probability distribution.When the conditional probability distribution is given,the fuzzy probability can be calculated through the following reasoning.The joint probability ofis the probability of fuzzy event,and can be calculated by Eq.(6):

    The probability of fuzzy event Aocan be calculated by Eq.(7):

    where T is the t-norm operator of weak conjunction calculated by the minimization process,× is the Cartesian product,and ∪isthe maximization operator.

    3.2.3 Fuzzy belief propagation algorithm

    In the BN reasoning algorithm,if Bayesian Network is small,simple marginal summation can meet the requirement,but the practical problem often involves a large scale.The overall operation complexity and data will show exponential growth.If the Belief Propagation algorithm is used to solve such a network problem,the operation complexity will only have a linear correlation with the number of nodes.Therefore,the Belief Propagation algorithm plays an increasingly important role when dealing with large-scale Bayesian network reasoning problems.

    The Clique Tree algorithm,also known as the Junction Tree algorithm,was proposed by Lauritzen et al.[Lauritzen and Spiegelhalter(1988)].The Clique Tree algorithm is the fastest Bayesian reasoning algorithm at present.The algorithm first converts the Bayesian network into a clique tree,and then performs calculation through message passing.The message will be broadcasted to every node of the clique tree to ensure consistency.

    In the Clique Tree Belief Propagation algorithm,every clique tree is set as a node to send message to neighbor nodes.Let clique treeTbe composed of clique C1,C2,…,Ck.Each φ∈Φ is assigned to clique α(φ),and the initial potential energy of Cjis defined by Eq.(9):

    In the equation,local evidence is corresponding to the prior probability or conditional probability of every node,i.e.,.For instance,P(a)=

    For clique Ci,βiis initialized as ψiaccording to Eqs.(9)and(3):

    Clique Cimultiplies all other neighbor messages with the initial clique potential energy,adds up all variables except variables in the cut set of Ciand Cj,and sends the result to Ci.Clique Cinode transmits the updated messages to other nodes after receiving messages of its neighbor nodes.The ultimate potential energy of Cican be obtained by Eq.(11):

    The messages of Ciand Cjare calculated by Eq.(12):

    where(1)bis Boolean number,indicating whether “0” exists in all mki.If “0” exists,b=t(true);otherwise,b=f(false);(2)zis the product of all messages whose value is not zero,i.e.,.

    In addition,the zero-perception table can also identify zero,and maps variable example into data that can detect zero.Suppose that Ψ is the table with zero detection and Φ is the standard table without zero detection,real(Ψ)is the standard table containing results obtained viareal(z,b)in table Ψ.

    The specific steps of Belief Propagation algorithm based on fuzzy probability are as follows:

    (1)Initialization

    The fuzzy set of(input or output)variable is obtained through fuzzification according to the language value of variable,and the generated SSFBN is divided.For every clique Ci,is calculated to initializeusing Eq.(10).The parameter of real(z,b)is initialized to(1,f).

    (2)Information transfer

    The message is transferred from the leaf node to the root node,and then transferred from the root node to the leaf node.In all directions,every cut set Si,jwill consider Φi,jas the previous message transmitted along the edge(i-j).This message can prevent repeated calculation,and every node will immediately update its belief after receiving messages from other nodes.It will multiply its belief with newly received message and divide the result by the message received from the same nodes last time.Hererealfunction is used to avoid meaningless message.

    The algorithm of Fuzzy-Probability Belief Propagation(FPBP)

    ProcedureJTree-FPBP(Φ,T)//Φ:set of factors,T:clique tree inΦ

    ProcedureInitialize-JTree(T)//Initialize the clique method

    ProcedureMessagePropagation(i,j)//Message passing method:isending clique,j receiving clique

    3.3 Establishment and reasoning of Situation-Specific Fuzzy Bayesian Network(SSFBN)

    The premise of MEBN reasoning is the generation of Situation-Specific Bayesian Network,i.e.,SSBN.The minimum BN that satisfies the query request is first obtained through MTheories of domain-specific knowledge and a group of observed results,and then the reasoning algorithm of Bayesian network is used to calculate the posterior probability of searching for the set of target random variables gathered from observation evidence.

    The algorithm of Fuzzy MEBN is a simple bottom-up algorithm and starts from the finite set containing target random variables and finite set of random variables gathered from observation evidence.The two sets are combined to build an approximate SSFBN.In the approximate SSFBN,if one node is the parent of another node or the context variable of its superior FMFrag,an edge between the two nodes is added.In each step,the algorithm builds a new approximate SSFBN by adding evidence nodes and instantiates superior FMFrags of random variables in the query sets and their ancestor nodes.The result random variables are added to the query set,any node irrelevant to query is removed,and the result sets of random variables are combined to create a new approximate SSFBN.The above process repeats until the approximate SSFBN remains unchanged or the stopping condition is met.If the algorithm has no stopping condition,the result of the algorithm is either an accurate response or a warning that the evidence information is inconsistent when the construction of SSFBN is done.If the algorithm does not stop,it produces many approximate SSFBNs that can generate correct query responses.Generally speaking,there is no evidence of finite length consistent with the query fragment,but inconsistent query fragments can be detected when SSFBN is constructed.

    To minimize the final SSFBN and improve the reasoning efficiency,while constructing SSFBN,evidence node,target node and internal node are retained while redundant node,barren node and d-separated node irrelevant to the ultimate result or with a small influence are removed.

    Suppose thatEis the set of observed evidences,Qis the set of queries,Bis Bayesian network,andVis the set of nodes.The construction process of SSFBN is consists of the following steps:

    (1)Initialization

    The set of queriesQis set to be the union set of random variables of target nodes and evidence nodes;the random variable exampleR0 is set to be Qand the set of evidencesE is set to be null;the maximum numberN0of states of every random variable is a positive integer user specified.

    (2)SSFBN Construction

    Suppose that the current SSFBNBiincludes nodes inRi,and all edges are defined in influential configurations.Remove any isolated nodes fromBi,and conduct dseparation from target nodes to evidence nodes,without updating the marginal distribution of nuisance nodes.Whenever a new evidence is added toE,a new SSFBN is constructed and any unnecessary node,such as barren node BNode(v)or worthless node NNode(v),including d-separated node connected to target node and marginal distribution node,is removed.Later,the FPBP algorithm in Section 3.2.3 can be used to calculate local distribution.Finally,the FBN algorithm is run according to Fuzzy Belief Propagation on the manicured networkBaccording toQandE.The goal of this algorithm is to construct the Bayesian network of context that meets the consistency condition.

    In the SSFBN construction algorithm,the situation-specific random variable node is generated based on query variableQand evidence setE.The context node of every random variable has passed satisfiability test.If the condition is met,this node and its parent node will be created.

    SSFBN construction algorithm

    ProcedureGenerate-SSFBN(Q,E)

    FBN reasoning algorithm

    ProcedureFBN-Reasoning(B,Q,E)

    (3)Local distribution construction

    We set local distribution inBi,modify the local distribution and restrict random variables toNipossible values,and estimate the influence of random variables not listed.Besides,the calculation should not exceedKisteps.

    (4)Reasoning

    Given evidence random variables,we calculate the local distribution of target random variables via the FPBP algorithm in Section 3.2.3.We run the Fuzzy Belief Propagation algorithm on the manicured networkBaccording toQandE,and calculate the conditional probability distribution of non-query nodes.Inconsistency of evidences is reported if the evidence random variable has probability 0,and if the query contents are consistent,the random variable of query set of the current SSFBN,Ri,is outputted.

    (5)Example enumeration and approximate parameter updating

    If the stopping condition is met,outputRi.Otherwise,the parent nodes of other random variables ofRiis added.As the distribution might be changed when extra parent nodes are added,NiandKiare added before going to Step 2.

    4 Experiment and analysis

    We used the tenfold cross validation method to evaluate the veracity of the FPBP algorithm by comparing it with the classic Belief Propagation algorithm(BP).Our test dataset is Iris,which consists of 50 samples from each of three species of Iris(Iris virginica,Iris setosa,andIris versicolor)and each sample has 4 attributes:sepal length:4.3 cm~7.9 cm;sepal width:2.0 cm~4.4 cm;petal length:1 cm~6.9 cm;petal width:0.1 cm~2.5 cm.

    The goal of our experiment was to predict the specie of an Iris flower from four attribute values including sepal length,sepal width,petal length,and petal width,as shown in Fig.6.

    Figure 6:Species fragment of Iris

    In the experiment,random variables,such as sepal lengthgetSepalLength(flower),are classified into three categories:short,medium and long.For the training data,the values in the sepal length column are first sorted in ascending order and the resulting sequence is bi,1,bi,2,…bi,nwhere the first subscriptiis the column index number,the second subscript represents the index in the sequence,andnis the size of training set.Then the sequence is equally divided into three parts and the boundaries of three categories can be determined,i.e.,short(< b1,n/3),medium(b1,n/3,b1,2n/3)and long(> b1,2n/3).Similar method can also be used to classify other random variables,as shown in Tab.3.

    Table 3:Category table of random variables of Iris

    By aiming at the characteristics of Iris dataset,the triangle membership degree function was used to calculate the membership degree from the values of random variables like petal length.Fig.7 shows the membership degree function of the Iris dataset,in whichais equivalent to the minimum value of random variable,bis the median bi,n/2of the attribute sequence,andcis the maximum value of this attribute.For instance,a,the membership degree of petal length random variable,has the minimum value(1 cm),bis the median(b3,n/2)of the petal length sequence b3,1,…,b3,n/2,…,b3,n,andcis the maximum value(6.9 cm).Suppose that the petal length of a specific line of data in training set isx.The membership degree vector can be calculated by using the membership degree function.

    Figure 7:Membership degree function of Iris

    We chose the tenfold cross validation method,which is the veracity test method of the common algorithm,to evaluate our proposed solution.We first create 10 sample data sets from the Iris dataset.9 of them were set as the training data in turn and 1 was used as the test data.The average classification accuracy(or error rate)of 10 runs was used to measure the performance of the algorithm.

    Tab.4 shows the tenfold cross validation results of the proposed FPBP algorithm(retain two digits after the decimal point).Meanwhile,we applied the same validation process to the BP algorithm using the same 10 sample data sets and compared the performance of two algorithms.As shown in Tab.5,the average classification accuracy of FPBP is 89.2%,which is nearly 34% higher than the average accuracy of BP(66.7%).Obviously,FPBP,the fuzzy probability algorithm that adds fuzzy membership degree calculation,has higher accuracy than the BP algorithm that simply calculates the probability,which demonstrates the veracity and validity of our algorithm.

    Table 4:Experimental results of tenfold cross validation of FPBP

    Table 5:Comparison of average classification accuracy between FPBP and BP

    The time complexity of the BP algorithm on the Iris dataset is T(n,s,p)=T1(n)*T3(p)*T2(s)=O f(n)* φ(p)*g(s),in whichnis the size of dataset,prepresents the number of random variables like petal length,andsis the number of variable states.By adding matrix multiplication related to number of categories,the time complexity of FPBP is as following:

    In practice,the size of data is often greater than the number of attributes and the number of categories,and the number of attributes is greater than the number of categories.Therefore,compared with the BP algorithm,the increase in time complexity of the FPBP algorithm is still within a reasonable range.

    5 Conclusions and future work

    In this paper,by integrating the Belief Propagation algorithm of Bayesian network with the fuzzy probability theory,a fuzzy belief propagation algorithm is proposed and applied to the representation and reasoning framework of Fuzzy PR-OWL ontology.Besides,the framework from ontology analysis to reasoning model was created for representation and reasoning of fuzzy probability knowledge.According to the tenfold cross validation experiment results and the time complexity analysis,the FPBP algorithm has much higher classification accuracy than the BP algorithm without increasing the time complexity too much.

    Although this paper has made some progress in reasoning study about fuzzy probability ontology,due to increasing complication and expansion of ontology,researchers in the field of uncertainty reasoning of ontology are still faced with many challenges.For example,the veracity of our algorithms highly depends on the selection of the membership degree function,and how to select appropriate membership degree function can be further explored.In addition,although our algorithm can handle most fuzzy probability reasoning problems of various scales,the efficiency can be further improved when the node number reaches a certain scale.

    Acknowledgments:The authors are grateful to the editors and reviewers for their suggestions and comments.This work was supported by National Key Research and Development Project(2018YFC0824400),National Social Science Foundation project(17BXW065),Science and Technology Research project of Henan(1521023110285),Higher Education Teaching Reform Research and Practice Projects of Henan(32180189).

    www.999成人在线观看| 国产av又大| av欧美777| 久久精品亚洲精品国产色婷小说| 日本vs欧美在线观看视频| 91麻豆精品激情在线观看国产 | www.999成人在线观看| 欧美人与性动交α欧美软件| 1024视频免费在线观看| 国产又爽黄色视频| 亚洲欧美一区二区三区久久| 久久久久视频综合| 大陆偷拍与自拍| 婷婷成人精品国产| 亚洲第一欧美日韩一区二区三区| 成在线人永久免费视频| 午夜免费成人在线视频| 国产精品1区2区在线观看. | 黄色毛片三级朝国网站| 亚洲熟女精品中文字幕| 日本精品一区二区三区蜜桃| 亚洲美女黄片视频| www日本在线高清视频| 女人精品久久久久毛片| 国产精品自产拍在线观看55亚洲 | 99国产综合亚洲精品| 无遮挡黄片免费观看| 香蕉久久夜色| 久久国产精品人妻蜜桃| 男人的好看免费观看在线视频 | 99riav亚洲国产免费| 久久狼人影院| 激情视频va一区二区三区| 欧美最黄视频在线播放免费 | 欧美国产精品va在线观看不卡| 久99久视频精品免费| 亚洲精品美女久久久久99蜜臀| 12—13女人毛片做爰片一| 国产成人精品久久二区二区免费| 十八禁人妻一区二区| 一进一出好大好爽视频| 亚洲av成人不卡在线观看播放网| 成人永久免费在线观看视频| www.精华液| x7x7x7水蜜桃| 中亚洲国语对白在线视频| 精品人妻1区二区| 777久久人妻少妇嫩草av网站| 亚洲精品久久午夜乱码| 免费不卡黄色视频| 久久精品亚洲av国产电影网| 国产有黄有色有爽视频| av不卡在线播放| 久久人人爽av亚洲精品天堂| 欧美另类亚洲清纯唯美| 亚洲免费av在线视频| 老司机午夜十八禁免费视频| 亚洲国产欧美一区二区综合| 亚洲熟妇熟女久久| 91麻豆精品激情在线观看国产 | 91字幕亚洲| 999久久久精品免费观看国产| 欧美人与性动交α欧美软件| 国产成人精品久久二区二区免费| 久久国产亚洲av麻豆专区| 国产精品电影一区二区三区 | 日本五十路高清| 亚洲精品国产区一区二| 亚洲国产精品合色在线| 亚洲国产欧美网| 免费一级毛片在线播放高清视频 | 亚洲视频免费观看视频| 亚洲精品久久成人aⅴ小说| 巨乳人妻的诱惑在线观看| 午夜精品国产一区二区电影| 亚洲精品国产区一区二| 午夜两性在线视频| 99国产精品一区二区三区| 亚洲av成人一区二区三| 亚洲中文日韩欧美视频| xxx96com| 欧美日韩亚洲国产一区二区在线观看 | 伊人久久大香线蕉亚洲五| 999久久久精品免费观看国产| 国产精品98久久久久久宅男小说| 国产精品一区二区精品视频观看| 国产精品国产高清国产av | 午夜福利一区二区在线看| 国产真人三级小视频在线观看| 欧美日韩黄片免| 亚洲色图av天堂| 免费在线观看视频国产中文字幕亚洲| 黄色a级毛片大全视频| 91大片在线观看| 女人被狂操c到高潮| 精品一区二区三区四区五区乱码| 嫁个100分男人电影在线观看| 这个男人来自地球电影免费观看| a级毛片在线看网站| 亚洲黑人精品在线| 成人免费观看视频高清| 捣出白浆h1v1| 一级a爱片免费观看的视频| 美女午夜性视频免费| 大香蕉久久网| 日韩大码丰满熟妇| 日本一区二区免费在线视频| 国产亚洲欧美在线一区二区| 黄色视频,在线免费观看| 一级片免费观看大全| 国产国语露脸激情在线看| 国产精品欧美亚洲77777| 欧洲精品卡2卡3卡4卡5卡区| 欧洲精品卡2卡3卡4卡5卡区| 宅男免费午夜| 亚洲国产毛片av蜜桃av| 午夜影院日韩av| 亚洲av成人不卡在线观看播放网| 亚洲三区欧美一区| 一二三四在线观看免费中文在| 老熟女久久久| av线在线观看网站| 飞空精品影院首页| 亚洲精品一卡2卡三卡4卡5卡| 岛国在线观看网站| 99国产精品一区二区三区| 美女国产高潮福利片在线看| 色播在线永久视频| 看黄色毛片网站| 久久久久久久久免费视频了| 久久国产精品人妻蜜桃| 一级毛片精品| 久久精品国产综合久久久| 久久 成人 亚洲| 国产激情欧美一区二区| 午夜免费鲁丝| svipshipincom国产片| 久久精品国产99精品国产亚洲性色 | 精品电影一区二区在线| 久热爱精品视频在线9| 日韩成人在线观看一区二区三区| 午夜两性在线视频| 老熟妇仑乱视频hdxx| tube8黄色片| 男人舔女人的私密视频| www.自偷自拍.com| 黑人巨大精品欧美一区二区蜜桃| 十八禁高潮呻吟视频| 色在线成人网| 一进一出好大好爽视频| 99精品欧美一区二区三区四区| 欧美日本中文国产一区发布| 国产亚洲欧美在线一区二区| 精品国产乱码久久久久久男人| 性色av乱码一区二区三区2| 亚洲黑人精品在线| 麻豆av在线久日| 香蕉国产在线看| 一级片免费观看大全| 午夜免费观看网址| 怎么达到女性高潮| 亚洲欧美色中文字幕在线| 亚洲精华国产精华精| 久久青草综合色| 可以免费在线观看a视频的电影网站| 最新在线观看一区二区三区| 美女扒开内裤让男人捅视频| 精品一区二区三区av网在线观看| 好看av亚洲va欧美ⅴa在| 久热爱精品视频在线9| 丝袜在线中文字幕| 午夜福利视频在线观看免费| 国产欧美日韩一区二区三| 19禁男女啪啪无遮挡网站| 最近最新中文字幕大全免费视频| 最新的欧美精品一区二区| 久99久视频精品免费| 两人在一起打扑克的视频| 免费久久久久久久精品成人欧美视频| 国产人伦9x9x在线观看| 日日摸夜夜添夜夜添小说| 热re99久久国产66热| 不卡av一区二区三区| 美女国产高潮福利片在线看| 久热爱精品视频在线9| 国产亚洲av高清不卡| 欧美日韩黄片免| 色综合欧美亚洲国产小说| 精品久久蜜臀av无| 精品欧美一区二区三区在线| 久久久久视频综合| 婷婷丁香在线五月| www.熟女人妻精品国产| 高清av免费在线| 精品国内亚洲2022精品成人 | 精品国产一区二区久久| 69精品国产乱码久久久| 99久久精品国产亚洲精品| 制服诱惑二区| 操美女的视频在线观看| 欧美黄色淫秽网站| 国产精品久久电影中文字幕 | 亚洲一区高清亚洲精品| 久99久视频精品免费| 国产精品亚洲av一区麻豆| 丰满人妻熟妇乱又伦精品不卡| 伦理电影免费视频| 欧美精品av麻豆av| 精品国产乱子伦一区二区三区| 亚洲av成人一区二区三| 在线十欧美十亚洲十日本专区| 建设人人有责人人尽责人人享有的| 欧美精品高潮呻吟av久久| 宅男免费午夜| 国产欧美日韩精品亚洲av| 久久人妻福利社区极品人妻图片| 久久中文字幕一级| 最近最新中文字幕大全免费视频| 最近最新中文字幕大全免费视频| 欧美日韩成人在线一区二区| 欧美日韩视频精品一区| 视频区图区小说| 成年女人毛片免费观看观看9 | 欧美国产精品一级二级三级| 国产精品永久免费网站| 亚洲 欧美一区二区三区| 搡老熟女国产l中国老女人| 午夜两性在线视频| 啦啦啦免费观看视频1| 国产三级黄色录像| a在线观看视频网站| 在线av久久热| 欧美久久黑人一区二区| 欧美人与性动交α欧美软件| 亚洲精品国产区一区二| tube8黄色片| 日韩 欧美 亚洲 中文字幕| 不卡一级毛片| 日韩欧美一区视频在线观看| 亚洲av日韩精品久久久久久密| 午夜精品久久久久久毛片777| √禁漫天堂资源中文www| 男男h啪啪无遮挡| 黄网站色视频无遮挡免费观看| 天天影视国产精品| 色94色欧美一区二区| 亚洲一区二区三区欧美精品| 亚洲欧美激情综合另类| 9色porny在线观看| 侵犯人妻中文字幕一二三四区| 国产一区在线观看成人免费| 最近最新免费中文字幕在线| 人妻久久中文字幕网| 一级毛片高清免费大全| 在线观看www视频免费| 国产主播在线观看一区二区| 国产一区二区三区在线臀色熟女 | 精品福利永久在线观看| 日本vs欧美在线观看视频| 两个人看的免费小视频| 欧美激情久久久久久爽电影 | 免费看十八禁软件| 久久精品国产99精品国产亚洲性色 | 老司机午夜福利在线观看视频| 怎么达到女性高潮| 水蜜桃什么品种好| 国产在线观看jvid| 天天添夜夜摸| 一级作爱视频免费观看| 国产精品欧美亚洲77777| 曰老女人黄片| 免费少妇av软件| 亚洲性夜色夜夜综合| 国产国语露脸激情在线看| 国产片内射在线| 老司机在亚洲福利影院| 巨乳人妻的诱惑在线观看| 99精国产麻豆久久婷婷| 99香蕉大伊视频| xxxhd国产人妻xxx| 丰满饥渴人妻一区二区三| 免费在线观看完整版高清| 在线视频色国产色| 99国产精品免费福利视频| 亚洲五月天丁香| 高清av免费在线| 亚洲va日本ⅴa欧美va伊人久久| 久久精品国产a三级三级三级| 亚洲av成人av| 夜夜爽天天搞| 欧美一级毛片孕妇| 搡老熟女国产l中国老女人| 国产亚洲精品第一综合不卡| 在线天堂中文资源库| 麻豆国产av国片精品| 亚洲在线自拍视频| 在线观看www视频免费| 久久精品国产亚洲av香蕉五月 | 真人做人爱边吃奶动态| 黄色 视频免费看| 亚洲精品一二三| a级毛片在线看网站| 成熟少妇高潮喷水视频| 不卡av一区二区三区| tube8黄色片| 午夜福利免费观看在线| 中文欧美无线码| 人人妻人人添人人爽欧美一区卜| 精品一区二区三区视频在线观看免费 | 狂野欧美激情性xxxx| 在线十欧美十亚洲十日本专区| 在线观看一区二区三区激情| 久久人妻熟女aⅴ| 校园春色视频在线观看| 九色亚洲精品在线播放| 国产主播在线观看一区二区| 久久久久久久久久久久大奶| 在线播放国产精品三级| 国产精品亚洲一级av第二区| 国产蜜桃级精品一区二区三区 | 久久精品亚洲熟妇少妇任你| 欧美在线黄色| 最新的欧美精品一区二区| 黄色毛片三级朝国网站| 久久久久久久国产电影| 亚洲精品粉嫩美女一区| 精品一区二区三区四区五区乱码| 一级毛片精品| 国产精品影院久久| 国产伦人伦偷精品视频| 校园春色视频在线观看| 757午夜福利合集在线观看| 国产一区有黄有色的免费视频| 大陆偷拍与自拍| 国产成人免费观看mmmm| 亚洲va日本ⅴa欧美va伊人久久| 亚洲欧美色中文字幕在线| 国产成人精品在线电影| 久久久精品国产亚洲av高清涩受| 久久国产亚洲av麻豆专区| 欧美日韩亚洲高清精品| 两个人免费观看高清视频| 桃红色精品国产亚洲av| 久久热在线av| 丰满饥渴人妻一区二区三| 日韩欧美三级三区| 成年人黄色毛片网站| 窝窝影院91人妻| 伊人久久大香线蕉亚洲五| 欧美日韩中文字幕国产精品一区二区三区 | 欧美黑人精品巨大| 久久久久久久久久久久大奶| 中文字幕av电影在线播放| 不卡一级毛片| av有码第一页| e午夜精品久久久久久久| 亚洲午夜理论影院| 亚洲美女黄片视频| 51午夜福利影视在线观看| 亚洲一区高清亚洲精品| 色婷婷久久久亚洲欧美| 国产男靠女视频免费网站| 亚洲九九香蕉| 久久久久国产一级毛片高清牌| 国产精品国产高清国产av | 在线观看免费日韩欧美大片| 国产精品欧美亚洲77777| 一边摸一边做爽爽视频免费| 在线免费观看的www视频| 一区在线观看完整版| 真人做人爱边吃奶动态| 婷婷精品国产亚洲av在线 | 免费观看精品视频网站| 黄色 视频免费看| 精品人妻熟女毛片av久久网站| 日韩有码中文字幕| 中文字幕高清在线视频| 男女午夜视频在线观看| 99精国产麻豆久久婷婷| e午夜精品久久久久久久| 热99国产精品久久久久久7| 国产一区有黄有色的免费视频| 欧美精品人与动牲交sv欧美| 国产成人啪精品午夜网站| 两性午夜刺激爽爽歪歪视频在线观看 | 久热这里只有精品99| 少妇裸体淫交视频免费看高清 | 十八禁人妻一区二区| 丰满迷人的少妇在线观看| 国产精品久久久久久精品古装| av有码第一页| 女人精品久久久久毛片| 午夜激情av网站| 两个人免费观看高清视频| 国产一区二区激情短视频| 国产熟女午夜一区二区三区| 亚洲精品av麻豆狂野| 亚洲美女黄片视频| 一级,二级,三级黄色视频| 夜夜躁狠狠躁天天躁| 免费看a级黄色片| 久久中文字幕一级| 国产高清国产精品国产三级| 国产高清视频在线播放一区| 精品免费久久久久久久清纯 | 三上悠亚av全集在线观看| 久久中文字幕人妻熟女| e午夜精品久久久久久久| 国产精品98久久久久久宅男小说| 捣出白浆h1v1| 两人在一起打扑克的视频| 亚洲成人免费电影在线观看| 精品国产一区二区三区久久久樱花| 91大片在线观看| 新久久久久国产一级毛片| 亚洲av熟女| 免费在线观看日本一区| 村上凉子中文字幕在线| 亚洲五月婷婷丁香| 欧美性长视频在线观看| 日本黄色日本黄色录像| 欧美中文综合在线视频| 中文字幕制服av| 久久精品国产亚洲av高清一级| 免费观看精品视频网站| 搡老岳熟女国产| 中文字幕人妻熟女乱码| 曰老女人黄片| 国产精品.久久久| 国产精品永久免费网站| av中文乱码字幕在线| 91国产中文字幕| 99国产精品一区二区三区| 精品人妻熟女毛片av久久网站| 伦理电影免费视频| 色在线成人网| 国产欧美日韩一区二区精品| 欧美国产精品va在线观看不卡| 亚洲精品美女久久久久99蜜臀| 久久草成人影院| 亚洲 国产 在线| 中国美女看黄片| 两性午夜刺激爽爽歪歪视频在线观看 | 狠狠狠狠99中文字幕| 大片电影免费在线观看免费| 99精品欧美一区二区三区四区| 黑人巨大精品欧美一区二区mp4| 亚洲精品美女久久av网站| 国产亚洲欧美精品永久| 精品电影一区二区在线| 丝袜人妻中文字幕| 91在线观看av| 国产精品二区激情视频| 十分钟在线观看高清视频www| 999久久久精品免费观看国产| 欧美在线一区亚洲| 热re99久久国产66热| 在线观看一区二区三区激情| 午夜精品在线福利| 大片电影免费在线观看免费| 纯流量卡能插随身wifi吗| 国产精品偷伦视频观看了| 九色亚洲精品在线播放| 99国产精品一区二区蜜桃av | 欧美激情极品国产一区二区三区| 欧美国产精品一级二级三级| 欧美av亚洲av综合av国产av| 免费观看a级毛片全部| 9热在线视频观看99| videosex国产| 老司机午夜福利在线观看视频| 国产高清视频在线播放一区| 国产伦人伦偷精品视频| 日日夜夜操网爽| 亚洲国产看品久久| 亚洲人成电影免费在线| 欧美乱色亚洲激情| 捣出白浆h1v1| 欧美中文综合在线视频| 999久久久国产精品视频| 中国美女看黄片| 欧美乱色亚洲激情| 一二三四社区在线视频社区8| 大片电影免费在线观看免费| 日韩免费av在线播放| 一级片免费观看大全| 黑人操中国人逼视频| 欧美黄色片欧美黄色片| 成人国语在线视频| 三上悠亚av全集在线观看| 久久天躁狠狠躁夜夜2o2o| 香蕉久久夜色| 免费在线观看完整版高清| 色综合欧美亚洲国产小说| 乱人伦中国视频| 亚洲熟女精品中文字幕| 亚洲国产精品合色在线| 美女高潮到喷水免费观看| xxx96com| 两个人免费观看高清视频| 欧美黄色淫秽网站| 久久人人97超碰香蕉20202| 国产av又大| 国产视频一区二区在线看| 久久久久久久精品吃奶| 久久精品国产综合久久久| 国产一区二区激情短视频| 夜夜躁狠狠躁天天躁| 久久天堂一区二区三区四区| 婷婷丁香在线五月| 91字幕亚洲| 国产av又大| 精品少妇一区二区三区视频日本电影| 精品免费久久久久久久清纯 | 亚洲av成人av| 搡老乐熟女国产| 在线观看一区二区三区激情| aaaaa片日本免费| 精品一区二区三区视频在线观看免费 | 免费一级毛片在线播放高清视频 | 国产99久久九九免费精品| 啦啦啦免费观看视频1| 久9热在线精品视频| 国产精品久久久久成人av| 悠悠久久av| 国产99白浆流出| 丰满饥渴人妻一区二区三| 丝袜在线中文字幕| 亚洲全国av大片| 嫩草影视91久久| 久久久久久久国产电影| 色老头精品视频在线观看| 99re6热这里在线精品视频| 777米奇影视久久| 欧美av亚洲av综合av国产av| 午夜老司机福利片| 丰满饥渴人妻一区二区三| 免费看十八禁软件| 久热爱精品视频在线9| 男女下面插进去视频免费观看| 欧美乱色亚洲激情| 精品国产乱码久久久久久男人| 欧美日韩国产mv在线观看视频| 欧美日韩中文字幕国产精品一区二区三区 | 脱女人内裤的视频| 免费在线观看完整版高清| 亚洲精品国产色婷婷电影| 国产精品香港三级国产av潘金莲| 精品久久蜜臀av无| 国产日韩欧美亚洲二区| 老司机午夜十八禁免费视频| 久久人人97超碰香蕉20202| av福利片在线| 中文亚洲av片在线观看爽 | 亚洲欧洲精品一区二区精品久久久| 久久精品熟女亚洲av麻豆精品| 麻豆国产av国片精品| 水蜜桃什么品种好| 免费人成视频x8x8入口观看| av免费在线观看网站| av超薄肉色丝袜交足视频| av免费在线观看网站| 亚洲一区二区三区欧美精品| 久热爱精品视频在线9| 一区二区三区精品91| 精品国产超薄肉色丝袜足j| tocl精华| 女人被躁到高潮嗷嗷叫费观| 久久香蕉激情| 久久久久久人人人人人| 国产极品粉嫩免费观看在线| 伊人久久大香线蕉亚洲五| 无人区码免费观看不卡| 国产在线一区二区三区精| 日本vs欧美在线观看视频| tube8黄色片| 免费久久久久久久精品成人欧美视频| 久久精品aⅴ一区二区三区四区| 男女高潮啪啪啪动态图| 1024香蕉在线观看| 久久久久久久久免费视频了| 女人久久www免费人成看片| xxx96com| 久久国产精品大桥未久av| 色播在线永久视频| 美女高潮到喷水免费观看| 亚洲av片天天在线观看| 日韩欧美免费精品| www日本在线高清视频| 交换朋友夫妻互换小说| 一级a爱视频在线免费观看| 电影成人av| 国产亚洲一区二区精品| 久9热在线精品视频| 久久久久久久国产电影| x7x7x7水蜜桃| 国产在线观看jvid| 免费av中文字幕在线| 十八禁网站免费在线| 中文字幕高清在线视频| 精品国产乱码久久久久久男人| 又黄又爽又免费观看的视频| 视频在线观看一区二区三区| 老司机靠b影院| 国产成人一区二区三区免费视频网站| 国产男靠女视频免费网站| 777米奇影视久久| 国产精品国产av在线观看| 50天的宝宝边吃奶边哭怎么回事| 国产精品九九99| 欧美乱色亚洲激情| 国产精品久久视频播放| 亚洲精品中文字幕在线视频| 99久久国产精品久久久| 无遮挡黄片免费观看|