• <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 ?

    Two-Stage Robust Optimization Under Decision Dependent Uncertainty

    2022-07-18 06:17:26YunfanZhangFengLiuYifanSuYueChenZhaojianWangandJoCatal
    IEEE/CAA Journal of Automatica Sinica 2022年7期

    Yunfan Zhang, Feng Liu,,, Yifan Su, Yue Chen,,,Zhaojian Wang,,, and Jo?o P. S. Catal?o,,

    Abstract—In the conventional robust optimization (RO)context, the uncertainty is regarded as residing in a predetermined and fixed uncertainty set. In many applications, however,uncertainties are affected by decisions, making the current RO framework inapplicable. This paper investigates a class of twostage RO problems that involve decision-dependent uncertainties.We introduce a class of polyhedral uncertainty sets whose righthand-side vector has a dependency on the here-and-now decisions and seek to derive the exact optimal wait-and-see decisions for the second-stage problem. A novel iterative algorithm based on the Benders dual decomposition is proposed where advanced optimality cuts and feasibility cuts are designed to incorporate the uncertainty-decision coupling. The computational tractability,robust feasibility and optimality, and convergence performance of the proposed algorithm are guaranteed with theoretical proof.Four motivating application examples that feature the decisiondependent uncertainties are provided. Finally, the proposed solution methodology is verified by conducting case studies on the pre-disaster highway investment problem.

    I. INTRODUCTION

    A. Background

    UNCERTAINTIES widely exist in real-world decisionmaking problems and various mathematical programming techniques, including scenario-based or chance-constrained stochastic programs, robust optimization (RO) [1] and distributionally robust optimization [2], have been developed to fit in different characteristics of the applications [3]. RO seeks a risk-averse solution by explicitly considering the worst-case effect of all possible realizations of the uncertain parameter within a pre-determined uncertainty set. It appeals especially when the decision maker has no knowledge of the probability distributions of the uncertain parameters, or when the feasibility of the system over the entire uncertainty set is prioritized. Due to its salient advantages on modeling capability, feasibility, and computational tractability [4], RO has gained increasing popularity over the recent decades and encompassed a wide variety of applications including process scheduling [4]–[6], power system planning and scheduling[7]–[10], and network optimization [3], [11], [12], etc.

    In the context of RO, the uncertainty sets are usually assumed to be a priori and fixed. The underlying assumption is that the decision maker’s strategies would not alter the range of uncertainty realization. In many real-world decisionmaking problems, however, uncertainties can depend on the strategies chosen by the decision makers and are assumed as endogenous. For example, in a batch-process scheduling problem, the processing time or the production yield of a task is endogenously uncertain since it retains a physical meaning only when the optimizer decides to operate the associated task in a given period [4]–[6], [13]. Another example is the demand response program on buildings’ electricity consumption [8]. For buildings participating in the program, the reserve demand requested from the system operator is uncertain and of endogenous nature due to its dependency on the reserve capacity provided by the building in the day-ahead market. The uncertainties affected by decisions are called decision-dependent uncertainties (DDUs) or endogenous uncertainties. In this paper, we use these two terms interchangeably without distinction. Also, we refer to decisionindependent uncertainties (DIUs) or exogenous uncertainties as those not altered by decisions. Consideration of DDUs in RO problems can provide considerably less conservative solutions, attributed to the fact that DDUs can be proactively controlled by the optimizer. However, differently from DIUs,the presence of DDUs brings great challenges to solving RO problems, mainly due to the mutual influences between uncertainties and decisions. In this paper, we propose a convergence-guaranteed algorithm to efficiently solve the two-stage RO problems with DDUs.

    B. Literature Review

    Regarding RO under DDUs, existing works can be categorized into two classes, based on whether there are recourse or wait-and-see decisions after the revelation of uncertainty realization:

    1)Static RO Under DDUs:In static RO, which is also called single-stage RO, all the decisions are here-and-now,i.e., to be determined before observing the uncertainty realization. In [14] and [15], variable budgeted static RO is studied, where the budget parameter that regulates the conservatism of the uncertainty set is modeled as an affine function with respect to the decision variables. In [16],polyhedral and conic uncertainty sets are introduced where decisions control the upper bounds of the uncertain variables.In [3] and [4], uncertainties are governed by binary-valued materialization indicator variables. The materialization indicator attains the values of 1 if the corresponding uncertain parameter retains a physical meaning in the problem and 0 otherwise. The DDU-involved RO models in the above works maintain the advantage of computational tractability of conventional RO problems, since a robust counterpart of mixed-integer linear programming (MILP) formulation can be derived by applying the strong duality theory and McCormick envelopes convex relaxation.

    2)Adaptive RO Under DDUs:Adaptive robust optimization(ARO), or called adjustable RO [17], is established in a twostage or a more general multi-stage setting where recourse or wait-and-see decisions are incorporated in response to actual uncertainty realization. The computational complexity of ARO lies in the fact that multi-level programming problems with more than two decision levels are non-deterministic polynomial-time (NP)-hard [18]. To overcome this intrinsic computational burden as well as the challenges raised by the coupling relationship between the here-and-now decisions and the uncertainties, affine decision rules parameterized in the uncertainty realizations and the here-and-now decisions are postulated on the wait-and-see decisions in [5], [19]–[21].Differently from the aforementioned approximate affine policies, the decision rules in [22] are generated with the aid of multi-parametric programming technique, rendering the exact solution to the ARO problem; however, computation burden of solving the second-stage problem parametrically as an explicit function of the here-and-now decisions and the uncertainty is transferred off-line but not eliminated. In the line of primal cut algorithms [23], [24], [9] proposes an improved column-and-constraint generation (C&CG) algorithm based on a worst-case scenario mapping technique. Its application, however, is limited to a high-dimensional rectangle decision-dependent uncertainty set (DDUS).

    Compared with conventional ARO problems, the challenges of solving DDUs-incorporated ARO problems reside in the fact that the existing cutting plane algorithms [23]–[26], with constraints generated by whether primal or dual information of the second-stage problem, fail to warrant finite convergence to the global optimum when the uncertainty set depends on the here-and-now decision variable which changes with iterations. This is because in that case the optimality cuts and feasibility cuts generated with concrete uncertain parameter values may become invalid and ruin the optimality of the solution, or even render an infeasible master problem which falsely implies that the original ARO problem is not robust feasible. Moreover, since the vertices of the uncertainty set change with the here-and-now decision, it is difficult to justify the finite convergence of these cutting plane algorithms by completely enumerating the vertices of the polyhedral uncertainty set. To the best of the authors’ knowledge, the solution approach for ARO under DDUs with a generic decision-dependency and without any assumption on approximation policies has not been addressed.

    C. Contribution and Organization

    The main contributions of this paper are twofold:

    1)From a Modeling Perspective:This paper addresses a generic two-stage robust optimization model with a class of polyhedral DDUSs whose right-hand-side vector has a dependency on the here-and-now decisions. Differently from[3], [4], [6], [14]–[16] that contemplate the static RO involving endogenous uncertainty, our two-stage RO grants more feasibility by the implementation of wait-and-see decisions after the revelation of uncertainty realization.Compared with [5], [20], [21] that are established in a twostage or multi-stage setting and postulate affine decision rules on the wait-and-see decisions, our model seeks to derive the exact optimal solution to the second-stage problem. In view of the DDUS, we extend the polyhedral uncertainty sets with reduced decision-dependency considered in [6], [8], [9], [15],[16] into a generic form where both the shape and size of the uncertainty set may be altered by decisions. Due to the aforementioned improved modeling capability, our model can considerably eradicate the conservatism effects of RO by proactively controlling the level of uncertainties and inherently capture the trade-off between the robustness and conservatism of the solution.

    2)From a Solution Standpoint:To solve the two-stage RO model with DDUs, we propose a novel iterative solution algorithm involving one master problem and two subproblems, based on the Benders dual decomposition.Compared with the C&CG algorithm [23], [24] which is widely applied to conventional two-stage RO problems, the proposed algorithm overcomes the challenges stemming from endogenous worst-case uncertainty scenarios by using only dual information of the second-stage problem. Also differently from classical Benders dual cutting-plane algorithms[25]–[27], advanced optimality cuts and feasibility cuts are designed to accommodate the coupling between uncertain parameters and decision variables. Performance of the proposed algorithm, including convergence and optimality, is guaranteed with a strict proof. To implement the algorithm,we derive the robust counterpart of the master problem and the reformulations of the sub-problems, maintaining the computational tractability of the proposed solution algorithm.

    The remainder of the paper is organized as follows. The remainder of this section introduces the notation. The mathematical formulation of the two-stage RO with DDUs is provided in Section II. Section III proposes the solution approach based on modified Benders decomposition. Section IV addresses some implementation issues of the proposed algorithm, including the derivation of a tractable robust counterpart of the master problem which is a static RO and the reformulation of the bi-level sub-problems into MILPs. In Section V, four examples are provided as the applications of the proposed model and solution methodology. And we present a case study on the pre-disaster highway network investment problem. We conclude the paper in Section VI.

    Notation:Rnis the set ofn-dimensional real vectors and Rm×nis the set ofm×nreal matrices. Z ( Z+) denotes the set of(positive) integers. [n]?{1,...,n} denotes the set of integers from 1 tonand [0]??.Fora vectorx∈Rn(amatrixA∈Rm×n),xT(AT)denotes its transpose. Weuse1and0 to denote vectors of ones and zeros, respectively. Forx,y∈Rn,we denote theinnerproductxTy=xiyiwherexi,yistand for thei-th entries ofxandy, respectively. The Hadamard product is defined asx?y=(x1y1,...,xnyn)T, i.e., the elementwise product of two equally-sized vectors or matrices. For a polytopeX?Rn, vert(X) is the set of the vertices ofX. ?a」 is the largest integer less than or equal toa. We define the difference of setXandYasXY?{x:x∈X,x?Y}.

    II. PROBLEM FORMULATION

    A. Two-Stage RO With DDU

    The two-stage RO with DDU (TRO-DDU) problem is formulated as

    where

    The here-and-now decisionx∈RnR×ZnZ is a mixed-integer variablevectorwitha dimensionofnx=nR+nZ.X?RnR×ZnZis the feasibleregionofx.f(x):RnR×ZnZ →R1is the costfunction withrespecttox.TheDDU parameteris denoted byw∈Rnwand the DDUSofwisW(x):RnR×ZnZ2Rnw,which is essentially a set-valued map parameterized byxwhereG∈Rr×nwis a constant matrix,g∈Rris a constant vector andh(x):RnR×ZnZ →Rris a vector-valued function with respect tox. The wait-and-see decision vector isy∈Rny.We consider the case that the second-stage decision problem isalinearprogram(LP) andc∈Rnydenotesthecost coefficientofy.Y(x,w):RnR×ZnZ×Rnw2Rny,whichis essentially a set-valued map parameterized byxandw,denotes the feasible region ofywith a specific form in (1d)whereA∈Rm×nx,B∈Rm×ny,C∈Rm×nwandb∈Rmare constant parameters. The constraint setXR?RnR×ZnZ defined in (1e) is the set ofxthat are robust feasible. A hereand-now decisionxis called robust feasible [1], [26] if for any realization of the uncertainwwithinW(x) there exists at least one feasible wait-and-see decisionylying withinY(x,w). The TRO-DDU problem (1) aims at minimizing the total cost of the two stages, which isf(x)+cTy, under the worst-case uncertainty realization. The following assumptions are made on problem (1).

    Assumption 1:1)f(x) is a convex function; 2)Xis a convex set; 3)X∩XR≠?; 4)Xis bounded,W(x) is bounded for anyx∈XandY(x,w) is bounded for anyx∈Xandw∈W(x).

    Under Assumption 1-1) and 1-2), the nominal problem of(1) is a convex minimization problem. Assumption 1-3) and 1-4) imply that the TRO-DDU problem (1) is feasible and bounded, thus there exists the optimal solution to (1).

    The TRO-DDU problem distinguishes itself from the existing literature by the consideration of the DDUwand the DDUSW(x). Regarding the generality of the DDU modeling in the TRO-DDU problem (1), we have the following remark.

    Remark 1(Decision-Dependent Uncertainty Set):The polyhedral DDUSW(x) in (1c), whose right-hand-side vector has a dependency onx, covers the formulations proposed in the existing literature [8], [9], [14]–[16], [20], [21]. It provides generic modeling capability since both the shape and size of the uncertainty set can be altered byx. Moreover,W(x) in (1c)applies readily with DIUs by settingh(x)=0. Though the coefficient matrixGis fixed in (1c), the solution strategy proposed in Sections III and IV can be easily extended to the cases in [3]–[5] wherexaffects the coefficient matrix through certain binary-valued functions with respect tox.

    Regarding the difficulty in solving TRO-DDU problem (1),we have the following remark.

    Remark 2:The well-known C&CG algorithm is no longer applicable to TRO-DDU problem (1) since finite convergence to the global optimum is not warranted in the presence of DDUs.

    1)Failure in Robust Feasibility and Optimality:The cutting planes in the C&CG algorithm with recourse decision variables in the primal space are generated with concrete values of the uncertain parameter for each identified scenario;and these cuts can become invalid when the uncertainty set depends on the here-and-now decisionxwhich varies with iterations. Forexample,given ahere-and-now decisionx1,the identified worst-case uncertaintyrealizationw1∈W(x1)would nevercometrueifanotherhere-and-nowdecisionx2is adopted andw1?W(x2);andinthat case, thefeasibilityand optimality cuts generated according tow1no longer necessarily perform a relaxation to problem (1) and may ruin the optimality of the solution, or even render an infeasible master problem.

    2)Failure in Finite Convergence:The finite convergence of C&CG algorithm is justified by a complete enumeration of the vertices of the uncertainty set. However, in problem (1), the vertices set of polytopeW(x) changes withxwhich varies with iterations, thus the C&CG algorithm no longer necessarily terminates within a finite number of iterations.

    III. SOLUTION METHOD

    A. Equivalent Transformation

    1)Robust Optimality:Looking at the inner-level problem in

    (1), we denote the following max-min bi-level optimization byS(x):

    By noting the duality of the inner problem inS(x),S(x) can be equivalently transformed into the following bi-linear maximization problem:

    whereu∈Rmis the dual variable that corresponds with the constraintsAx+By+Cw≤b. Then, problem (1) is equivalent to

    Introducing a supplementary variable α ∈R1, then problem(4) can be further rewritten in an epigraph form as

    2)Robust Feasibility:The robust feasibility of decisionx,i.e., whetherxlies inXR, can be examined by solving the following relaxed bi-level problem:

    wheres∈Rmis the slack variable vector introduced to relax the feasible region ofydefined in (1d). IfR(x)≤0,xis robust feasible, i.e.,x∈XR.ElseifR(x)>0, theremustexista realizationw?∈W(x)suchthat nofeasiblewait-and-see decisionyis available. It is useful to write the dual of the inner minimization problem inR(x) . ThenR(x) is equivalent to the single-level bi-linear optimization problem as follows:

    where π ∈Rmis the dual variable vector corresponding with constraints (6c). Sincex∈XRif and only ifR(x)≤0, problem(5) can be further written as

    From the above transformation, we derive problem (8)which is the surrogate model of TRO-DDU problem (1).

    B. Modified Benders Decomposition Algorithm

    Next, we have the overall Algorithm 1 to solve problem (8)where the master problem (MP) involved is formulated as

    Regarding the difference between the proposed modified Benders dual decomposition (Algorithm 1) and the existing algorithms such as the C&CG algorithm and classic Benders dual cutting-plane algorithms, we give a remark as below.

    Algorithm 1 Modified Benders Dual Decomposition Algorithm for TRO-DDU Problem (1)

    Remark 3(Advanced Optimality and Feasibility Cuts):Constraints (9c) and (9d), which are appended to the master problem with iterations, are called optimality cuts and feasibility cuts, respectively. They are designed to have the following salient features to adapt to the DDUS: 1) Concrete worst-case uncertainty, i.e., the solution toR(xk) orS(xk)denoted bywk, is not involved in the optimality cuts or the feasibility cuts. This is different from the C&CG algorithm, as it incorporates the coupling relationship betweenxandw;2) Dual information of sub-problemsR(xk) andS(xk), i.e.,πkanduk, are utilized to generate the feasibility cuts and optimality cuts, inspired by the Benders dual decomposition.However, these cuts are designed to be no longer hyperplanes,but a set of static robust constraints, to comprise a cluster of endogenous worst-case uncertainty realizations.

    Regarding the extensive adaptability of Algorithm 1, we have the following remark.

    Remark 4(Adaptability to DDU Formulations):Algorithm 1 is designed to be applicable to a wide range of DDU formulations. Note that the explicit formulation of DDUSW(x)is not involved in Algorithm 1, nor in the theoretical justification of its performance. Thus, Algorithm 1 does not preclude the implementation of more DDUS structures besides polyhedrons, such as ellipsoidal and conic sets.However, considering the difficulty in solving master problem(9) and sub-problemsR(x) andS(x), here we focus on a class of polyhedral uncertainty sets with decision dependency in the right-hand-side vector, as formulated in (1c).

    We justify the convergence and optimality of Algorithm 1 by the following theorem.

    Theorem 1:Letqbe the number of vertices ofUdefined in(3b) andpbe the number of vertices of Π defined in (7c). Letx?denote the optimal solution to the TRO-DDU problem (1).Then, Algorithm1terminateswithinO(p+q)iterationsand outputsthe solutionxksuch thatxk∈X∩XRand|f(xk)+S(xk)?f(x?)?S(x?)|<ε.

    Theorem 1 indicates that the proposed modified Benders dual decomposition algorithm is finitely convergent to the optimum. Proof of Theorem 1 can be found in Appendix.

    IV. ROBUST COUNTERPART AND BIG-M REFORMULATION

    In this section, we show that the implementation of Algorithm 1 is computationally tractable, by deriving the robust counterpart of the master problem (9) and reformulations of sub-problemsR(xk) andS(xk).

    In Algorithm 1, the master problem (9) involves static robust constraints (9c) and (9d). Without loss of generality,we illustrate how to deal with the robust constraint (9c) in the master problem by substituting it with its robust counterpart.

    The following robust constraint with respect to (α,x) with given:

    A. Robust Counterpart of the Master Problem

    is equivalent to

    wherewesubstitutewbyfordenotingvariablein constraintformed by.Next,thefollowing twocasesare discussed:

    1)A Bi-Linear TermλTh(x)can be Precisely Linearized Through the Big-M Method Whereλ ∈Rris an r-Dimensional Variable and x is annx-Dimensional Variable:To illustrate this, we takex∈{0,1}nxandh(x)=Hxas an example. Duality of the inner-level problem in (11) is deployed, thus (11) can be reformulated as

    where λj∈Rris the dual variable corresponding with constraint (11b). Note that constraint (12) is equivalent to

    bydroppingthe minimizationoperation.Sincexisbinary,the bi-lineartermxT HTλjcanbeexactlylinearizedthroughthe big-Mmethodby introducingsupplementaryvariableηj∈Rnxand a largeenoughpositivenumberM,whereηj=x?(HTλj).Then, constraint (13) has the following equivalent formulation:

    which constitutes the robust counterpart of constraint (10).The robust counterpart of feasibility cuts (9d) can be derived in a similar way. Thus the master problem (9) has the following MILP robust counterpart:

    where the convex functionf(x) is substituted by its piecewise linearization form as in (15c) and the optimality cuts and the feasibility cuts are substituted by their robust counterparts.∈Rnx,∈R1(i∈[I])areconstant parameters for the piecewise linearizationoff(x).

    2)Otherwise:If a bi-linear term λTh(x) cannot be exactly linearized by the big-M method, we deploy the Karush-Kuhn-Tucker (KKT) conditions of the inner-level problem in (11) as follows:

    where λj∈Rris the dual variable corresponding with constraints (11b), and (16b) denotes the complementary relaxation condition. The nonlinearity of complementary condition (16b) can be eliminated through the big-M method by introducing the binary supplementary variablezj∈{0,1}r.Thus, the optimality cut (11) has the following equivalent robust counterpart:

    where the complementary relaxation condition (16b) is substituted by (17c) and (17d). The robust counterpart of the feasibility cut (9d) can be derived similarly. Thus the master problem (9) has the following robust counterpart which is a mixed-integer programming problem.

    B. Reformulation of Sub-Problems

    In Algorithm 1, the robust feasibility examination subproblemR(x) in (7) and the robust optimality sub-problemS(x)in (3) both have bi-linear objective, imposing difficulties on the solving. Next, we provide linear surrogate formulations ofS(x) andR(x). Since they have similar structure, without loss of generality we only focus onR(x).

    R(x) in (7) can be equivalently rewritten into

    where the here-and-now decisionxis given by solving the master problem. We deploy the KKT condition of the innerlevel LP problem in (19) as follows:

    where ζ ∈Rris the dual variable corresponding with constraintw∈W(x) in (19). The complementary constraints(20a) can be linearlized by introducing binary supplementary variablez∈{0,1}r. Then, the complementary constraints (20a)can be substituted with its equivalent linear formulations, like what we have done to (16b). Moreover, since strong duality holds, we substitute the ? πTCwin (19) by (g+h(x))Tζ. Then,sub-problemR(x) is equivalent to the following MILP problem:

    V. APPLICATIONS

    The proposed model and solution algorithm cater for a variety of application problems. In this section, we provide four motivating DDU-featured examples that can be formulated into the proposed TRO-DDU model (1). For the first three applications, we focus on the endogeneity of the uncertainties by constructing the DDUS and do not go into the details of problem formulation. For the last application, predisaster network investment, detailed problem statements and numerical case studies are provided with an out-of-sample analysis.

    A. Application 1: Batch Scheduling Problem

    This case is from [13] (Section 8) where stochastic programming formulations with endogenous uncertainties are established. Here we reform the characterization of DDUs to fall in the scope of RO. Consider a chemical process network as shown in Fig. 1.

    Fig. 1. Illustration of Application 1.

    Here, chemical A is produced in Process 3 from chemical B while chemical B can be produced in Process 2 from chemical D or produced in Process 1 from chemical C. If needed,chemicals A, B, C, and D can be purchased from market. Now there is a demand for chemical A that must be satisfied and decisions on which specific processes should be operated are made to maximize the net profit. The per unit yields of the three processes, denoted by θi,tfor each processi∈I={1,2,3} and timet∈T, are uncertain. The endogeneity stems from the fact that θi,tretains a physical meaning only when processiis operated at time periodt. Letxi,t∈{0,1}denote the binary decision of operating processiat timet,then the DDUS forθis constructed as

    B. Application 2: Shortest Path Over an Uncertain Network

    The second case comes from [16] (Example 2), considering to find the shortest path from the origin node A to the destination node B over a network with uncertain arc lengths,as shown in Fig. 2.

    Fig. 2. Illustration of Application 2.

    Let E denote the arcset of the graph and the uncertain length for an arce∈E is denoted byde. The uncertain arc lengthsdlie in the DDUS as follows:

    C. Application 3: Frequency Reserve Provision

    The third application comes from [8]. Consider a smart building participating in the demand response program on electricity consumption. The building is asked by the grid operator to adjust its electricity consumption within a prespecified reserve capacity. The building is rewarded for providing this reserve service and the payment depends on both the size of the reserve capacity and the actually deployed reserve energy.

    Consider a symmetric reserve capacityRt(Xt)=[?Xt,Xt]whereXt≥0 is the size of reserve capacity at timet. For the smart building, the operator’s real-time reserve deployment requestrtis endogenously uncertain at the planning stage. The endogeneity stems from the fact that the range of reserve deployment that the smart building admits is fundamentally affected by its decisions on the reserve capacity. Specifically,rt∈Rt(Xt) andXtis the building’s day-ahead decision. It is assumed that providing reserve capacity of sizeXtis rewarded byXtandofferingreserveenergyofsizertis compensatedbyrt.Thefrequencyreserveprovision problem is formulated to maximize the building’s profit while always satisfying inner technical and comfort constraints.

    D. Application 4: Pre-Disaster Network Investment

    In this subsection, a pre-disaster highway network investment problem is provided, including its problem formulation and a case study. This application is motivated by the work of [28] where the decision maker aims to strengthen the highway system in advance to prevent link failures due to earthquakes. By contrast, we model the problem into a twostage RO with DDUs for the case that link failure probabilities are not known.

    The goal is to guarantee the existence of a path between a given origin-destination (O-D) pair after the earthquake and concurrently minimize the post-disaster travel cost from the origin node to the destination node and the costs of necessary pre-disaster investment. Endogenous uncertainties arise from the fact that the post-disaster state of each link, either functional or non-functional, is uncertain but can be altered by pre-disaster reinforcement.

    1)Problem Formulation:The high-way network is modeled into a directed graphN=(V,E) with node setVand arc setE.To characterize the functionality of the highway system after the disaster, two nodes in the graphNare specified:Ddenotes the destination node which is the district with the highest expected damage in the earthquake, andOdenotes the origin node which usually refers to the district with the most support resources.

    We use binary-valued variablexe∈{0,1}1to denote the investment decision on linke∈E.xetakes the values of 1 if thereisaninvestmentonlinkeand0otherwise.Letwe∈{0,1}1denotethedecision-dependentuncertainpostdisaster state of linke∈E.weattains the value of 1 if the linkeis not functional after the occurrence of the disaster and the value of 0 otherwise. Strengthened links have guaranteed functionality after the disaster and the non-strengthened links are subject to random failures. Thus the realization ofweis restricted by

    where ψ ∈[0,1] is the robustness budget to reduce conservativeness. Note that the constraint matrix of (24)satisfies total unimodularity, thus the binary-valuedwecan be relaxed to be continuous without any compromise [3]. Thus,the budgeted DDUS forweis constructed as

    To verify the post-disaster connectivity fromOtoDas well as find the path fromOtoDwith minimal traversal costs after the disaster, the following second-stage optimization problem is considered:

    whereY(w)?R|E|is defined through the following constraints:

    andce∈R1denotes the length of linke. The objective is to minimize the traversal cost fromOtoDin the network. The wait-and-see decision variableye∈{0,1}1takes the values of 1 if the path fromOtoDgoes through linkeand 0 otherwise.Note that in (27), the binary-valuedyeis substituted by its continuous relaxation (27a) since the constraint matrix ofysatisfies total unimodularity property. There exists a path fromOtoDafter the occurrence of the disaster if and only if (26)has at least one feasible solution and the optimal solution to(26) implies the path with least traversal cost fromOtoDin the network.

    Next, the mathematical formulation of the two-stage robust pre-disaster highway network investment problem is given

    whereae∈R1denotes the investment cost of linke.W(x) andY(w)are defined in (25) and (27), respectively. The objective is to minimize the traversal costs between the O-D pair under the worst-case uncertainty realization as well as the necessary investment cost. Constraint (28c) guarantees the existence of at least one path fromOtoDafter the disaster under any realization ofwin the uncertainty setW(x). The proposed modified Benders dual decomposition algorithm is applied to solve problem(∑28).Notethat givenavariableλ∈R1, thebilinear item λ?ψe∈E(1?xe)」 canbepreciselylinearized,thus the robust counterpart of the master problem in Algorithm 1 can be formulated into an MILP problem.

    2)Basic Results:The computational results are based on a network with 8 nodes and 9 links, as demonstrated in Fig. 3.The relevant data are from [28]. The traversal costs (lengths)and investment costs of links are provided in Table I. The O-D pair is chosen as (1,6), i.e.,O=1,D=6. For the nominal 9-link network, there are 4 possible paths fromOtoD, as listed in Table II. If all links remain functional, the shortest path fromOtoDis Path 1 (1-3-5-9) with a length of 13.52. The program runs on an Intel Core-i5 1.6-GHz computer and is coded with YALMIP. CPLEX 12.6.0 is utilized as the solver.

    The proposed modified Benders dual decomposition algorithm is applied with an initialization ofUB0=4000,LB0=0, ε=0.01, ψ=0.3. The algorithm reaches convergence ofUB=LB=1100.65 after 8 rounds of iterations and the evolution process ofUBkandLBkis shown in Fig. 4. The optimal robust investment (the here-and-now decision) is on Link 3, Link 8, and Link 9. For all possible network realizations under this pre-disaster investment scheme, the worst-case is the failure on Link 5 after the earthquake. In that case, the shortest post-disaster path fromOtoD(the waitand-see decision) is Path 3 with a length of 20.65. The correctness of the computational results can be verified by enumeration.

    Fig. 3. The 8-node 9-link network.

    Link e 1 2 3 4 5 6 7 8 9 ce 6.41 8.09 1.97 6.35 2.87 4.11 2.27 3.91 2.27 ae 500 620 160 780 260 220 500 120 800

    TABLE II PATHS FOR O-D PAIR (1,6)

    Fig. 4. Evolution of the U Bk and L Bk .

    3)Comparative Study:To emphasize the necessity and superiority of the proposed algorithm for DDU-involved twostage RO problems, existing cutting plane algorithms that are theoretically applicable to only the case of DIUs are adopted to solve the same problem in (28). In both the classical Benders decomposition algorithm and the C&CG algorithm,the master problem becomes infeasible at the second round,then the solution procedure interrupts without convergence.This is because the worst-case uncertainty realization identified on the first round is a failure in Link 9, under which there is no path fromOtoD. This situation can be eliminated by strengthening Link 9; however, the invalid feasibility cutgenerated with this concrete scenario is appended to the master problem and the consequent infeasibility falsely implies that there is no feasible solution to the original problem (28). The observations in this comparative case study verify the statements in Remark 2.

    TABLE III IMPACT OF ROBUSTNESS PARAMETER ψ

    4)Sensitivity Analysis:We present the impact of robustness budgetψo(hù)n the optimal solution to problem (28). In this case,7 robustness parametersψfrom 0 to 0.6 with a gradient of 0.1 are introduced and the results are displayed in Table III. It is observed that, asψincreases, the optimal objective value including investment costs and worst-case traversal costs increases accordingly. This is because the decision maker has to enhance more links to hedge against the increasing uncertainties on link failures. However, the optimal predisaster investment saturates whenψis greater than or equal to 0.5. This is due to the fact that no matter how largeψis,there always exists a path between the O-D pair if links 1,3,5,9 (Path 1) are all reinforced.

    5)Out-of-Sample Analysis:Finally, an out-of-sample analysis is carried out to verify the performance of the robust investment decision. Steps of the out-of-sample assessment areasfollows [29]:a)Consideringallpossible statesofthe 9 links,wegenerate29=512scenarios, namelywo∈{0,1}9,o∈[512]. b) Givenx?derived by solving (28), for each scenarioo, the following relaxed second-stage problem is solved:

    where non-negative slack variablesso+,so?are introduced to the flow conservation constraints (29d)–(29f) and penalized in the objective (29a). The penalty cost coefficientP∈R1is set as 5000 in this case. Problem (29) is a modified version of the second-stage problem (26) with a given here-and-now decisionx?andthegiven uncertaintyrealizationwo. Denote the solutionto(29)byyo?,so+?,so??.c)Atlast,the average sampled second-stage costCavand the average sampled infeasibility levelsavare computed as follows:

    Table IV shows the out-of-sample assessment results for the robust optimal solution under different values of the robustness budgetψ. It is observed thatCavandsavdecrease with the increasingψ, indicating that a biggerψleads to a more robust solution that can hedge against a higher level of uncertainties, but also accordingly generates higher investment cost in the first stage. Comparison with the deterministic model is also provided in Table IV . We can see that disregarding uncertainty would give rise to significantly high second-stage costs (due to the penalty on the slack variablesand)andinfeasibilitylevel.By choosingaproper r obustnessbudgetψ,the decisionmaker can achievethetradeoff between the first-stage investment costs and the secondstage feasibility level.

    TABLE IV RESULTS FROM THE OUT-OF-SAMPLE ANALYSIS

    VI. CONCLUSIONS

    In this paper, a novel two-stage RO model with DDUs is proposed. We introduce a class of polyhedral DDUSs whose right-hand-side vectors are in a dependency of the here-andnow decision. Solution methodology for the problem is designed based on modified Benders decomposition, robust counterpart derivation, and linearization techniques. Computational tractability and convergence performance of the proposed algorithm is guaranteed by a strict proof.

    Case studies on the pre-disaster highway network investment problem verify that the proposed DDU-incorporated two-stage RO model is an amenable framework for addressing decision-making problems under endogenous uncertainties.The optimality and feasibility of the robust solution are validated by enumeration in this case. Furthermore, the computational studies elucidate that the DDU-involved twostage RO model inherently captures the trade-off between uncertainty mitigation (line investment) and the corresponding expenses (line investment costs) and provides less conservative robust results.

    Though we focus on polyhedral DDUS whose right-handside vector has a general correlation with the here-and-now decision, the proposed solution framework based on the modified Benders decomposition in Section III is also applicable to DDUSs with ellipsoidal or conic structure. The limitation of the work in this paper lies in the fact that the duality-based Benders decomposition is conditional upon the second-stage problem which is an LP. Also, uncertain parameters that are of discrete or binary nature would prohibit us from formulating the robust counterpart of the master problem as described in Section IV. These issues remain to be addressed in our future study. In addition, as an extension of this work, future study could combine the optimality and feasibility cuts (9c) and (9d) with the so-called Pareto cuts[30], or extend the robust counterpart (17) to incorporate primal cuts with recourse decision variables in the primal space, to further improve the convergence rate in practice.

    APPENDIX

    Before we give the proof of Theorem 1, a crucial lemma is provided as follows.

    Lemma 2:Letx?denote the optimal solution to TRO-DDU problem (1). For anyk∈Z+:

    Since the TRO-DDU problem (1), problem (8) and problem(31) are equivalent,f(x?)+S(x?) is also the optimal objective value of (31). Also note that the master problem (9) is a relaxation to minimization problem (31) since∈Ufor allj∈[ku] and∈Π for allj∈[kπ]. Thus, the optimal objective value of problem (9), which is denoted byLBk, is less than or equaltotheoptimal objective value of problem (31), i.e.,LBk≤f(x?)+S(x?).

    Recall the updating rule of the upper bound which isUBk=UBk?1whenR(xk)>0 andUBk=f(xk)+S(xk) whenR(xk)=0 . Next, we proveUBk≥f(x?)+S(x?) by induction.First of all,UB0=M>f(x?)+S(x?). Suppose for the sake of induction thatUBk?1≥f(x?)+S(x?) wherek∈Z+. Then, ifUBk=UBk?1, we haveUBk≥f(x?)+S(x?) directly. Else ifUBk=f(xk)+S(xk) whenR(xk)=0, then we haveUBk=f(xk)+S(xk)≥f(x?)+S(x?) sincexkis a feasible solution to the TRO-DDU problem (1) (by recalling thatxk∈XRif and only ifR(xk)=0 ) andf(x?)+S(x?) is the optimal objective value of TRO-DDU problem (1).

    Assertion b):∈vert(Π) canbeeasily verified bynoting thattheoptimalsolutionofabi-linear programwith polyhedral feasible region can be achieved at one of the vertices of the polytopes [31].

    which contradicts with (33).

    Assertion c):Similarly to the proof of Lemma 2–b),∈vert(U)can be verified by noting that the optimal solution of bi-linear programming over a polytope can be achieved at one of its vertices.Suppose there existj1,j2∈[ku] andj1≠j2such that=. We assume thatis the optimum toS(xk j1) andis the optimum toS(xkj2). Without loss of generality, we assume thatj1

    then, we have

    which is equivalent to

    Recalling the updating rule of lower bound, we have

    wherei) comes from the relationship in (38),ii) comes from the fact that=,iii) comes from the fact thatis the optimum toS(xk j2), andiv) comes from the updating rule of the upper bound. Thus, we haveUBk j2 ≤LBk j2. According to Lemma 2-a),LBk j2 ≤f(x?)+S(x?)≤UBk j2. Thus, we haveLBk j2 =UBk j2 =f(x?)+S(x?)and the Algorithm 1 terminates at iterationkj2. This completes the proof of the statement in Lemma 2-c). ■

    The proof of Theorem 1 is given as follows.

    Proof of Theorem 1:First, we prove that Algorithm 1 terminates within a finite round of iterations through a complete enumeration of the vertices of the polyhedral feasible regions of the dual multipliers. According to Lemmas 2-b) and 2-c), no vertex of Π orUwill be appended twice to the master problem. Thus, the termination condition must be met within O (p+q) iterations.

    Next, we prove the feasibilityofsolutionxk.Sincexkis ge nerated through master problem(9),xk∈X. Moreover,since Algorithm 1 terminates withR(xk)=0 , wehavexk∈XRbyrecallingthatx∈XRif and only ifR(x)=0. Thus,xk∈X∩XR.

    Finally, we prove the optimality of solutionxk. According to Lemma 2-a),LBk≤f(x?)+S(x?)≤UBk. Together with the fact that the Algorithm 1 terminates with |UBk?LBk|<ε andR(xk)=0 , we have |UBk?f(x?)?S(x?)|<ε. Recalling thatUBk=f(xk)+S(xk), we have|f(xk)+S(xk)?f(x?)?S(x?)|<ε, which justifies the optimality ofxk.■

    欧美精品人与动牲交sv欧美| 中国三级夫妇交换| 99九九线精品视频在线观看视频| 国产69精品久久久久777片| 青春草视频在线免费观看| 女人久久www免费人成看片| 国产精品国产三级专区第一集| 最近最新中文字幕免费大全7| 免费观看无遮挡的男女| 国产v大片淫在线免费观看| 欧美高清成人免费视频www| 肉色欧美久久久久久久蜜桃| 亚洲自偷自拍三级| 一区二区三区免费毛片| 国产深夜福利视频在线观看| 欧美日韩综合久久久久久| 亚洲美女黄色视频免费看| 亚洲精品国产色婷婷电影| 亚洲精品日本国产第一区| 亚洲精品一区蜜桃| 日本免费在线观看一区| av又黄又爽大尺度在线免费看| 能在线免费看毛片的网站| 精品99又大又爽又粗少妇毛片| 国产伦在线观看视频一区| 亚洲av中文av极速乱| 多毛熟女@视频| 国产片特级美女逼逼视频| 最近最新中文字幕大全电影3| 国产黄片美女视频| 成人国产av品久久久| 五月伊人婷婷丁香| 91精品伊人久久大香线蕉| 免费看日本二区| 久久亚洲国产成人精品v| 久久热精品热| 成人特级av手机在线观看| 乱系列少妇在线播放| 久久婷婷青草| freevideosex欧美| 久久久精品94久久精品| 成年免费大片在线观看| 国产欧美亚洲国产| 少妇人妻久久综合中文| 久久国产乱子免费精品| 精品一区在线观看国产| 热re99久久精品国产66热6| 国产精品不卡视频一区二区| 一区二区三区精品91| 男女国产视频网站| 水蜜桃什么品种好| 91狼人影院| 2022亚洲国产成人精品| 日本av手机在线免费观看| 色视频www国产| 亚洲国产精品999| 亚洲真实伦在线观看| 成人影院久久| 亚洲av不卡在线观看| 国产一区二区三区综合在线观看 | 国产美女午夜福利| 少妇人妻精品综合一区二区| 一级片'在线观看视频| 中国三级夫妇交换| 五月玫瑰六月丁香| 在线观看三级黄色| 丰满迷人的少妇在线观看| 亚洲aⅴ乱码一区二区在线播放| 亚洲精品国产av成人精品| 亚洲国产精品成人久久小说| 婷婷色av中文字幕| 伊人久久国产一区二区| 国产淫语在线视频| 欧美老熟妇乱子伦牲交| 欧美区成人在线视频| av在线观看视频网站免费| 最近最新中文字幕免费大全7| 成人18禁高潮啪啪吃奶动态图 | 熟妇人妻不卡中文字幕| 亚洲av二区三区四区| 草草在线视频免费看| 亚洲成人手机| 欧美一级a爱片免费观看看| 亚洲av日韩在线播放| 国产国拍精品亚洲av在线观看| 亚洲精品国产av成人精品| 亚洲第一av免费看| 天天躁夜夜躁狠狠久久av| 免费av中文字幕在线| 麻豆乱淫一区二区| 91精品一卡2卡3卡4卡| 国产黄频视频在线观看| 色吧在线观看| 亚洲av.av天堂| 国产伦在线观看视频一区| 成人综合一区亚洲| 成人美女网站在线观看视频| 大又大粗又爽又黄少妇毛片口| 蜜桃在线观看..| 99精国产麻豆久久婷婷| freevideosex欧美| 亚洲综合色惰| 在线观看国产h片| 久久午夜福利片| 亚洲精品第二区| 久久久久久久国产电影| 成人一区二区视频在线观看| 亚洲国产日韩一区二区| 久久久久久九九精品二区国产| 一区二区av电影网| 一级毛片电影观看| 麻豆国产97在线/欧美| 国产av码专区亚洲av| 永久免费av网站大全| 草草在线视频免费看| 国产成人精品婷婷| a级毛片免费高清观看在线播放| 乱系列少妇在线播放| 日日摸夜夜添夜夜添av毛片| 久久国产乱子免费精品| 精华霜和精华液先用哪个| 国产伦精品一区二区三区四那| 少妇的逼水好多| 高清黄色对白视频在线免费看 | 我的老师免费观看完整版| 我的女老师完整版在线观看| 观看美女的网站| 国产成人aa在线观看| 久久久久久久亚洲中文字幕| 久久av网站| 国产乱人视频| 欧美精品国产亚洲| 一区二区三区乱码不卡18| 极品少妇高潮喷水抽搐| 国产av一区二区精品久久 | 自拍偷自拍亚洲精品老妇| 国产视频首页在线观看| 久久精品夜色国产| 日韩电影二区| 尾随美女入室| 欧美日韩一区二区视频在线观看视频在线| 日韩不卡一区二区三区视频在线| 在线观看美女被高潮喷水网站| 男人和女人高潮做爰伦理| 狠狠精品人妻久久久久久综合| 日日啪夜夜爽| 久久毛片免费看一区二区三区| h日本视频在线播放| 欧美日韩精品成人综合77777| xxx大片免费视频| 免费观看无遮挡的男女| 男人狂女人下面高潮的视频| 国产日韩欧美亚洲二区| 熟女人妻精品中文字幕| 欧美97在线视频| 九九在线视频观看精品| 日韩不卡一区二区三区视频在线| 欧美日韩国产mv在线观看视频 | 日韩强制内射视频| 亚洲伊人久久精品综合| 一边亲一边摸免费视频| 一级毛片我不卡| 韩国av在线不卡| 国产精品国产av在线观看| 身体一侧抽搐| 精品国产三级普通话版| 97在线视频观看| 久久韩国三级中文字幕| 国产中年淑女户外野战色| 国产淫语在线视频| 成人毛片60女人毛片免费| 国产精品国产三级国产av玫瑰| 在线观看美女被高潮喷水网站| 中文乱码字字幕精品一区二区三区| 国产深夜福利视频在线观看| 国产在线视频一区二区| 日本av免费视频播放| 久久精品国产a三级三级三级| 男人添女人高潮全过程视频| 大话2 男鬼变身卡| 亚洲美女视频黄频| 国产av国产精品国产| 99热这里只有是精品在线观看| 国产日韩欧美在线精品| av福利片在线观看| 五月伊人婷婷丁香| 一级二级三级毛片免费看| 熟女人妻精品中文字幕| 亚洲最大成人中文| 亚洲国产欧美人成| 国产日韩欧美亚洲二区| 欧美日韩在线观看h| 18禁裸乳无遮挡免费网站照片| 少妇的逼水好多| 亚洲怡红院男人天堂| 下体分泌物呈黄色| 美女脱内裤让男人舔精品视频| 国产精品爽爽va在线观看网站| 内地一区二区视频在线| 国产黄色免费在线视频| 色婷婷久久久亚洲欧美| av天堂中文字幕网| 国产黄片视频在线免费观看| 久久久精品免费免费高清| 久久精品国产自在天天线| 亚洲国产色片| 欧美区成人在线视频| 久久影院123| 涩涩av久久男人的天堂| 国产午夜精品久久久久久一区二区三区| 欧美xxxx黑人xx丫x性爽| 亚洲人成网站在线观看播放| 国产精品人妻久久久影院| 秋霞在线观看毛片| 亚洲av二区三区四区| 国产视频首页在线观看| 精品亚洲乱码少妇综合久久| 免费观看a级毛片全部| 婷婷色综合www| 色5月婷婷丁香| 亚洲一级一片aⅴ在线观看| 麻豆精品久久久久久蜜桃| 在线天堂最新版资源| 久久综合国产亚洲精品| 在线观看一区二区三区激情| 欧美bdsm另类| 建设人人有责人人尽责人人享有的 | 欧美日本视频| 国产亚洲欧美精品永久| 高清在线视频一区二区三区| 成人黄色视频免费在线看| av不卡在线播放| 亚洲国产日韩一区二区| 欧美性感艳星| 亚洲三级黄色毛片| 亚洲国产精品专区欧美| 女的被弄到高潮叫床怎么办| 国产黄频视频在线观看| 亚洲人与动物交配视频| 寂寞人妻少妇视频99o| 亚洲色图综合在线观看| 亚洲国产精品999| av又黄又爽大尺度在线免费看| 女的被弄到高潮叫床怎么办| 亚洲欧美精品专区久久| 国产中年淑女户外野战色| 人体艺术视频欧美日本| 少妇猛男粗大的猛烈进出视频| 国产在线男女| 少妇的逼好多水| 中文字幕免费在线视频6| av福利片在线观看| 国产精品一及| 在线亚洲精品国产二区图片欧美 | videos熟女内射| av播播在线观看一区| 日韩亚洲欧美综合| 精品国产露脸久久av麻豆| 黄色怎么调成土黄色| 国产淫语在线视频| 大码成人一级视频| 一个人免费看片子| 日日撸夜夜添| 欧美高清性xxxxhd video| 亚洲一区二区三区欧美精品| 国产淫片久久久久久久久| 亚洲色图av天堂| 国产高清国产精品国产三级 | 搡老乐熟女国产| 丝袜喷水一区| 精品久久久久久电影网| 国产91av在线免费观看| 国产黄频视频在线观看| 51国产日韩欧美| 高清午夜精品一区二区三区| 久久午夜福利片| 国产伦在线观看视频一区| 高清av免费在线| 又爽又黄a免费视频| 成人黄色视频免费在线看| 少妇 在线观看| 免费观看av网站的网址| 午夜老司机福利剧场| 不卡视频在线观看欧美| 狂野欧美激情性xxxx在线观看| videos熟女内射| 夜夜看夜夜爽夜夜摸| 久久精品久久精品一区二区三区| 久久久久久久久久成人| 欧美一区二区亚洲| 色5月婷婷丁香| 多毛熟女@视频| 久久久久久久亚洲中文字幕| 亚洲伊人久久精品综合| 人妻系列 视频| 欧美精品一区二区大全| 免费观看的影片在线观看| 亚洲中文av在线| 偷拍熟女少妇极品色| 国产淫语在线视频| 一级片'在线观看视频| videos熟女内射| 久久精品久久久久久噜噜老黄| 99热全是精品| 汤姆久久久久久久影院中文字幕| 亚洲欧美日韩另类电影网站 | 欧美精品一区二区免费开放| 色婷婷久久久亚洲欧美| 网址你懂的国产日韩在线| 日韩成人av中文字幕在线观看| 久久久精品免费免费高清| 菩萨蛮人人尽说江南好唐韦庄| 91精品国产九色| 亚洲精品久久久久久婷婷小说| av女优亚洲男人天堂| 国产精品久久久久久久电影| 大片免费播放器 马上看| 欧美激情国产日韩精品一区| 超碰av人人做人人爽久久| 久热这里只有精品99| 97超视频在线观看视频| 涩涩av久久男人的天堂| 我的女老师完整版在线观看| 在线观看人妻少妇| 精品久久久久久电影网| av在线app专区| 搡女人真爽免费视频火全软件| 婷婷色av中文字幕| 免费观看a级毛片全部| 国产精品一二三区在线看| 男女边摸边吃奶| 久久久久网色| 1000部很黄的大片| 丝袜脚勾引网站| 免费观看在线日韩| 日韩亚洲欧美综合| 日韩欧美 国产精品| 婷婷色综合www| 欧美日韩综合久久久久久| 午夜福利视频精品| 国产成人精品福利久久| 日日摸夜夜添夜夜爱| 免费看光身美女| 午夜免费鲁丝| 人妻制服诱惑在线中文字幕| 亚洲aⅴ乱码一区二区在线播放| 国产探花极品一区二区| 久久99热这里只频精品6学生| 久久av网站| 99九九线精品视频在线观看视频| 人妻一区二区av| 成人特级av手机在线观看| 精品一品国产午夜福利视频| 亚洲国产欧美人成| 蜜桃在线观看..| 伦理电影免费视频| kizo精华| 18禁在线播放成人免费| 色5月婷婷丁香| 欧美性感艳星| 国产一区二区三区综合在线观看 | 人体艺术视频欧美日本| 我要看黄色一级片免费的| 国产av码专区亚洲av| 我要看日韩黄色一级片| 老熟女久久久| 免费人妻精品一区二区三区视频| 国内精品宾馆在线| 亚洲国产精品999| 少妇人妻 视频| 水蜜桃什么品种好| 亚洲经典国产精华液单| av一本久久久久| 97在线人人人人妻| 日日啪夜夜撸| 中文字幕精品免费在线观看视频 | 2022亚洲国产成人精品| 能在线免费看毛片的网站| 日日摸夜夜添夜夜添av毛片| 丰满人妻一区二区三区视频av| 青春草国产在线视频| 熟女电影av网| av在线播放精品| 亚洲人成网站在线观看播放| 搡老乐熟女国产| 亚洲无线观看免费| 久久97久久精品| 久久鲁丝午夜福利片| 18禁在线播放成人免费| 男女无遮挡免费网站观看| 日本av免费视频播放| 国产精品嫩草影院av在线观看| 国产成人精品福利久久| 国产精品精品国产色婷婷| 成人一区二区视频在线观看| 日韩大片免费观看网站| 免费黄网站久久成人精品| 成人毛片a级毛片在线播放| 国精品久久久久久国模美| 午夜免费观看性视频| 99久久精品热视频| 色视频www国产| 天堂中文最新版在线下载| 日韩成人av中文字幕在线观看| 欧美97在线视频| 国产精品秋霞免费鲁丝片| 日本黄色片子视频| 国产免费一级a男人的天堂| 卡戴珊不雅视频在线播放| 蜜桃在线观看..| 黑人高潮一二区| 精品99又大又爽又粗少妇毛片| 日韩一本色道免费dvd| 一级毛片我不卡| 国产男女超爽视频在线观看| 亚洲四区av| av网站免费在线观看视频| 免费黄色在线免费观看| 哪个播放器可以免费观看大片| 国产爽快片一区二区三区| 美女国产视频在线观看| 久久久a久久爽久久v久久| 国产亚洲91精品色在线| av视频免费观看在线观看| 成人18禁高潮啪啪吃奶动态图 | 建设人人有责人人尽责人人享有的 | 国产视频首页在线观看| 国产成人freesex在线| 一二三四中文在线观看免费高清| 高清午夜精品一区二区三区| 日韩 亚洲 欧美在线| 中文字幕av成人在线电影| 少妇熟女欧美另类| 欧美bdsm另类| 久久久久久久亚洲中文字幕| 91久久精品国产一区二区三区| 久久久欧美国产精品| 日韩成人av中文字幕在线观看| 国产精品.久久久| 插阴视频在线观看视频| 国产女主播在线喷水免费视频网站| 狂野欧美激情性bbbbbb| av免费观看日本| 精品亚洲乱码少妇综合久久| 亚洲内射少妇av| 少妇的逼水好多| 人人妻人人添人人爽欧美一区卜 | 中国国产av一级| 国产在线视频一区二区| 插阴视频在线观看视频| 日韩一本色道免费dvd| 国产精品一二三区在线看| 久久精品国产自在天天线| 乱码一卡2卡4卡精品| 一级a做视频免费观看| 国产精品不卡视频一区二区| 卡戴珊不雅视频在线播放| 欧美区成人在线视频| 一区二区三区乱码不卡18| 国产av码专区亚洲av| av女优亚洲男人天堂| 一区二区三区精品91| 色视频www国产| 亚洲美女视频黄频| 身体一侧抽搐| 色婷婷久久久亚洲欧美| 日韩一本色道免费dvd| 婷婷色av中文字幕| 亚洲av.av天堂| 久久婷婷青草| 久久青草综合色| 国产av码专区亚洲av| 精品人妻熟女av久视频| 欧美日韩视频精品一区| 欧美极品一区二区三区四区| 亚州av有码| 免费看不卡的av| 青青草视频在线视频观看| 成人高潮视频无遮挡免费网站| 国产免费一区二区三区四区乱码| 我的老师免费观看完整版| 建设人人有责人人尽责人人享有的 | 久久人人爽av亚洲精品天堂 | 精品一区二区三区视频在线| av在线观看视频网站免费| 欧美高清性xxxxhd video| 在线观看三级黄色| 欧美激情国产日韩精品一区| 亚洲va在线va天堂va国产| 午夜福利高清视频| 久热久热在线精品观看| 日韩中文字幕视频在线看片 | 国产精品久久久久久av不卡| 永久免费av网站大全| 亚洲av综合色区一区| 国产成人免费观看mmmm| 80岁老熟妇乱子伦牲交| 日日摸夜夜添夜夜爱| 亚洲欧美精品自产自拍| 欧美3d第一页| 国产成人免费观看mmmm| 国产av一区二区精品久久 | 91精品一卡2卡3卡4卡| 久久国产亚洲av麻豆专区| 国产精品一区二区性色av| 欧美3d第一页| 18禁裸乳无遮挡动漫免费视频| kizo精华| 能在线免费看毛片的网站| 91午夜精品亚洲一区二区三区| 国产欧美日韩一区二区三区在线 | 亚洲欧美成人综合另类久久久| 精品一区二区三卡| 久久精品久久精品一区二区三区| 国产成人精品久久久久久| 欧美3d第一页| 国产欧美另类精品又又久久亚洲欧美| 色5月婷婷丁香| 国语对白做爰xxxⅹ性视频网站| 天天躁夜夜躁狠狠久久av| 欧美丝袜亚洲另类| 美女国产视频在线观看| 亚洲精品一二三| 国模一区二区三区四区视频| 婷婷色麻豆天堂久久| 成年av动漫网址| 日韩三级伦理在线观看| 国产毛片在线视频| 在线观看免费高清a一片| 亚洲av欧美aⅴ国产| 美女视频免费永久观看网站| 亚洲精品乱久久久久久| 国产精品99久久久久久久久| 2021少妇久久久久久久久久久| 中文欧美无线码| 又黄又爽又刺激的免费视频.| 丰满迷人的少妇在线观看| 人体艺术视频欧美日本| 99热这里只有是精品在线观看| 日日撸夜夜添| 亚洲性久久影院| 18禁裸乳无遮挡免费网站照片| 汤姆久久久久久久影院中文字幕| 中文字幕人妻熟人妻熟丝袜美| 国产精品爽爽va在线观看网站| 久久国产精品大桥未久av | 久久这里有精品视频免费| 国产成人免费无遮挡视频| 国产爱豆传媒在线观看| 丝瓜视频免费看黄片| 男女国产视频网站| 黑丝袜美女国产一区| 亚洲欧美日韩另类电影网站 | 亚洲av不卡在线观看| 九色成人免费人妻av| 成人免费观看视频高清| 国产精品99久久99久久久不卡 | av播播在线观看一区| 亚洲国产毛片av蜜桃av| 嫩草影院新地址| 日本黄大片高清| 免费看av在线观看网站| 777米奇影视久久| 欧美激情极品国产一区二区三区 | 国产午夜精品久久久久久一区二区三区| 丝袜喷水一区| 嫩草影院新地址| 中文天堂在线官网| 直男gayav资源| 99九九线精品视频在线观看视频| 免费观看av网站的网址| 晚上一个人看的免费电影| 一级a做视频免费观看| 国产精品一区二区在线观看99| 久久精品国产亚洲网站| 亚州av有码| 黄色配什么色好看| 丝袜脚勾引网站| 久久久成人免费电影| 亚洲精品一二三| 亚洲四区av| 少妇丰满av| 欧美高清成人免费视频www| 夜夜骑夜夜射夜夜干| 日本午夜av视频| 亚洲丝袜综合中文字幕| 美女内射精品一级片tv| av播播在线观看一区| 亚洲三级黄色毛片| 大片免费播放器 马上看| 亚洲精品,欧美精品| 国产色婷婷99| 男男h啪啪无遮挡| 这个男人来自地球电影免费观看 | 91精品国产九色| 夫妻午夜视频| 在线天堂最新版资源| 极品教师在线视频| 成人特级av手机在线观看| 一级毛片黄色毛片免费观看视频| 极品教师在线视频| a级毛片免费高清观看在线播放| 国产熟女欧美一区二区| 特大巨黑吊av在线直播| av一本久久久久| 免费观看av网站的网址| 午夜精品国产一区二区电影| av一本久久久久| 永久网站在线| 亚洲精品乱码久久久v下载方式| 一级二级三级毛片免费看| 纵有疾风起免费观看全集完整版| 久久久久久久久久久丰满| 人人妻人人添人人爽欧美一区卜 | 高清欧美精品videossex|