• <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 ?

    Spatiotemporal Input Control: Leveraging Temporal Variation in Network Dynamics

    2022-04-15 04:20:16YihanLinJiaweiSunGuoqiLiGaoxiXiaoSenior
    IEEE/CAA Journal of Automatica Sinica 2022年4期

    Yihan Lin, Jiawei Sun, Guoqi Li,, Gaoxi Xiao, Senior,

    Changyun Wen, Fellow, IEEE, Lei Deng, Member, IEEE, and H. Eugene Stanley

    Abstract—The number of available control sources is a limiting factor to many network control tasks. A lack of input sources can result in compromised controllability and/or sub-optimal network performance, as noted in engineering applications such as the smart grids. The mechanism can be explained by a linear timeinvariant model, where structural controllability sets a lower bound on the number of required sources. Inspired by the ubiquity of time-varying topologies in the real world, we propose the strategy of spatiotemporal input control to overcome the source-related limit by exploiting temporal variation of the network topology. We theoretically prove that under this regime,the required number of sources can always be reduced to 2. It is further shown that the cost of control depends on two hyperparameters, the numbers of sources and intervals, in a trade-off fashion. As a demonstration, we achieve controllability over a complex network resembling the nervous system of Caenorhabditis elegans using as few as 6% of the sources predicted by a static control model. This example underlines the potential of utilizing topological variation in complex network control problems.

    I. INTRODUCTION

    NETWORKED structure is a common feature in natural and artificial systems, many of which contain complex topology and higher-order interactions among nodes. The analysis of empirical networks has inspired research fields such as community detection [1], influence maximization [2],and information dissemination, which in return boost applications like epidemic prediction [3]–[5] and recommender systems [6]. Many complex networks are also control systems of practical value [7]–[11], which inspired the study of network control, including theoretical efforts to understand network behaviors under control [12]–[14] and applicationdriven work on specific control solutions or algorithms[15]–[18]. However, part of these efforts are confounded by the complexity originating from network topology, dynamics,and sometimes mere scale [13], [19], [20]. The unprecedented network scale and diversity encountered in modern engineering call for control strategies that are not dependent on specific feature or structure but can handle network complexity in a general way.

    Many networks found in nature and human societies show highly dynamic interaction patterns [21]–[24] and are referred to as temporal networks. Among them, there are systems that function robustly under control and can offer inspiration for new control principles. It is suggested that the time-varying topology is essential to a network’s efficiency and flexibility[25]. Although the majority of control problems are studied using approximated linear systems, e.g., the linear timeinvariant (LTI) model [26], [27], there are fundamental differences between a temporal system and what a linear model can capture. An LTI system is governed by the equationx˙(t)=Ax(t)+Bu(t), which assumes slow dynamics of the network connection specified by matricesAandBand approximates them as constant matrices, whereas in the temporal case, the dynamics of these matrices can have a characteristic time scale comparable tox(t) andu(t).

    Despite the attractive properties, there is yet to be a general framework to exploit the advantages of temporal structures for network control. Here, we propose a control strategy featuring time-dependent input connection (i.e., node-source connection) and by modeling and simulation, evaluate its potential in controlling large complex networks. Specifically,we consider a model that extends from the linear network model but incorporates a piece-wise matrixBthat shows variable connectivityBkin different time periods:x˙(t)=Ax(t)+Bkuk(t)fort∈[tk?1,tk) (Fig. 1). In this model,the time intervals, the input matrices, and the control input vectors are open to optimization, which resembles the realworld cases where input signals can be directed to different sets of nodes during control [28]–[31]. Here, matrixBis assumed to evolve at a much higher rate than the adjacency networkAitself, andAis modeled as a constant. Despite the simplicity, the model captures the fast and slow dynamics of network structures and is expected to offer clues on the general principles of control with temporal input connections.

    Fig. 1. An example comparing the spatiotemporal input control with static network control. (a)–(b) Schematics illustrating an 8-node network (nodes are colored in light blue; directed edges are represented by black arrows) made controllable by 2 sources (dark blue) in 4 time intervals or by 3 sources via static input connection. (c)–(d) Optimized node trajectories of the 2 control strategies. Although the target is reached at 8.4 seconds in both cases, the spatiotemporal input gives much more compact trajectories.

    We term the proposed control strategy instantiated with this model as thespatiotemporal input control, given that it integrates the spatial and temporal patterning of the nodesource connection. It is proved that given a sufficient number of time intervals, controllability can always be achieved with 2 control sources; in many cases, 1 source is sufficient. This lifts the lower bound predicted by the LTI model, which can be computed using the maximum matching algorithm [32].The difference in controllability could be explained by the delivery of control signals to the network: in contrast to the LTI model where all nodes are under effective control [32],[33], the spatiotemporal input controls different subsets of nodes in different time intervals.

    We further optimize the input matrices and signals with a proposed semi-analytic method to generate control solutions.By simulating more than 1000 randomly generated networks,we find that the cost of control, or the control energy, drops with the number of intervals and the number of sources. The dependence of energy on intervals and sources could be summarized into a trade-off relationship, where more intervals may compensate for the lack of sources. Practically, this indicates a promising strategy to tackle large complex networks if dividing intervals can be done with a much lower cost than adding control sources. To demonstrate the potential of the method in handling complex networks, we present a test case using the network topology extracted from the nervous system ofCaenorhabditis elegans[34]–[36], and show that controllability can be achieved with fewer than 6% of the nodes computed by the maximum matching algorithm. A fast construction method is proposed to efficiently generate input connections for large networks like this.

    II. MODEL

    A. Temporal Network Model

    Consider the following piece-wise linear model with temporal input connection. Specifically, it contains a constant directed network ofNnodes controlled byMindependent sources. The time duration is split intoKintervals, i.e.,[tk?1,tk),k=1,...,K. For simplicity, we only consider intervals of equal length, giving Δt=tK/Kfor all intervals.The dynamics is determined by the following system of ordinary differential equations:

    B. Controllability

    A prerequisite for successful control is the system’s controllability [32]. Given proper input, a controllable system can access any target in the RNspace starting from any initial state. In contrast to the LTI system of which controllability is determined by the Kalman rank condition [37], networks with temporal input structure have a slightly different condition of controllability. Assuming that the control sources are homogeneous and can be connected to any node at any time,the following lemmas give the controllability condition of the model in (1).

    C. Control Energy as an Objective Function

    D. Optimal Control Problem

    III. METHODS

    A. Constructing Input Connections for Controllability

    For a classical LTI network, a graph-based method,maximum matching, gives a node-source mapping that uses the fewest control sources to ensure controllability. However,this lower bound may not hold true for temporal input connections given that a control source can deliver signals to multiple sets of nodes by switching connections across intervals. In this section, the lower bound of control sources is recalculated by constructing input matrices. The following definitions and lemmas are presented for the proof of Theorem 1.

    Algorithm 1 Construct Input Matrices If the Adjacency Matrix A Has Complex Eigenvalues{Bk}As{tk}Kk=1 Require: ,{Bk}Kk=1 Ensure:function MATRIXGENERATOR ( , )K ≥γ+「ζ As{tk}Assign 2 ■,γ= β 2 As Jt=P?1AsP Real Jordan canonical form transform of : .Use j as the row number as the Jordan block.k=1 →K kth for do k ≤「ζ if then Cbk=zeros(2,j)2■Cbk[1:2,j]=ones(2,2)Ok=zeros(2,j)Ck=[O1 ··· Cbk ··· Oχ]T end if「ζ 2■+γ ≥k>「ζ if then Cbk=zeros(1,j)2■Cbk[1,j]=1 Ok=zeros(1,j)Ck=[O1 ··· Cbk ··· Oχ]T end if k>「ζ if then Ck=O 2■+γ end if{Bk}Kk=1=P×{Ck}Kk=1 Calculate .Note: the computed should have no more than 2 columns end for end function Bk(k=1,...,K)

    Algorithm 2 Construct Input Matrices If the Adjacency Matrix A Has Only Real Eigenvalues{Bk}As{tk}K Require: ,{Bk}K k=1 Ensure:k=1 function MATRIXGENERATORR K ≥ζ(As,{tk})Assign As Js=P?1AP Real Jordan canonical form transform of : .Use j as the row number as the Jordan block.k=1 →K kth for do Cbk=zeros(1,j)Cbk[1,j]=1 Ok=zeros(1,j)if then Ck=[O1 ··· Cbk ··· Oζ]T 1 ≤k ≤ζ end if k>ζ if then Ck=0.end if end for{Bk}K k=1=P×{Ck}K Calculate .Bk(k=1,...,K)k=1 Note: the computed should have only 1 column.end function

    B. Optimization

    The functionalEin (6) is introduced both as a metric for evaluating the control input and as an objective for problem(7). In this subsection, problem (7) is solved, and in the following sections, the minimized energy is calculated under various conditions to investigate the system’s hyperparameter dependence. This is of practical interest since energy is often an analogy to real-life control effort.

    After solving (7), we seek to find a general relationship between energyE, total number of intervalsK, and total number of sourcesM. In order to compare these cases of different hyperparameters, the factors of specific control input, namely {Bk} and {uk(t)}, need to be eliminated by optimization. In other words, it is much less meaningful to compare the energy of systems that are far away from their optima. Therefore, a series of optimal control problems in the form of (7) with differentKs andMs are solved to obtain the minimized energy. Note that the case ofK=1 reduces to an LTI model.

    To start with, we have the following theorems and methods for optimization.

    Theorem 2:Givenx˙(t)=Akx(t)+Bkuk(t), the input vector minimizingEkof time interval [tk?1,tk) is given by

    Hence we have

    Proof:Plugging (8) and (10) into (6) yields,

    update , = ControlEnergy(A, , , ).^El^Γl{Bl+1 k } tkΛl?Λ =^ΓlΛl compute .?E wlk compute projected gradient using (49) and obtain unit vector .ˉwlk apply Adam algorithm to determine the step size and update .Δ=∑Λl+1=Λl ?μlkˉwlkμlk k ?Blk‖+‖Λl+1 ?Λl‖compute .‖Bl+1 end while end function k

    C. Coordinate Descent Algorithm and Simulation

    In order to combine the optimization of the input matrices{Bk}and the vector Λ that are related tox(t) and {uk(t)}, a coordinate descent algorithm is proposed based on the methods described in the last subsection to solve the optimization problem in (38), as shown in Algorithm 3.Specifically, {Bk} and Λ will be updated in an alternating fashion in each iteration. Note that this algorithm also applies to more general cases where bothAandBare piece-wise constant matrices.

    To implement the algorithm, {Bk} can be initialized using the constructive methods described in Algorithms 1 and 2 to ensure controllability. In order to avoid converging to suboptimal local minima, we launch gradient descent at multiple initial states with different step sizes and take the case with the lowest energy as an approximation of the optimal solution.As a test case, control energy is optimized for an example network given different numbers of intervals, and a significant decrease ofEwith an increasingKis observed (Fig. 2).

    To better illustrate the effect of optimization, a video is provided showing examples of randomly generated networks containing 8 nodes and 10% of all possible connections among nodes (see demo1.mp4). The dynamics of the system is visualized using 4 two-dimensional trajectories representing the state of the 8 nodes. The first and the last 4 nodes are visualized as thexandycoordinates, respectively.

    D. Control by Network Partition

    It is challenging to solve control problems of large networks. In particular, the computational complexity of the proposed Algorithms 1 to 3 scale with the network size. In addition, the split of time into several intervals also brings about the challenge of temporal complexity to the finite computational resources one can possibly utilize. In this subsection, we discuss the complexity of the proposed algorithms and present a method to circumvent this challenge while constructing and optimizing input connections.

    In Algorithm 1, the time complexity of calculating a Jordan canonical form isO(N4) [43]. Other operations in Algorithm 1 are linear except for matrix multiplication, of which the time complexity isO(N3). Regarding space complexity, the most costly operation is also calculating the Jordan canonical form,whose complexity is bounded byO(N2).

    Fig. 2. More time intervals result in lower control energy in randomly generated 8-node networks. (a) Histograms of control energy before and after optimization using the proposed coordinate descent algorithm. Networks are randomly generated and are driven from the origin to the same target state within K = 4 time intervals. Although the specific network topology has a substantial impact on the energy, the algorithm overall reduces energy by multiple orders of magnitude. (b) Histograms of optimized energy of more than 1000 randomly generated 8-node networks given different number of time intervals. The total time is divided into K = 4, 5, 7 or 8 intervals. The increase in the number of intervals significantly reduces control energy.

    Fig. 3. An example illustrating the method of control by network partition. A 16-node temporal network is split into two 8-node subnetworks (shown in green and pink), with each controlled by one input source (s1 or s2). The 2 subnetworks, with the internal edges (blank arrows), are controllable in 5 time intervals.The combined 16-node network is also controllable regardless of the edges connecting the 2 subnetworks (grey allows), which is consistent with Theorem 5.

    The time complexity of Algorithm 3 is hard to determine due to the nonlinear operations such as numerical integration and matrix exponential, which is also dependent on numerical precision. When ignoring the unpredictable iterative steps of gradient descent, the theoretical time complexity is bounded by the calculation of the energy gradient, which isO((NK)3×N×M×K)=O(N4K4M). The space complexity isO(N2K2), which is due to the storage of the gradient.

    Remark 1:Note that the above theorems only guarantee structural controllability [45], [46], which is a slightly weaker condition than controllability. However, it is practically useful for controlling a temporal network model, since the probability of having a set of weights that make a network with structural controllability uncontrollable is often 0. Thus,without solving the original optimal control problem, which can be costly, we may instead solve a series of simpler optimal control problems based on the partition of the large network, although the performance of the combined input signals and connections may vary according to the partition.Nonetheless, this indicates the possibility of controlling a large-scale network by controlling parts of the network, and the general methodology is not limited to temporal networks.

    IV. NUMERICAL EXAMPLES

    A. The Trade-Off Between Energy, Time Intervals and Input Sources

    So far, the results regarding controllability and energy in the above sections suggest two pairs of trade-offs, namely betweenKandMand betweenEandK. Here, by simulating an example network, we investigate the relations between energy and the two important hyperparameters, in search of general principles governing temporal network control. By solving the optimal control problems with various combinations ofMandK, we are able to approximately determine the minimal energy in each case, based on which we provide a visualization of the hyperparameter space by fitting a 2D surface from the discrete data points (Fig. 4).Although the results are based on a simple network model, it may explain the power of temporal input connections, which can be partly boiled down to the trade-off among energy,sources and intervals. Compared with static input connections,the spatiotemporal input introduces an extra dimension ofKto the hyper-parameter space, allowing new possibilities of reducing energy or the total number of sources.

    B. A Test Using the Neural Network of Caenorhabditis Elegans

    Complex networks, a common structure studied across fields, typically have a large size and a non-trivial topology.Here, we consider the applicability of the spatiotemporal input as a general strategy for controlling such networks.Particularly, based on the above-mentioned trade-off relations,we focus on the choice ofMandKin search of a proper balance.

    A biological network from the nervous system ofC. elegansis used as an example. Its complexity and the extensive research on its cellular connections make it a perfect test case[46]. It is known to be a sparse network of 2177 nodes and 4007 edges (multiple edges are counted as 1 if they are between the same 2 neurons). In the simulation, the network is treated as a directed graph, while the biological context is neglected. First, we demonstrate the feasibility of achieving controllability and optimizing control over a small 30-node subnetwork. Then, we show how the entire network can be controlled by a small number of control sources. The partitioning method introduced in the previous section is applied to the full network.

    Fig. 4. (a) Visualization of the trade-off relationship among energy, the number of sources, and the number of time intervals. The solid dots represent simulation results from an 8-node example network, while the circles are obtained by spline interpolation from the simulated data points due to limited computational power. (b) Projection of (a) on the M-K plane. The dashed line in the lower-left corner shows the boundary of controllability. (c) Projection of (a)on the E-K plane. (d) Projection of (a) on the E-M plane.

    Fig. 5. Optimized trajectories of a 30-node sub-network extracted from the nervous system of C. elegans. (a)–(b) Trajectories of all 30 nodes controlled by 1 or 3 sources, respectively. The right panels provide zoom-in views near the target states, to which the relative error of each node is contained within 0:1%. (c) A graph of the sub-network. The nodes are labeled by their cellular code in [47]. Each node is colored differently according to its degree.

    1) Test on a Subnetwork:A small subnetwork of 30 cells with sparse connection is extracted from the large neural network of an adultC. eleganshermaphrodite. The cells include ADELa1~ADELb5, AS01a1~AS01b5, and MDL05a1~MDL05b5, which together form 23 edges in total,with an average in-degree of 0.756 (Fig. 5(c)). According to Theorem 1, the network is controllable by 1 source. A set of input matrices {Bk} are generated that spans acrossK=11 intervals. The optimal control problem is solved, giving the trajectories in Fig. 5(a). Another set of parametersM=3 andK=4 yield a much lower energy after optimization (Fig. 5(b)).In contrast to these cases, the maximum matching algorithm predicts a lower bound of 11 sources connected to cells ADELa1, ADELa2, ADELa3, ADELa4, ADELa5, MDL5a5,AS01a1, AS01a2, AD01a3, AD01a4 and AD01a5,respectively.

    Fig. 6. Graphical demonstration of the node-node and node-source connections in the controlled neural network of C. elegans. The first and last 3 intervals out of 34 are shown here. The nodes and the sources are colored in blue and black, respectively. Each interval is of 1 second. See connection.mov in the supplementary materials for more details.

    It is noted that given 3 sources, ‖x(t)‖ is smaller by roughly an order of magnitude on overage than the 1 source case,while the overall “complexity” of control input, which can be thought of asM×K, is similar. It is possible that a non-trivial interplay exists between the numbers of sources and intervals in terms of energy scaling, which requires further studies.

    2) Test on the Entire Network:It is important to understand how nature exploits limited resources and energy for efficient manipulation of complex systems. In living matter, it is reasonable to speculate that systems deliver or respond to control signals in a time-dependent fashion with only a small number of sources being used or activated at the same time,given examples like the sparse and temporal firing patterns in cortical activities [48]–[51]. Here we show that in agreement with this intuition, the spatiotemporal input can achieve controllability over the entire network of 2177 nodes by frequently switching the connections from a limited number of sources.

    According to Fig. 4, we can balance the numbers of sources and time intervals given a certain level of energy. For small networks, this can be done by generating random instances of{Bk}until the controllability condition (3) is satisfied. Here,given a large network, we construct input connections by the method of network partition, as introduced in the previous section. For each subnetwork, the input connections are generated by Algorithm 1 or 2.

    The full network is split intoR=44 subnetworks, each of which is controllable by 1 source, givingM=R=44. This results inK=34 time intervals. The node-node and nodesource connections are illustrated in Fig. 6. Controllability is tested directly using equation (3). As a comparison, the same network requiresM≥754 nodes if controlled by static input.Therefore, the spatiotemporal input strategy saves more than 94%of the control sources. It is worth noting that there are numerous ways of partitioning the network, and the decision will likely be reflected in the control energy.

    To accelerate the simulation, we increaseRto 223 and still assumeM=R. The network is controllable withinK=5 intervals. Although we have successfully achieved controllability, the trajectories are not optimized due to the incapability of the optimization algorithm to overcome network scaling, which is a typical phenomenon for gradient descent algorithms, known as the curse of dimension. Due to the limited computational resources, it is difficult to achieve an optimal result within an acceptable amount of time.Instead, a few iterations of gradient descent are carried out for partial optimization (Fig. 7(a)). A short animation(connection.mov) is provided in the supplementary materials to show the temporal sequence of the node-node and nodesource connections for this case.

    Fig. 7. Simulated dynamics of the C. elegans neural network controlled by 223 sources in 5 intervals. The target state is randomly generated. Trajectories for 2177 nodes are plotted in different colors. A zoom-in view near the target state is provided in the right panel, underlining the accuracy of the control process.

    V. CONCLUSIONS

    This work explores the idea of using temporal input connections in complex network control. Specifically, a piecewise linear model is studied, which yields a novel control strategy termed the spatiotemporal input control. Regarding controllability, we prove that as few as 2 input sources are sufficient for an arbitrary network given sufficient intervals.Regarding the energy cost of control, we show by simulation that having more intervals can effectively reduce control energy after optimizing the input. Further simulation suggests that the advantage of temporal topology may be attributed to the trade-offs between energy, control sources and time intervals. By making the input matrix temporal, we are essentially introducing an extra dimension of intervals to the hyperparameter space, expanding the space of possible control solutions. It is worth noting that although the above properties are revealed with a relatively simple model, the trade-off relationships may generally apply to different forms of temporal topology.

    The example of theC. elegansnervous system underlines the potential of temporal input connections in controlling large networks. In real-world applications where the time intervals can be engineered, this strategy can prevent the number of input sources from scaling proportionally with the network size and thus becoming unaffordable. However, solving the optimal control problem can pose challenges on computational resources due to the time complexity and the gradient descent method. It remains an interesting problem to develop a faster algorithm for optimizing the temporal topology in general.

    In summary, the paper contributes a) conceptually by suggesting a shift from static to temporal input connections,which is tested with a piece-wise linear model; b) theoretically by proving that the lower bound of the number of input sources can be relaxed to 1 or 2, and presenting the trade-off relationship among energy, sources and intervals; c)methodologically by solving the optimal control problem for temporal networks and presenting the methods for constructing temporal input matrices. The results suggest a promising control strategy for complex networks and a possible explanation for the highly efficient control systems observed in nature. The methods proposed with our model may also be of practical value in future applications.

    APPENDIX

    And according to Lemmas 1 and 2, the controllable space of the system in (1) is

    国产精品一区二区三区四区免费观看| 成年av动漫网址| 蜜桃久久精品国产亚洲av| 可以在线观看毛片的网站| 白带黄色成豆腐渣| 欧美日本亚洲视频在线播放| 日韩强制内射视频| 国产黄色视频一区二区在线观看 | 成人一区二区视频在线观看| 国产精品1区2区在线观看.| 亚洲精品日韩av片在线观看| 日韩成人av中文字幕在线观看| 亚洲成人久久爱视频| 真实男女啪啪啪动态图| 九草在线视频观看| 国产真实伦视频高清在线观看| 久久久久久国产a免费观看| 男人和女人高潮做爰伦理| 午夜久久久久精精品| 网址你懂的国产日韩在线| 国产亚洲5aaaaa淫片| 少妇的逼好多水| 禁无遮挡网站| 亚洲三级黄色毛片| 国产极品精品免费视频能看的| 免费av毛片视频| 99久久无色码亚洲精品果冻| 99久久人妻综合| 黄色一级大片看看| 免费观看a级毛片全部| 一级二级三级毛片免费看| 一进一出抽搐动态| 大又大粗又爽又黄少妇毛片口| 午夜激情欧美在线| 男人舔奶头视频| 你懂的网址亚洲精品在线观看 | 伦精品一区二区三区| 两性午夜刺激爽爽歪歪视频在线观看| 变态另类成人亚洲欧美熟女| 亚洲自拍偷在线| 成人特级av手机在线观看| 午夜精品在线福利| 青春草视频在线免费观看| 国产高清不卡午夜福利| 久久99热这里只有精品18| 久久久精品欧美日韩精品| 国产一区二区亚洲精品在线观看| 丝袜喷水一区| 在现免费观看毛片| 国产真实乱freesex| 久久精品夜色国产| 久久精品久久久久久久性| 中出人妻视频一区二区| 国产精品一区www在线观看| 久久99热6这里只有精品| 日本三级黄在线观看| 麻豆一二三区av精品| 久久久久免费精品人妻一区二区| 亚洲乱码一区二区免费版| 黄色视频,在线免费观看| 人妻系列 视频| 日韩在线高清观看一区二区三区| 美女大奶头视频| 校园人妻丝袜中文字幕| 精品日产1卡2卡| 国内精品一区二区在线观看| 国产精品一及| 亚洲最大成人手机在线| 中文字幕av在线有码专区| 此物有八面人人有两片| 狠狠狠狠99中文字幕| 嘟嘟电影网在线观看| 久久精品国产亚洲av涩爱 | 亚洲成人久久爱视频| a级一级毛片免费在线观看| 一进一出抽搐gif免费好疼| 国产成人精品婷婷| 99热6这里只有精品| 赤兔流量卡办理| 日韩av不卡免费在线播放| 久久精品综合一区二区三区| 成年免费大片在线观看| 久久久久久久久久久免费av| 久久中文看片网| 最近手机中文字幕大全| 看片在线看免费视频| 91午夜精品亚洲一区二区三区| 变态另类丝袜制服| 少妇熟女欧美另类| 久久久久国产网址| 我的女老师完整版在线观看| 精品一区二区三区视频在线| 99久久精品一区二区三区| 日韩人妻高清精品专区| 久久99精品国语久久久| 嫩草影院新地址| 晚上一个人看的免费电影| 欧美不卡视频在线免费观看| 黑人高潮一二区| 好男人在线观看高清免费视频| 99精品在免费线老司机午夜| 久久99蜜桃精品久久| 嫩草影院新地址| 亚洲国产精品久久男人天堂| 亚洲电影在线观看av| 99热网站在线观看| www.色视频.com| 别揉我奶头 嗯啊视频| 性欧美人与动物交配| 中出人妻视频一区二区| 可以在线观看毛片的网站| 亚洲色图av天堂| 国产av在哪里看| 免费电影在线观看免费观看| 久久韩国三级中文字幕| 高清毛片免费观看视频网站| 久久人人精品亚洲av| 婷婷六月久久综合丁香| av又黄又爽大尺度在线免费看 | 高清午夜精品一区二区三区 | 熟女人妻精品中文字幕| 日本三级黄在线观看| 小说图片视频综合网站| 九九热线精品视视频播放| 国产日本99.免费观看| 伊人久久精品亚洲午夜| 国产一区亚洲一区在线观看| 日韩 亚洲 欧美在线| 欧美日本视频| 午夜a级毛片| 美女cb高潮喷水在线观看| 欧美最黄视频在线播放免费| 国产 一区精品| 日日干狠狠操夜夜爽| 久久精品国产亚洲av天美| 精品熟女少妇av免费看| 亚洲av一区综合| 亚洲不卡免费看| 午夜福利在线观看免费完整高清在 | 免费人成视频x8x8入口观看| 美女被艹到高潮喷水动态| 国产精品福利在线免费观看| 成人无遮挡网站| 成人欧美大片| 久久精品夜夜夜夜夜久久蜜豆| 亚洲欧美清纯卡通| a级一级毛片免费在线观看| 久久久午夜欧美精品| 欧美一区二区国产精品久久精品| 性欧美人与动物交配| 国产伦一二天堂av在线观看| 国产熟女欧美一区二区| 国模一区二区三区四区视频| 国产探花在线观看一区二区| 天天躁夜夜躁狠狠久久av| 国产亚洲91精品色在线| 久久精品久久久久久久性| 两性午夜刺激爽爽歪歪视频在线观看| 最新中文字幕久久久久| 我要搜黄色片| 亚洲欧美日韩高清专用| 国产精品免费一区二区三区在线| 亚洲色图av天堂| 亚洲精品日韩在线中文字幕 | 国产精品爽爽va在线观看网站| 最近2019中文字幕mv第一页| 人人妻人人看人人澡| 亚洲电影在线观看av| 插逼视频在线观看| 欧美日韩在线观看h| 26uuu在线亚洲综合色| 国产精品国产三级国产av玫瑰| 免费大片18禁| 三级毛片av免费| 免费人成在线观看视频色| 日本黄色视频三级网站网址| 久久这里只有精品中国| 韩国av在线不卡| 欧美不卡视频在线免费观看| 亚洲最大成人av| 国产精品久久电影中文字幕| 99国产极品粉嫩在线观看| 欧美成人a在线观看| 亚洲在久久综合| 日韩制服骚丝袜av| 午夜亚洲福利在线播放| 麻豆乱淫一区二区| 美女被艹到高潮喷水动态| 26uuu在线亚洲综合色| 一区二区三区免费毛片| 中国美白少妇内射xxxbb| 久久99热这里只有精品18| 久99久视频精品免费| 国产成人freesex在线| 国产精品,欧美在线| 久久精品国产亚洲网站| 午夜福利在线观看免费完整高清在 | 日韩欧美精品免费久久| 亚洲第一电影网av| 又爽又黄无遮挡网站| 黄色视频,在线免费观看| 春色校园在线视频观看| 夜夜夜夜夜久久久久| 人人妻人人看人人澡| 精品人妻偷拍中文字幕| 变态另类丝袜制服| 久久久国产成人免费| 亚洲欧美成人精品一区二区| 亚洲熟妇中文字幕五十中出| 直男gayav资源| 男女啪啪激烈高潮av片| 美女cb高潮喷水在线观看| 亚洲无线观看免费| 国产精品国产三级国产av玫瑰| 色视频www国产| 日本黄色视频三级网站网址| 免费搜索国产男女视频| 97超碰精品成人国产| 国产精品久久久久久精品电影| 中文字幕精品亚洲无线码一区| 夫妻性生交免费视频一级片| 91久久精品国产一区二区三区| 99热这里只有精品一区| 国产精品福利在线免费观看| 国产精华一区二区三区| 亚洲经典国产精华液单| 超碰av人人做人人爽久久| 国产色爽女视频免费观看| 国产 一区精品| 在线免费十八禁| 白带黄色成豆腐渣| 免费观看人在逋| 最近最新中文字幕大全电影3| 麻豆久久精品国产亚洲av| 日韩三级伦理在线观看| 国产一区二区在线av高清观看| 欧美性感艳星| 一进一出抽搐动态| 亚洲人成网站高清观看| 黄色一级大片看看| 亚洲性久久影院| 3wmmmm亚洲av在线观看| 好男人视频免费观看在线| 亚洲三级黄色毛片| 免费观看a级毛片全部| 插阴视频在线观看视频| 免费一级毛片在线播放高清视频| 国产爱豆传媒在线观看| 一级二级三级毛片免费看| 久久精品夜色国产| 中文字幕av在线有码专区| 婷婷色综合大香蕉| 欧美成人a在线观看| av在线老鸭窝| 成人高潮视频无遮挡免费网站| av国产免费在线观看| 最近中文字幕高清免费大全6| 老司机福利观看| 欧美日韩综合久久久久久| 国产男人的电影天堂91| 午夜激情福利司机影院| 内射极品少妇av片p| 国产伦精品一区二区三区视频9| 99riav亚洲国产免费| 丝袜喷水一区| 亚洲中文字幕日韩| 国产成人福利小说| 夜夜夜夜夜久久久久| 成年版毛片免费区| 亚洲av免费在线观看| 国产精品一区www在线观看| 尾随美女入室| 成人国产麻豆网| 99久国产av精品国产电影| 国产成人午夜福利电影在线观看| 免费看a级黄色片| 精品久久国产蜜桃| 久久草成人影院| 日本-黄色视频高清免费观看| 精品国产三级普通话版| 丝袜美腿在线中文| 天堂影院成人在线观看| 99久久久亚洲精品蜜臀av| 欧美一区二区国产精品久久精品| 亚洲人与动物交配视频| 欧美日韩国产亚洲二区| 国产爱豆传媒在线观看| 亚洲av成人av| 久久久久性生活片| 国内精品一区二区在线观看| 亚洲av.av天堂| 简卡轻食公司| 赤兔流量卡办理| 久久久久久久久久黄片| 亚洲美女搞黄在线观看| 欧美bdsm另类| 国产69精品久久久久777片| 日本-黄色视频高清免费观看| 麻豆国产97在线/欧美| 在线观看美女被高潮喷水网站| 菩萨蛮人人尽说江南好唐韦庄 | 亚洲成人精品中文字幕电影| 狠狠狠狠99中文字幕| 成熟少妇高潮喷水视频| 高清午夜精品一区二区三区 | 亚洲av电影不卡..在线观看| 91狼人影院| 哪个播放器可以免费观看大片| 欧美激情在线99| 别揉我奶头 嗯啊视频| .国产精品久久| 国产黄a三级三级三级人| 久久精品久久久久久久性| av福利片在线观看| 精品久久国产蜜桃| 美女被艹到高潮喷水动态| 一本—道久久a久久精品蜜桃钙片 精品乱码久久久久久99久播 | 一区二区三区高清视频在线| 久久国内精品自在自线图片| 精品不卡国产一区二区三区| 午夜福利在线观看吧| 九草在线视频观看| 狠狠狠狠99中文字幕| 伦精品一区二区三区| 亚洲最大成人av| 精品久久久久久久人妻蜜臀av| 最后的刺客免费高清国语| www.av在线官网国产| 一级二级三级毛片免费看| 国产在线男女| 国产一区二区三区在线臀色熟女| 91精品一卡2卡3卡4卡| 国产精品日韩av在线免费观看| 亚洲人成网站高清观看| 久久久久久久久久黄片| 男女那种视频在线观看| 高清日韩中文字幕在线| 尤物成人国产欧美一区二区三区| www.色视频.com| 高清毛片免费看| 精品久久久久久久久亚洲| 美女高潮的动态| 男的添女的下面高潮视频| 日韩av在线大香蕉| a级一级毛片免费在线观看| 欧美日本亚洲视频在线播放| 青春草亚洲视频在线观看| 欧美成人精品欧美一级黄| 婷婷亚洲欧美| 日韩三级伦理在线观看| 日韩一区二区三区影片| 白带黄色成豆腐渣| 国产精品日韩av在线免费观看| 精品久久久久久成人av| 一区二区三区高清视频在线| 校园人妻丝袜中文字幕| 大又大粗又爽又黄少妇毛片口| 免费av毛片视频| 免费看av在线观看网站| 人妻夜夜爽99麻豆av| 日本欧美国产在线视频| 12—13女人毛片做爰片一| 久久久久网色| 欧美+亚洲+日韩+国产| 最近2019中文字幕mv第一页| 日产精品乱码卡一卡2卡三| 最近2019中文字幕mv第一页| 亚洲精品久久久久久婷婷小说 | 两个人视频免费观看高清| 黑人高潮一二区| 久久久久网色| 欧美日韩国产亚洲二区| 日本色播在线视频| 18禁在线播放成人免费| 一卡2卡三卡四卡精品乱码亚洲| 网址你懂的国产日韩在线| 亚洲综合色惰| 人妻系列 视频| 亚洲中文字幕日韩| 美女xxoo啪啪120秒动态图| 又黄又爽又刺激的免费视频.| 国产午夜精品久久久久久一区二区三区| 免费观看的影片在线观看| 亚洲电影在线观看av| 99久国产av精品国产电影| 亚洲一区二区三区色噜噜| 国产片特级美女逼逼视频| 亚洲成人久久性| 亚洲欧美中文字幕日韩二区| 亚洲精品久久久久久婷婷小说 | 国产麻豆成人av免费视频| 又粗又爽又猛毛片免费看| 美女 人体艺术 gogo| 久久欧美精品欧美久久欧美| 久久精品国产99精品国产亚洲性色| 国产午夜精品久久久久久一区二区三区| 美女 人体艺术 gogo| 黄色日韩在线| 国产欧美日韩精品一区二区| 亚洲国产高清在线一区二区三| 亚洲内射少妇av| 又爽又黄无遮挡网站| 人妻夜夜爽99麻豆av| 美女内射精品一级片tv| 亚洲丝袜综合中文字幕| 亚洲国产欧洲综合997久久,| 国产又黄又爽又无遮挡在线| 精品人妻视频免费看| 禁无遮挡网站| 人妻制服诱惑在线中文字幕| 99热网站在线观看| 久久草成人影院| 男女啪啪激烈高潮av片| 亚洲精品国产av成人精品| 亚洲欧美日韩东京热| 国产精品爽爽va在线观看网站| 九九爱精品视频在线观看| 九九久久精品国产亚洲av麻豆| 乱码一卡2卡4卡精品| 国产男人的电影天堂91| 欧美高清成人免费视频www| 深爱激情五月婷婷| 看十八女毛片水多多多| 亚洲四区av| 边亲边吃奶的免费视频| 夜夜夜夜夜久久久久| 日韩中字成人| 久久久a久久爽久久v久久| 亚洲内射少妇av| 亚洲成av人片在线播放无| 观看美女的网站| 色综合色国产| 一本久久精品| 日韩av不卡免费在线播放| 麻豆av噜噜一区二区三区| 日韩中字成人| 观看免费一级毛片| 精品久久久久久久久av| 国产伦一二天堂av在线观看| 18禁在线无遮挡免费观看视频| 精品熟女少妇av免费看| 在线免费观看不下载黄p国产| 亚洲成人精品中文字幕电影| 国内久久婷婷六月综合欲色啪| 日韩三级伦理在线观看| 97超碰精品成人国产| 亚洲七黄色美女视频| 精品人妻一区二区三区麻豆| 国产精品国产三级国产av玫瑰| 可以在线观看毛片的网站| 我要看日韩黄色一级片| 看十八女毛片水多多多| 好男人在线观看高清免费视频| 亚洲av中文av极速乱| 久久草成人影院| 久久久成人免费电影| 天堂网av新在线| 丰满乱子伦码专区| 亚洲欧美日韩东京热| 97热精品久久久久久| 悠悠久久av| 日本撒尿小便嘘嘘汇集6| eeuss影院久久| 国产又黄又爽又无遮挡在线| 综合色av麻豆| 伦理电影大哥的女人| 国产亚洲精品久久久com| 国产三级中文精品| 一卡2卡三卡四卡精品乱码亚洲| 国产精品国产三级国产av玫瑰| 亚洲av中文字字幕乱码综合| av视频在线观看入口| 高清午夜精品一区二区三区 | 国产精品一区二区性色av| 日本撒尿小便嘘嘘汇集6| avwww免费| 日本一本二区三区精品| 黄色视频,在线免费观看| 网址你懂的国产日韩在线| 国产精品嫩草影院av在线观看| 99国产极品粉嫩在线观看| 99热这里只有精品一区| 99在线人妻在线中文字幕| 久久久久久久久久成人| 日韩高清综合在线| 久久精品人妻少妇| 免费黄网站久久成人精品| 中文精品一卡2卡3卡4更新| 国产伦理片在线播放av一区 | 一个人看视频在线观看www免费| 午夜福利在线观看吧| 亚洲第一区二区三区不卡| 精品欧美国产一区二区三| 全区人妻精品视频| 18禁在线播放成人免费| 午夜爱爱视频在线播放| 精品人妻一区二区三区麻豆| 亚洲18禁久久av| 久久久精品大字幕| 国产精品三级大全| 国产精品人妻久久久久久| 观看美女的网站| 久久久精品94久久精品| 看黄色毛片网站| 日韩强制内射视频| 亚洲无线观看免费| 一级毛片久久久久久久久女| 欧美激情国产日韩精品一区| 欧美激情在线99| 免费av不卡在线播放| 亚洲四区av| 又粗又爽又猛毛片免费看| 欧美3d第一页| 国产高清激情床上av| 久久久a久久爽久久v久久| 国产成年人精品一区二区| 国产在线男女| 综合色丁香网| 大香蕉久久网| 亚洲五月天丁香| 国产老妇女一区| 午夜福利视频1000在线观看| 小说图片视频综合网站| 亚洲精品久久国产高清桃花| 国产精品一区二区三区四区免费观看| 搡老妇女老女人老熟妇| 天堂影院成人在线观看| 亚洲av.av天堂| 欧美另类亚洲清纯唯美| 3wmmmm亚洲av在线观看| 男人狂女人下面高潮的视频| 自拍偷自拍亚洲精品老妇| 夜夜夜夜夜久久久久| 国产精品一及| 伦精品一区二区三区| 亚洲精品国产av成人精品| 日本与韩国留学比较| 老女人水多毛片| 成人漫画全彩无遮挡| 精品不卡国产一区二区三区| 久久99热6这里只有精品| 亚洲一级一片aⅴ在线观看| 91狼人影院| 免费观看的影片在线观看| 亚洲欧美日韩无卡精品| 日韩视频在线欧美| 日韩av在线大香蕉| 91午夜精品亚洲一区二区三区| 别揉我奶头 嗯啊视频| 日韩强制内射视频| 全区人妻精品视频| 久久精品国产99精品国产亚洲性色| 日韩三级伦理在线观看| 伊人久久精品亚洲午夜| av女优亚洲男人天堂| 色噜噜av男人的天堂激情| 免费人成在线观看视频色| 久久韩国三级中文字幕| 亚洲精品日韩在线中文字幕 | 搡女人真爽免费视频火全软件| 小蜜桃在线观看免费完整版高清| 亚洲精品久久国产高清桃花| 中文欧美无线码| 黄色欧美视频在线观看| 女人十人毛片免费观看3o分钟| 精品人妻偷拍中文字幕| 国产乱人偷精品视频| 国产一区二区三区在线臀色熟女| 欧美日韩一区二区视频在线观看视频在线 | 国产精品麻豆人妻色哟哟久久 | 淫秽高清视频在线观看| 变态另类成人亚洲欧美熟女| 久久精品久久久久久久性| 国产男人的电影天堂91| 亚洲色图av天堂| 免费在线观看成人毛片| 亚洲精品色激情综合| 两性午夜刺激爽爽歪歪视频在线观看| 国产精品久久久久久精品电影小说 | 国产精品免费一区二区三区在线| 久久精品国产亚洲av涩爱 | 亚洲精品成人久久久久久| 美女被艹到高潮喷水动态| 好男人视频免费观看在线| 久久99蜜桃精品久久| 日韩视频在线欧美| 嫩草影院入口| 自拍偷自拍亚洲精品老妇| 日韩视频在线欧美| 六月丁香七月| 中文欧美无线码| 一级毛片aaaaaa免费看小| 美女内射精品一级片tv| 亚洲乱码一区二区免费版| 啦啦啦韩国在线观看视频| 99久久无色码亚洲精品果冻| av在线观看视频网站免费| 成人综合一区亚洲| 国产精品久久视频播放| 黄片无遮挡物在线观看| 一级毛片aaaaaa免费看小| 97热精品久久久久久| 97人妻精品一区二区三区麻豆| 国产探花在线观看一区二区| 精品久久久久久久末码| 天天躁夜夜躁狠狠久久av| 久久99热6这里只有精品| 国产精品久久久久久亚洲av鲁大| 91久久精品国产一区二区三区| 久久精品国产99精品国产亚洲性色| 成人综合一区亚洲| 午夜亚洲福利在线播放| 身体一侧抽搐| 女的被弄到高潮叫床怎么办| 少妇的逼好多水|