• <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 ?

    Application of the edge of chaos in combinatorial optimization?

    2021-10-28 06:59:48YanqingTang唐彥卿NayueZhang張娜月PingZhu朱萍MinghuFang方明虎andGuoguangHe何國光
    Chinese Physics B 2021年10期
    關(guān)鍵詞:方明張娜國光

    Yanqing Tang(唐彥卿), Nayue Zhang(張娜月), Ping Zhu(朱萍),Minghu Fang(方明虎), and Guoguang He(何國光)

    Department of Physics,Zhejiang University,Hangzhou 310027,China

    Keywords: edge of chaos,chaotic neural networks,combinatorial optimization,travelling salesman problem

    1. Introduction

    Many problems in science,engineering and real life,such as traveling salesman problem(TSP),design of large-scale integrated circuits,[1]prediction of protein structure,[2]structural optimization of bimetallic nanoparticles,[3]and parameter estimation for chaotic systems[4]are related to the combinatorial optimization.Therefore,much attention has been attracted to designing a fast and efficient algorithm to obtain the globally optimal or near-optimal solutions to the combinatorial optimization problems. A variety of methods were proposed,such as cutting plane method,branch and bound method,and dynamic programming based on the Bellman’s principle of optimality.[1,5]However,the combinatorial optimization problems are usually NP-hard problems with time complexity in nondeterministic polynomial forms. With the increase of the scale of problems, the exact algorithms become powerless.Then, studies have been focused on the approximate algorithms or heuristic algorithms to seek near-optimal solutions of the problem with limited amount of computing time. Many algorithms were therefore proposed, such as stochastic simulated annealing (SSA) algorithm,[6–8]some algorithms based on biological phenomena,[3,8–13]and artificial neural network algorithms,[14–24]which have brought about some achievement in different scale problems.

    In 1982, Hopfield neural networks was put forward and then was successfully applied to TSP.[14,15]Hopfield neural networks can quickly converge to a stable state, but it usually is not the globally optimal solution. The solution obtained is highly related to the initial value and the parameters. It is found that, even in the problems of 10 cities, the networks often converge to the local optimal solution,and the networks become powerless for larger scale problems.[14,15]On the other hand,the chaotic neural networks based on biological electrophysiological experiments have been proved to have ergodic property and can reach the globally optimal solution during cruising,but their chaotic motions make the networks unstable in the globally optimal solution. The solution therefore could not be obtained. Chenet al.proposed the transient chaotic neural networks(TCNN)algorithm based on chaotic neural networks (CNN) and simulated annealing algorithm, and applied their algorithm to TSP.[20]By controlling the attenuation of the feedback term,the network transits from chaotic state to stable state, thus reaches to the optimal solution. Compared with SSA algorithm, TCNN algorithm uses chaotic dynamics instead of Monte Carlo random to conduct annealing,which leads to less computing time and much higher probability of global optimum in TCNN algorithm.The globally optimal solution can be obtained in medium scale TSP problems with TCNN. However, Chenet al.found that 95%of the TSP solutions for 48 cities fell on the local optimal solutions, and the algorithm lacked the ability to jump out of those local optimal regions.

    In order to improve TCNN algorithm,Wanget al. added noise into TCNN algorithm to enhance the randomness of search (stochastic chaotic neural network, SCNN) and the probabilities of the globally optimal solution for 10 cities and 21 cities were improved with the algorithm.[21]For larger scale problems (52 cities and 70 cities), the probabilities of valid solutions with SCNN algorithm were greater than that with TCNN algorithm. However, the probabilities of the globally optimal solutions for the two TSPs were not given in their work.[21]Besides, several other improved algorithms were proposed based on CNN,such as chaotic neural network with Gauss wavelet self-feedback,[22]frequency conversion sinusoidal chaotic neural network (FCSCNN),[23]and FCSCNN with hysteretic noise(HNFCSCNN).[24]Those algorithms enhanced the globally optimal probability in the TSPs of no more than 30 cities in certain degrees.However,no globally optimal solutions of larger scale problems were reported.[21–24]

    Studies show that the algorithms based on chaotic neural networks have higher searching efficiency than many other algorithms.[20–24]At the initial stage of optimization,a chaotic neural network is in chaotic states. As time evolves, the chaotic neural network gradually degenerates into a stable state, optimization then is achieved. However, a shortcoming exists in the degradation of chaotic motion. At the moment that the chaotic neural network degenerates into a stable state,the system may be far away from the globally optimal solution, which leads to the degenerated neural network falling into a local optimal state.The globally optimal or near-optimal solutions thus cannot be obtained. In their work of controlling chaos in chaotic neural networks,Heet al. found that the controlled networks can converge on a stable mode related to the initial mode when the networks operate at the critical region between chaos and non-chaos,and believed that the edge of chaos (EC) plays a very important role in memory recall process.[25]In fact,many natural systems operate in the critical states between order and disorder,such as gene expression,morphogenesis, optimal cell growth, bacterial colonies and bird colonies.[26,27]In particular, a large number of theoretical models and experiments in neuroscience show that brain runs in the critical states between order and disorder,that is,on the edge of chaos.[28,29]Therefore,it has better abilities of information transmission, information processing, information storage and learning,and is relatively stable.[26,28]In addition,studies have revealed that artificial neural networks show better computing power when operating on the edge of chaos.[30]Recently, the critical characteristics of deep learning neural networks have attracted much attention. Computer scientists have found that in order to achieve faster learning and training and to reduce the loss of information in learning, machines can only run on the edge of chaos.[31]The possible reason is that the neural networks exhibit greater mutual information at the critical region and contains more metastable states, and the neurons show stronger correlation.[32]All the above studies illustrate that the neural networks have better ability on operation and information processing on the edge of chaos.Up to now, there is no report about the application of chaotic neural networks on the edge of chaos in the field of combinatorial optimization. TSP is one of the typical combinatorial optimization problems, and is often used as a touchstone of new algorithms. In this work,an algorithm for the combinatorial optimization problems is proposed based on chaotic neural networks on the edge of chaos,and the algorithm is applied to TSPs from 10 cities to 70 cities.

    2. Model

    In order to improve previous algorithms based on chaotic neural networks,we propose a method to solve TSP by using neural networks on the edge of chaos(ECNN).In the optimal process with the algorithm,firstly,the transient chaotic neural network is used. When the chaotic neural network degenerates to the stable neural network, the parameters of the neural network are modulated to make the neural network fall in a chaotic state close to the critical region between chaos and non-chaos again and start a new round of degradation from transient chaotic states to a stable state.Such switch ends after several rounds of degradation process according to the problem to be solved. In this way, the network runs in the critical states between chaos and non-chaos, that is, on the edge of chaos, where the network has the best searching ability.Thus, the ability for the network to obtain globally optimal or near-optimal solutions is enhanced. Aihaira chaotic neural network is established based on biological electrophysiological experiments.And its chaotic dynamics is similar to biological systems.[33]Therefore, Aihara chaotic neural network is selected to solve the combinatorial optimization problems.[20]Thei-th neuron of chaotic neural network on the edge of chaos can be described by the following equations:

    wherexiis the output of thei-th neuron,yiis the internal state of thei-th neuron,kis the damping factor of nerve membrane,αis the positive scaling parameter for inputs,εis the steepness parameter of the output function,Iiis the input bias of thei-th neuron,I0is a positive constant,andz(t)is the time-dependent self-feedback connecting weight or refractory strength.z1is the initial value of the self-feedback coefficient when the network becomes chaotic again.βis the damping factor ofz,which is chosen randomly for each annealing process to enhance the randomness of the network,that is,β=rand(βmin,βmax).β0is the damping factor of the first annealing process.wijis the connection weight from thei-th neuron to thej-th neuron,which satisfies

    whereEis the energy function of a problem and the energy function of TSP will be defined in Eq.(7).σEis the variance of energy functions among the nearest 10 iterations. Large variance represents that the neural network is in the chaotic state, otherwise the network is in the stable state.σEis defined as follows:

    where ˉEis the average energy function among the nearest 10 iterations.

    Here, we use Eq. (3) to replace the feedback termz(t+1)=z(t)(1?β)in TCNN algorithm. In this way,when the network becomes stable, the sudden change of the feedback parameterzwill make the system chaotic again, so that the system has chance to escape from the local minimum region. The parameterz1is chosen close to the critical region that the network runs between chaotic state and stable state,namely on the edge of chaos, where the network has the best searching ability.

    TSP is a typical combinational optimization problem. It seeks the shortest route for a traveling salesman to visit a certain number of cities and then to return to the starting point with the constrains of visiting each city once. For TSP withncities,a network withn×nlattice structure is constituted. Its neural output,xi j, represents to visit cityiin visiting orderj.The minimized tour length of TSP satisfying all constrains can be expressed by the energy function of the neural network as following:[20]

    wherexi0=xin,xin+1=xi1,di jrepresents the distance between cityiand cityj,W1andW2represent the coupling parameters corresponding to the constraints and the cost function of the tour length, respectively. The first two terms of the equation indicate the constraints (each city can be accessed once only),and the last term represents the total length of the journey. When the energy function takes the minimum value, we get the shortest valid route satisfying the constrains.

    Combining Eqs.(2),(4),and(7),one can find that based on ECNN,the TSP dynamics of a certain number of cities can be described by the following equation:

    The initialyijare randomly generated with values between?1.0 and 1.0. Once the initial values are given, an optimal solution of the problem can be obtained through the evolution of the neural network.

    3. Application of ECNN in TSP

    We perform the simulations on the computer with the following specs: Dell PowerEdge R730 Server. Its CPU is E5-2650 V4@2.20 GHz and its memory size is 64G.The program is written in C language in a single threaded manner, that is,the output signals of the neurons are updated in turn in each iteration.

    Firstly,ECNN algorithm is applied to TSP with 10 cities.The coordinate data of inter-city distance is from Ref. [15],and the globally optimal solution of the problem is given in Fig. 1. The parameters of the network are set ask= 0.9,α=0.015,z(0)=0.10,z1=0.035,I0=0.75,ε=1/250,W1=W2=1,β0=0.5,andβ=0.003.

    Fig. 1. The optimal tour of the 10-city TSP, each number in the tour represents one city.

    Fig.2. (a)The output distribution of one neuron(x11)in the chaotic neural network for 10-city TSP with the change of parameter z. (b)The largest Lyapunov exponent. (c)Evolution of the parameter z. (d)The output Evolution of one neuron(x11). (e)Evolution of the energy function E.

    Table 1. Results with three algorithms for the 10-city TSP.

    Since the behaviors of the neurons in the network are similar,we analyze the behavior dynamics of one neuron(x11)in the network as an example. Figure 2(a)shows output distribution of one neuron(x11)in the chaotic neural network for the 10-city TSP at different self-feedback parameterz.Figure 2(b)gives the largest Lyapunov exponent of the system calculated as the definition of Ref.[34]. Figures 2(a)and 2(b)are separated to two parts. In the right side with the self-feedback parameter less than 0.029,the largest Lyapunov exponent of the system is negative and the output signals of the neuron have certain discrete values, which means that the network is in a stable state. In the left side with the self-feedback parameter larger than 0.029,the largest Lyapunov exponent of the system is positive and the output signals of the neuron are randomized with the value range from 0 to 1 as seen in the dark region of Fig.2(a),which means that the network is in chaos. Based on Figs. 2(a) and 2(b), one can choose parameterzto make the system run on the edge of chaos. Figures 2(c)–2(e)show the evolution of parameterzin ECNN algorithm,the corresponding output evolution of one neuron (x11) and the corresponding evolution of the energy functionE,respectively. With the decrease of the feedback parameterz, the output of the neuron gradually converges to a stable point, the energy of the system gradually converges to a local minimum,and network becomes stable at the end of the first annealing process. For the TCNN algorithm and the SCNN algorithm,the simulation ends right now and an optimal solution is obtained. However,the system may be far away from the globally optimal solution and the degenerated neural network actually falls into a local minimum. For the ECNN algorithm, the self-feedback parameter is adjusted to be 0.035(z1), which is slightly larger than its critical value of 0.029. As a result,the output signals of the neuron and the energy of the system become random again, which means the system enters chaotic states near the critical region between chaos and non-chaos and a new round of annealing process begins. The system thus successfully escapes from the previously local minimum state. Then the system gradually converges to another stable state as the decrease ofz. Such switch ends after several rounds of degradation process according to the given ending rules. In this way, the network works in the critical region between chaos and nonchaos,that is,the network is on the edge of chaos.From above results, one can conclude that the chaotic neural network can be controlled to operate on the edge of chaos by modulating the self-feedback parameterz.

    We simulate 5000 times for 10-city TSP with initial conditions ofyijgenerated randomly in the region [?1.0, 1.0].After three rounds of annealing process for each simulation,the optimal solution is obtained. It is found that the optimization efficiency is related to the value of the damping parameterβ. Whenβ ≤0.003, the rate of the globally optimal solution among all simulations can reach 100%. As a comparison,TCNN algorithm[20]and SCNN algorithm[21]are used for 10-city TSP. In SCNN simulations, the noise amplitudeA[n(0)]=0.002, and the damping factor of the noise isβ2=0.00005.[21]Other parameters are taken the same as those in ECNN algorithm of 10-city TSP.5000 simulations are made in both TCNN and SCNN algorithms,and the results are given in Table 1. It can be seen from Table 1 that compared with the other algorithms, ECNN algorithm shows the advantage in consuming time for optimization and the probability of the globally optimal solution.

    Then, we solve the 21-city TSP, and the coordinate data of inter-city distance is obtained from Ref. [35]. The parameters are set asz1=0.035,I=0.5,W1=1.0,W2=1.0/980,β=0.0001,β=rand(0.00001,0.0003),and other parameters are the same as those we use in the 10-city TSP.12 rounds of annealing process are taken before each simulation ends and the globally optimal tour length we get is 2707. The globally optimal route is plotted in Fig.3.

    Fig.3. The optimal tour of the 21-city TSP with tour length 2707. Each number in brackets represents one city, whereas other numbers represent the distances between two cities.

    For 48-city TSP,the data of inter-city distance is also derived from Ref. [35]. In order to save computing time, we modify the ECNN algorithm. The feedback term in Eq. (2)is added only to the randomly chosenm×m(10

    The parameters in the simulations of the 48-city TSP are set asz1= rand (0.027, 0.028),I= 0.75,W1= 1.0,W2=1.0/3160,β=0.00005, andβ=rand (0.000005, 0.00005).Other parameters are the same as those used in the 10-city TSP.In simulations,parameterz1is randomized between 0.027 and 0.028 for each annealing process, which makes the randomness of the network increase. Simulations end if more than 30 annealing had been done and there is no improvement in the nearest 20 consecutive annealing process. The globally optimal solution of the problem is derived in our simulations. The globally optimal tour length is 10628 and corresponding route is plotted in Fig.4.

    Fig. 4. The optimal tour of the 48-city TSP, each number in the tour represents one city.

    TCNN algorithm and SCNN algorithm are employed to 21-city and 48-city TSPs too. The results with 100 simulations for each case are summarized in Table 2. For SCNN algorithm, the noise amplitude is set asA[n(0)]=0.002 for 21-city TSP andA[n(0)]=0.02 for 48-city TSP and the noise damping factor is set asβ2=0.00005. The damping factors of the two algorithms are set asβ=0.000005 for 21-city TSP,andβ=0.00005 for 48-city TSP.Other parameters are taken the same as those in ECNN algorithm of 21-city and 48-city TSP. We find that smallerβcould not improve the calculating results for 48-city TSP.The reason is that with TCNN and SCNN algorithms,less iterations are taken for the 48-city TSP than for the 21-city TSP as seen in Table 2.

    Table 2. Results of three algorithms for the 21-and 48-city TSPs.

    As seen from Table 2, the probabilities of the global optimal solutions for both 21-city TSP and 48-city TSP with ECNN algorithm are higher than those with TCNN algorithm or SCNN algorithm. For 21-city TSP,the rate of the globally optimal solution with ECNN algorithm is only slightly higher that with TCNN and SCNN algorithm. However, much less computing time is demanded for ECNN algorithm to obtain the globally optimal solution of the problem. In particular,for 48-city TSP,the probability of the global optimal solutions with ECNN algorithm can reach 88%, while the probability with TCNN and SCNN algorithm is 7% and 13%, respectively. However, because of repeated switch between chaotic and non-chaotic states, the computing time in ECNN algorithm is the longest among the three algorithms. Considering the dominant superiority on the probability of the globally optimal solution,we believe that ECNN algorithm has a great improvement over TCNN and SCNN algorithms in solving 48-city TSP.

    At last, we apply our algorithm to the 70-city TSP, and the coordinate data of inter-city distances are from Ref. [35]too. The parameters of the algorithm are set asz1=0.028,I0=0.75,W1=1.0,W2=1.0/100,β=0.00005, andβ=rand(0.000005, 0.00005). Other parameters are the same as those used in the 10-city TSP. Simulations end based on the same rule as the 48-city TSP.The smallest tour length of 70-city TSP among 100 simulations is 679.However,the globally optimal tour length of the problem is 675.[35]There is a gap of 0.59%between our best solution and the globally optimal one.As a comparison, TCNN algorithm and SCNN algorithm are used for the 70-city TSP.In simulations, the damping factors of these algorithms are set as 0.000003,the noise amplitude of the SCNN algorithm is set asA[n(0)]=0.020,and the corresponding noise damping factor is set asβ2=0.00005. Other parameters are taken the same as those in ECNN algorithm of 48-city TSP. For each algorithm, we do 100 simulations.Both TCNN algorithm and SCNN algorithm fail in getting the globally optimal solution and the gaps of their best tour length from the globally optimal one are 0.59% and 1.03%, respectively. The summarized results for the 70-city TSP with three algorithms are listed in Table 3.

    Table 3. Results of three algorithms for the 70-city TSP.

    One can see from Table 3 that ECNN algorithm takes the least time to search the valid solutions among three algorithms. The probability of valid solutions with ECNN algorithm reaches 100%, while the probability with the other two algorithms is only at 54%. Besides, ECNN algorithm shows superiority in getting the best solution. The rate of the best solution is 21% with ECNN algorithm, while the rate is 1%and 5% with TCNN algorithm and SCNN algorithm, respectively. Considering the computing time and accuracy, we believe that ECNN algorithm demonstrates great advantages in solving TSP with a larger scale.

    4. Discussion

    For TSPs of 10 cities,21 cities,48 cities and 70 cities,our results prove that ECNN algorithm is much better in terms of optimization efficiency and accuracy than TCNN algorithm or SCNN algorithm,and provides an improvement in the solution of the combinatorial optimization problems.

    In order to clarify the superiority of our algorithm further, we firstly compare ECNN algorithm to several other recent works on TSP with improved algorithms based on CNN. The globally optimal probability for TSP of 30-city is 46% with the algorithm based on CNN with Gauss wavelet self-feedback,[22]30.5%with the FCSCNN algorithm,[23]and 49%with the HFCSCNN algorithm.[24]ECNN algorithm can drive the network away from local optimal states. Even in 48-city TSP, the globally optimal solution can be obtained at a probability of 88%with ECNN algorithm. This result is much better than those obtained by other algorithms based on CNN.

    Some algorithms based on biological phenomena were also proposed to solve the combinatorial optimization problems.[3,8–13]However, crossover, mutation and other extra operators for evolutionary are usually needed in these algorithms,[11,13,24]and the computing time thus are relatively longer. The local optimal solutions could not be avoided as well. In the work of Qiaoet al.,[24]the results of HNFCSCNN algorithm were compared with those of genetic algorithm(GA)[13]and particle swarm optimization(PSO)algorithm,[12]and it was found that the globally optimal solution for 30-city TSP were obtained by GA and PSO algorithm at a probability of 32.5%and 29%respectively. In the work of Denget al.,[11]a hybrid genetic algorithm was proposed and the best solution for 48-city TSP had a gap of 0.005%from the global optimal solution. With ECNN algorithm, the global optimal solution is obtained at a rate of 88%for 48-city TSP.Therefore,ECNN algorithm show advantage in the probability of the globally optimal solutions.

    In recent years, deep learning algorithms have attracted much attention and been introduced to the combinatorial optimization problems.[16–19]Globally optimal solutions can be reached in 20-city TSP[19]with deep learning algorithms,but for medium and large-scale TSP, it failed to obtain globally optimal solutions. For 50-city TSP and 100 city-TSP, the gap between the global optimal solution and the best valid solution obtained by deep learning algorithms is 0.12% and 0.87%, respectively.[19]Meanwhile, a large number of samples are required in deep learning algorithms. For example,in the work of Costaet al., about 10000 samples were used for each problem.[19]Samples are not easy to obtain as well,and it would require a lot of time to train samples. But samples are not required in the ECNN algorithm. Once the initial values are given, an optimal solution of the problem can be reached through the evolution of the neural network. Furthermore,a large probability of the globally optimal solutions can be obtained with ECNN algorithm for small and mediumsized problems. The algorithm based on the edge of chaos is therefore a good way to solve the combinatorial optimization problems.

    The number of iterations and the computing time is dependent on the scale of problems,the parameters of networks and the ending rule of a simulation as well. In general, more computing time is needed for larger scale problems. For TSP of 48 cities,we concentrate on finding the global optimal solutions of the problem at a high probability,and the parameters and ending rule are set to raise the traversal time in the chaotic region. The computing time with ECNN algorithm thus is longer than with TCNN algorithm or SCNN algorithm. However, for TSP of 70 cities, the global optimal solution could not be obtained with the ECNN algorithm since it is not easy to set suitable parameters. We therefore focus on searching for the best valid solutions of the problem. Traversal time in the chaotic region is shorten under the chosen parameters and ending rules.Thus,less computing time is needed with ECNN algorithm compared to the other two algorithms.

    It has been noticed that the parameters of networks are important to obtain optimal solutions in ECNN algorithm as in TCNN,[20]SCNN[21]and other algorithms based on chaotic neural networks.[22–24]For ECNN algorithm, solutions are sensitive to the initial value of the self-feedback coefficientz1,the damping factorβand the coupling parametersW1andW2. Parameterz1is chosen by a rule thatz1should be chosen to make the network near but not too close to the critical region, that is, on the edge of chaos. Larger values ofz1will cause the network to enter into the chaotic region deeper,thus reduces the optimization efficiency as well as the calculation accuracy. On the other hand,ifz1makes networks very close to the critical region between chaos and non-chaos,the traversal time in the chaotic region is too short to escape from a local minimum,thus reduces calculation accuracy. Studies have revealed that artificial neural networks show strong computing power when operating on the edge of chaos.[30]Thus by modulating the self-feedback parameterzto make the neural network operate on the edge of chaos, the optimization efficiency is improved and the probability of the globally optimal solution increase. The damping factorβand the coupling parametersW1andW2are selected according to the experience.However, it is not easy to find a set of appropriate parameters for TSPs with a large scale. The same problem exists in TCNN,[20]SCNN[21]and other algorithms based on chaotic neural networks.[22–24]Efforts have been made to conduct parameter selection. For example, the algorithm hybridization has been tested in optimization problems.[11,12]In the future,we will combine ECNN algorithm with other algorithms to improve parameter selection. Meanwhile, modulating important parameters by searching dynamics is another way we plan to try.

    5. Conclusion

    In this work,an algorithm for the combinatorial optimization problems is proposed based on neural networks on the edge of chaos(ECNN),and is then applied to traveling salesman problems of 10 cities, 21 cities, 48 cities and 70 cities.The results show that for small scale TSPs of 10 cities and 21 cities,ECNN algorithm has higher probability of the globally optimal solutions and less time consumption than TCNN algorithm and SCNN algorithm. Since ECNN algorithm has more power to drive the networks away from local minimum,it shows stronger advantage in searching the globally optimal solution or near-optimal solution for larger scale TSP than TCNN algorithm and SCNN algorithm. In TSP with 48 cities,the probability of the globally optimal solution from ECNN algorithm can reach 88%,while the probabilities with TCNN and SCNN algorithm are only 6% and 13%, respectively. In 70-city TSP,ECNN algorithm provides much higher the probabilities of valid solutions and the best optimal solution,while spends less computing time than TCNN and SCNN algorithm.Moreover, ECNN algorithm is superior to other algorithms based on CNN,some algorithms based on biological phenomena and deep learning algorithms in searching for the optimal solutions of TSPs. To conclude,ECNN algorithm provides an effective way for solving the combinatorial optimization problems.

    猜你喜歡
    方明張娜國光
    Effect of cognitive training on brain dynamics
    動作不可少(下)
    動作不能少(上)
    凝心固本 引智聚力 創(chuàng)新開拓
    著名旅法詩人 方明
    鴨綠江(2020年23期)2020-10-13 13:29:04
    解三角方程題的常用方法
    松樹梢
    叫桃的女人
    尋找一個叫桃的女人
    短篇小說(2014年7期)2014-09-20 15:08:53
    Detection of promoter methylation of p27 gene in gastric carcinoma by methylation-specific PCR technique
    女的被弄到高潮叫床怎么办 | 日本-黄色视频高清免费观看| 99热这里只有是精品50| 一本久久中文字幕| 亚洲黑人精品在线| 免费观看人在逋| 中文字幕熟女人妻在线| 丰满人妻一区二区三区视频av| 少妇被粗大猛烈的视频| 日韩中字成人| 一级a爱片免费观看的视频| 男人舔奶头视频| 麻豆成人av在线观看| 老熟妇乱子伦视频在线观看| 欧美精品啪啪一区二区三区| 麻豆一二三区av精品| 精品一区二区三区av网在线观看| 久久人妻av系列| 老女人水多毛片| 高清在线国产一区| 精品99又大又爽又粗少妇毛片 | 国产白丝娇喘喷水9色精品| 亚洲性久久影院| 国产精华一区二区三区| 国产黄色小视频在线观看| 嫩草影院新地址| 亚洲av第一区精品v没综合| 亚洲狠狠婷婷综合久久图片| 黄色视频,在线免费观看| 亚洲国产欧美人成| 国产黄a三级三级三级人| 精品久久久久久久久av| 99在线人妻在线中文字幕| 日韩欧美一区二区三区在线观看| 国产一区二区三区av在线 | 成年免费大片在线观看| 乱系列少妇在线播放| 高清毛片免费观看视频网站| 午夜福利在线观看吧| 亚洲精华国产精华精| 亚洲狠狠婷婷综合久久图片| 欧美性猛交黑人性爽| 我要看日韩黄色一级片| 色综合站精品国产| 国产精品自产拍在线观看55亚洲| 精品久久久久久久久av| 九色成人免费人妻av| 我的女老师完整版在线观看| 久久久久久久久大av| 亚洲精品日韩av片在线观看| 永久网站在线| 欧美精品啪啪一区二区三区| 能在线免费观看的黄片| 男女啪啪激烈高潮av片| 欧美最新免费一区二区三区| 赤兔流量卡办理| 久久草成人影院| 精品欧美国产一区二区三| 天美传媒精品一区二区| 欧美日韩中文字幕国产精品一区二区三区| 成人特级黄色片久久久久久久| 搡老熟女国产l中国老女人| 午夜福利成人在线免费观看| 欧美激情在线99| 亚洲18禁久久av| av国产免费在线观看| 国产精品久久久久久亚洲av鲁大| 久久久午夜欧美精品| 亚洲人成网站在线播放欧美日韩| 波多野结衣巨乳人妻| 久久久久久久久久久丰满 | 亚洲国产高清在线一区二区三| 免费一级毛片在线播放高清视频| 久久久国产成人精品二区| 色精品久久人妻99蜜桃| 人妻夜夜爽99麻豆av| 别揉我奶头 嗯啊视频| 免费在线观看影片大全网站| 最近视频中文字幕2019在线8| 欧美日韩瑟瑟在线播放| 国产精品一区二区免费欧美| 亚洲av电影不卡..在线观看| 99热这里只有是精品在线观看| 亚洲国产欧美人成| 美女高潮喷水抽搐中文字幕| 久久这里只有精品中国| 亚洲精品在线观看二区| 亚洲第一区二区三区不卡| 亚洲欧美日韩东京热| 国产亚洲精品av在线| 直男gayav资源| 国产亚洲欧美98| 1024手机看黄色片| 在线观看午夜福利视频| 免费看美女性在线毛片视频| 国产欧美日韩精品亚洲av| 91在线观看av| 啦啦啦韩国在线观看视频| 国产大屁股一区二区在线视频| 一进一出抽搐动态| 亚洲精华国产精华精| 国产伦人伦偷精品视频| 久久精品综合一区二区三区| 丰满人妻一区二区三区视频av| 搡女人真爽免费视频火全软件 | 国产真实乱freesex| 观看美女的网站| 麻豆国产97在线/欧美| 99久久中文字幕三级久久日本| 婷婷色综合大香蕉| 不卡一级毛片| 亚洲精品456在线播放app | 真实男女啪啪啪动态图| 国产v大片淫在线免费观看| 日韩 亚洲 欧美在线| 日韩高清综合在线| 免费看美女性在线毛片视频| 噜噜噜噜噜久久久久久91| 国产一区二区亚洲精品在线观看| 丰满人妻一区二区三区视频av| 午夜福利欧美成人| 18禁黄网站禁片免费观看直播| 97碰自拍视频| 成人永久免费在线观看视频| 99热这里只有是精品50| 国产 一区 欧美 日韩| 特大巨黑吊av在线直播| 真人一进一出gif抽搐免费| 午夜精品在线福利| 免费人成视频x8x8入口观看| 欧美黑人巨大hd| 亚洲av一区综合| 综合色av麻豆| 长腿黑丝高跟| 国产精品一区二区三区四区久久| 亚洲av.av天堂| 国产伦精品一区二区三区视频9| 亚洲一级一片aⅴ在线观看| 色综合亚洲欧美另类图片| 99热网站在线观看| 久久婷婷人人爽人人干人人爱| 久久人人爽人人爽人人片va| 久久久久久伊人网av| av黄色大香蕉| 变态另类丝袜制服| 亚洲精品在线观看二区| 亚洲成人久久性| 一进一出抽搐gif免费好疼| 久久香蕉精品热| 99视频精品全部免费 在线| 女同久久另类99精品国产91| 欧美高清成人免费视频www| 久久久久久久久中文| 精品久久久久久久人妻蜜臀av| 国产探花在线观看一区二区| 免费一级毛片在线播放高清视频| 联通29元200g的流量卡| 天堂av国产一区二区熟女人妻| 免费人成在线观看视频色| 亚洲国产精品合色在线| 欧美绝顶高潮抽搐喷水| 看片在线看免费视频| 99热这里只有是精品50| 亚洲无线观看免费| 欧美日韩国产亚洲二区| 日韩,欧美,国产一区二区三区 | 国产主播在线观看一区二区| 久久人人精品亚洲av| 男女啪啪激烈高潮av片| 欧美+亚洲+日韩+国产| 国产午夜精品久久久久久一区二区三区 | 久久久久久久久久黄片| 久久婷婷人人爽人人干人人爱| 午夜福利18| 免费看a级黄色片| 午夜福利在线在线| 日韩人妻高清精品专区| 熟妇人妻久久中文字幕3abv| 少妇的逼水好多| 窝窝影院91人妻| 级片在线观看| av在线老鸭窝| 级片在线观看| 成人美女网站在线观看视频| 老女人水多毛片| 观看免费一级毛片| 精品久久久久久成人av| 精品久久久久久久久久久久久| 午夜精品一区二区三区免费看| 国产成人a区在线观看| av黄色大香蕉| 久久久精品大字幕| 色5月婷婷丁香| 日韩在线高清观看一区二区三区 | 99热6这里只有精品| 啪啪无遮挡十八禁网站| 中国美女看黄片| 91在线精品国自产拍蜜月| 国产伦人伦偷精品视频| 91久久精品国产一区二区成人| 一级毛片久久久久久久久女| 欧美性猛交╳xxx乱大交人| 免费观看人在逋| 两个人视频免费观看高清| 午夜亚洲福利在线播放| 舔av片在线| 我要看日韩黄色一级片| 国内毛片毛片毛片毛片毛片| 我要搜黄色片| 性插视频无遮挡在线免费观看| 亚洲精华国产精华精| 真人一进一出gif抽搐免费| 久久午夜亚洲精品久久| 男插女下体视频免费在线播放| 一本久久中文字幕| 亚洲精华国产精华精| 一本久久中文字幕| 日本一二三区视频观看| 午夜激情欧美在线| 99精品久久久久人妻精品| 两人在一起打扑克的视频| 亚洲av电影不卡..在线观看| 日本黄色片子视频| 欧美另类亚洲清纯唯美| 亚洲avbb在线观看| 别揉我奶头 嗯啊视频| 又黄又爽又刺激的免费视频.| 日日摸夜夜添夜夜添小说| 赤兔流量卡办理| 久久国产精品人妻蜜桃| 波多野结衣巨乳人妻| 日本-黄色视频高清免费观看| 热99在线观看视频| 一夜夜www| 国产亚洲欧美98| 国产在线男女| 精品一区二区三区视频在线观看免费| 免费av毛片视频| 一个人免费在线观看电影| 亚洲成a人片在线一区二区| 日韩大尺度精品在线看网址| 在线a可以看的网站| 夜夜爽天天搞| 男女下面进入的视频免费午夜| 国产精品久久久久久亚洲av鲁大| www日本黄色视频网| 精品午夜福利在线看| 欧美性猛交黑人性爽| 极品教师在线免费播放| 69人妻影院| 黄色一级大片看看| 中文字幕av成人在线电影| 男女啪啪激烈高潮av片| av在线亚洲专区| 国产精品99久久久久久久久| 国产av麻豆久久久久久久| 成人美女网站在线观看视频| 18+在线观看网站| 看免费成人av毛片| 日本成人三级电影网站| 国产精品亚洲一级av第二区| 国产高清视频在线播放一区| 草草在线视频免费看| 他把我摸到了高潮在线观看| 欧美一区二区国产精品久久精品| 嫩草影院入口| 午夜激情欧美在线| 午夜福利成人在线免费观看| 淫妇啪啪啪对白视频| 欧美精品啪啪一区二区三区| 免费在线观看影片大全网站| 亚洲五月天丁香| 很黄的视频免费| 午夜精品一区二区三区免费看| 成人午夜高清在线视频| 听说在线观看完整版免费高清| 国产乱人视频| 国产成年人精品一区二区| 午夜精品在线福利| av在线老鸭窝| 国产精品98久久久久久宅男小说| 男女做爰动态图高潮gif福利片| 久久人人爽人人爽人人片va| 美女xxoo啪啪120秒动态图| 精品免费久久久久久久清纯| 在线免费十八禁| 欧美成人免费av一区二区三区| 啦啦啦韩国在线观看视频| 国产黄片美女视频| 久久久午夜欧美精品| 免费不卡的大黄色大毛片视频在线观看 | 久久精品夜夜夜夜夜久久蜜豆| 久久精品国产亚洲av香蕉五月| 久久久成人免费电影| 女的被弄到高潮叫床怎么办 | videossex国产| 成人午夜高清在线视频| 精品欧美国产一区二区三| 欧美高清性xxxxhd video| 最好的美女福利视频网| 最新在线观看一区二区三区| 天堂√8在线中文| 国产爱豆传媒在线观看| 国产探花在线观看一区二区| 精品久久久噜噜| 国产色爽女视频免费观看| 色尼玛亚洲综合影院| 国产单亲对白刺激| 午夜精品久久久久久毛片777| 一区二区三区免费毛片| 国产精品久久视频播放| 亚洲国产色片| 免费观看的影片在线观看| 美女免费视频网站| 少妇丰满av| 国产真实伦视频高清在线观看 | 日日撸夜夜添| 男人舔女人下体高潮全视频| 免费电影在线观看免费观看| 午夜影院日韩av| 日韩强制内射视频| 美女高潮喷水抽搐中文字幕| 九九热线精品视视频播放| 高清日韩中文字幕在线| 久久久久国内视频| 午夜精品一区二区三区免费看| 日本黄大片高清| 一个人看视频在线观看www免费| 日日撸夜夜添| 久久久久九九精品影院| 成人精品一区二区免费| 免费电影在线观看免费观看| 国产成人aa在线观看| 人妻制服诱惑在线中文字幕| 亚洲国产欧美人成| 成熟少妇高潮喷水视频| 亚洲精品一卡2卡三卡4卡5卡| 听说在线观看完整版免费高清| 亚洲男人的天堂狠狠| 人人妻,人人澡人人爽秒播| 日韩av在线大香蕉| 国内精品久久久久精免费| 男女做爰动态图高潮gif福利片| 毛片一级片免费看久久久久 | 最近最新中文字幕大全电影3| 搡老岳熟女国产| 国产亚洲av嫩草精品影院| 美女被艹到高潮喷水动态| 精品人妻视频免费看| 午夜老司机福利剧场| 日韩在线高清观看一区二区三区 | 亚洲av免费高清在线观看| 麻豆一二三区av精品| 国产亚洲91精品色在线| 亚洲av电影不卡..在线观看| 永久网站在线| 毛片女人毛片| 欧美不卡视频在线免费观看| 欧美日韩乱码在线| 日日夜夜操网爽| 美女高潮喷水抽搐中文字幕| 亚洲不卡免费看| 国产亚洲精品综合一区在线观看| 国产成人福利小说| 特大巨黑吊av在线直播| 中文资源天堂在线| a级毛片a级免费在线| 亚洲精品乱码久久久v下载方式| 久久精品91蜜桃| 欧美成人a在线观看| 亚洲天堂国产精品一区在线| 人妻夜夜爽99麻豆av| a在线观看视频网站| 亚洲性久久影院| 欧美黑人欧美精品刺激| 午夜久久久久精精品| 一进一出抽搐gif免费好疼| 两个人视频免费观看高清| 精品久久久久久久久久免费视频| 国产一级毛片七仙女欲春2| 美女xxoo啪啪120秒动态图| 精品午夜福利视频在线观看一区| 久久热精品热| 一本久久中文字幕| 午夜福利在线观看免费完整高清在 | 18+在线观看网站| 精品久久国产蜜桃| 少妇的逼水好多| 国产一区二区亚洲精品在线观看| 免费一级毛片在线播放高清视频| 国产一区二区亚洲精品在线观看| 亚洲精华国产精华液的使用体验 | 亚洲精品色激情综合| 国产 一区 欧美 日韩| 又粗又爽又猛毛片免费看| 免费看美女性在线毛片视频| 欧美激情国产日韩精品一区| 国产一区二区三区在线臀色熟女| 国产成人一区二区在线| 国产高清有码在线观看视频| 老熟妇乱子伦视频在线观看| 国产精品久久视频播放| 小说图片视频综合网站| 久久天躁狠狠躁夜夜2o2o| 欧美日韩瑟瑟在线播放| 日本免费a在线| 别揉我奶头 嗯啊视频| 在线观看一区二区三区| 又爽又黄无遮挡网站| 国产精品国产高清国产av| 女人十人毛片免费观看3o分钟| 亚洲三级黄色毛片| 亚洲国产欧美人成| 久久久精品大字幕| 哪里可以看免费的av片| 老师上课跳d突然被开到最大视频| 一进一出好大好爽视频| 日本撒尿小便嘘嘘汇集6| 亚洲中文字幕日韩| 免费观看的影片在线观看| 久久精品国产亚洲av涩爱 | 欧洲精品卡2卡3卡4卡5卡区| 国产一区二区三区视频了| 精华霜和精华液先用哪个| 国产精品一区二区性色av| 最近在线观看免费完整版| 成人欧美大片| 国产精品,欧美在线| 精品午夜福利视频在线观看一区| 成人毛片a级毛片在线播放| 亚洲精品一区av在线观看| 国内精品一区二区在线观看| 午夜福利在线观看免费完整高清在 | 日韩欧美国产一区二区入口| 啦啦啦啦在线视频资源| 欧美日韩精品成人综合77777| 国产日本99.免费观看| 黄色配什么色好看| 嫩草影院入口| 亚洲av五月六月丁香网| 在线观看午夜福利视频| 看十八女毛片水多多多| 一个人免费在线观看电影| 两性午夜刺激爽爽歪歪视频在线观看| or卡值多少钱| 大又大粗又爽又黄少妇毛片口| 最新中文字幕久久久久| 国产单亲对白刺激| 婷婷六月久久综合丁香| 欧美另类亚洲清纯唯美| 免费av毛片视频| 99热这里只有精品一区| 久久久久久伊人网av| 国产高清有码在线观看视频| 三级毛片av免费| 最近最新中文字幕大全电影3| av福利片在线观看| 日日干狠狠操夜夜爽| 国产乱人伦免费视频| 最近中文字幕高清免费大全6 | 久久久午夜欧美精品| 成年人黄色毛片网站| 22中文网久久字幕| 成人av在线播放网站| 国产极品精品免费视频能看的| 高清毛片免费观看视频网站| 亚洲欧美日韩东京热| 女人十人毛片免费观看3o分钟| 日韩欧美国产在线观看| www.色视频.com| 日本一本二区三区精品| 久久精品综合一区二区三区| 亚洲av中文字字幕乱码综合| 琪琪午夜伦伦电影理论片6080| 国产精品永久免费网站| 成人国产综合亚洲| 少妇丰满av| 国产精华一区二区三区| 老熟妇仑乱视频hdxx| 最近在线观看免费完整版| 精品人妻1区二区| 午夜福利视频1000在线观看| 91午夜精品亚洲一区二区三区 | 婷婷精品国产亚洲av在线| 国产单亲对白刺激| 午夜福利18| 中文字幕免费在线视频6| 亚洲欧美日韩无卡精品| 国产69精品久久久久777片| 国产不卡一卡二| 婷婷丁香在线五月| 色综合亚洲欧美另类图片| 九九爱精品视频在线观看| 成人特级黄色片久久久久久久| 久久精品国产清高在天天线| 亚洲av.av天堂| 婷婷精品国产亚洲av| 天堂动漫精品| 能在线免费观看的黄片| 成人美女网站在线观看视频| 亚洲一区二区三区色噜噜| 波多野结衣高清作品| aaaaa片日本免费| 人人妻人人看人人澡| 免费看a级黄色片| 综合色av麻豆| av黄色大香蕉| 少妇的逼好多水| 国产成人a区在线观看| 国产蜜桃级精品一区二区三区| 成人亚洲精品av一区二区| 亚洲精品在线观看二区| АⅤ资源中文在线天堂| 乱系列少妇在线播放| 搡老熟女国产l中国老女人| 老熟妇仑乱视频hdxx| 国产精品爽爽va在线观看网站| 欧美色欧美亚洲另类二区| 国产av麻豆久久久久久久| 亚洲电影在线观看av| 日日撸夜夜添| 久99久视频精品免费| 老熟妇仑乱视频hdxx| 人妻少妇偷人精品九色| 12—13女人毛片做爰片一| 别揉我奶头~嗯~啊~动态视频| 三级国产精品欧美在线观看| 两人在一起打扑克的视频| 成人毛片a级毛片在线播放| 乱系列少妇在线播放| 少妇被粗大猛烈的视频| 亚洲五月天丁香| 一本精品99久久精品77| 中文字幕av成人在线电影| 久久久午夜欧美精品| 久久婷婷人人爽人人干人人爱| 日日啪夜夜撸| 日韩欧美精品免费久久| 日日摸夜夜添夜夜添小说| 神马国产精品三级电影在线观看| 听说在线观看完整版免费高清| 亚洲av电影不卡..在线观看| 欧美人与善性xxx| 少妇被粗大猛烈的视频| 亚洲五月天丁香| 又黄又爽又免费观看的视频| 亚洲最大成人av| 免费无遮挡裸体视频| 一区二区三区免费毛片| 国产一级毛片七仙女欲春2| 久久草成人影院| 少妇丰满av| 久久精品国产亚洲av天美| 男女边吃奶边做爰视频| 色5月婷婷丁香| 嫁个100分男人电影在线观看| 国产午夜福利久久久久久| 好男人在线观看高清免费视频| 亚洲精品成人久久久久久| 一卡2卡三卡四卡精品乱码亚洲| 久久精品夜夜夜夜夜久久蜜豆| 啦啦啦韩国在线观看视频| 亚洲狠狠婷婷综合久久图片| 欧美人与善性xxx| 国产精品嫩草影院av在线观看 | 一区二区三区激情视频| 亚洲无线在线观看| 女同久久另类99精品国产91| 欧美色欧美亚洲另类二区| 丰满人妻一区二区三区视频av| 欧美人与善性xxx| 日韩亚洲欧美综合| 久久精品久久久久久噜噜老黄 | 亚洲内射少妇av| 国产精品一区二区三区四区免费观看 | 国产精品嫩草影院av在线观看 | 麻豆精品久久久久久蜜桃| 国产私拍福利视频在线观看| 国产精品99久久久久久久久| 一本一本综合久久| 男人舔女人下体高潮全视频| 麻豆一二三区av精品| 97热精品久久久久久| 人人妻,人人澡人人爽秒播| 久久国产精品人妻蜜桃| 男女之事视频高清在线观看| 悠悠久久av| 亚洲avbb在线观看| 狂野欧美激情性xxxx在线观看| 成人精品一区二区免费| a在线观看视频网站| 国内少妇人妻偷人精品xxx网站| 日本在线视频免费播放| 免费高清视频大片| 99久久九九国产精品国产免费| 99riav亚洲国产免费| 国语自产精品视频在线第100页| 午夜激情欧美在线| 国产精品无大码| 国产探花在线观看一区二区| 美女xxoo啪啪120秒动态图| 亚洲va日本ⅴa欧美va伊人久久| 成人无遮挡网站| 波野结衣二区三区在线| 在线看三级毛片| 久久精品国产自在天天线| 少妇的逼水好多| 校园人妻丝袜中文字幕| 九色国产91popny在线| 国产一区二区三区在线臀色熟女| 国产午夜精品论理片| 91久久精品国产一区二区成人| 成人美女网站在线观看视频| 亚洲第一区二区三区不卡| 在线国产一区二区在线|