• <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 ?

    Airport gate assignment problem with deep reinforcement learning①

    2020-04-13 07:06:04ZhaoJiaming趙家明WuWenjunLiuZhimingHanChanghaoZhangXuanyiZhangYanhua
    High Technology Letters 2020年1期
    關(guān)鍵詞:趙家

    Zhao Jiaming(趙家明),Wu Wenjun,Liu Zhiming,Han Changhao,Zhang Xuanyi,Zhang Yanhua

    (*Faculty of Information Technology,Beijing University of Technology,Beijing 100124,P.R.China)(**IT Department,Beijing Capital International Airport Co.Ltd,Beijing 100124,P.R.China)

    Abstract

    Key words:airport gate assignment problem (AGAP),deep reinforcement learning (DRL),Markov decision process (MDP)

    0 Introduction

    In recent years,the air transportation has developed rapidly.However,this development has also brought great challenges to the operation and management of civil aviation.Airport is the terminal of any airline,bearing great pressures of the rapid growth of air traffic.The limited resources and operating abilities of airport has become one of the main causes of flight delays.In the airport operation,gates including fixed gates next to the terminal and remote gates located on apron are key resources.The gate assignment directly affects the airport operation efficiency and the quality of experience (QoE) of passengers.Therefore,the optimal assignment of gates plays an important role in practice.

    In 1974,Steuart modelled the airport gate assignment problem (AGAP) for the first time[1].Dorndorf summarized the optimization objectives and constraints of the AGAP problem in 2007[2].As a result of his contribution,the academic research on AGAP has turned to the innovation of algorithm.The AGAP algorithms are mainly divided into 2 categories:mathematical programming algorithm and heuristic algorithm.The representative algorithms in the mathematical programming algorithm are the column generation method and the branch and bound algorithm.Both 2 algorithms were used to solve AGAP[3,4]and achieved some positive results.The advantage of mathematical programming algorithm is that it can be found the global optimal solution of the problem.Meanwhile,the shortage is also very obvious,i.e.these algorithms are only applicable to the AGAP with a small number of flights and gates due to their high complexity.The heuristic or modern heuristic algorithms are more popular in academics.The object such as minimizing the idle time of gates,reducing the number of flights that have not been assigned to gate,equalizing the idle time of aircraft,minimizing the distance traveled of passengers and the number of flights assigned to remote gate are considered.The greedy algorithm,genetic algorithm,Tabu search algorithm,bee colony algorithm and several other heuristic algorithms are used to solve these problems[5-9].Ref.[10] has also developed an airport gate assignment framework to deal with the issue of aircraft assignment under random flight delay and adopted a heuristic algorithm module to iteratively solve the AGAP problem.

    Although researches have been carried out in the academia,the AGAP still faces many problems in practice.In most cases,the AGAP is divided into 2 stages in practice,i.e.the pre-assignment stage based on known flight plans and the dynamic assignment stage to adjust the random change of flight time.The existing algorithms can be used to solve the pre-assignment problems.However,the contradiction between the complexity of the existing algorithms and the timeliness of the dynamic assignment are both difficult problems.

    In the development of artificial intelligence (AI),deep reinforcement learning (DRL) technology has received extensive attention[11,12],which can be applied to solve complex decision problems in various fields.In Ref.[13],DRL method was used to solve the problem of scheduling tasks with multiple resource demands.The results show that the proposed deep RM performs better than the heuristic algorithm and adapts to different conditions.This has significant implications for the research.AGAP is also a kind of task scheduling and resource assignment problem,in which tasks are flights and resources are the airport gates.The important point is that DRL based algorithm can satisfy the timeliness requirements of the dynamic AGAP.

    In this research,a DRL based method to solve the AGAP with random flight delay is proposed.Considering the passenger’s QoE,the distance traveled of passengers is reduced by maximizing the ratio of flights assigned to fixed gates next to the terminal.The AGAP is translated into a Markov decision process (MDP) which can be solved as a learning problem.The proposed DRL based method is validated via simulation and the solution using the software Gurobi is given for comparison.Simulation results show that the optimization performances of the proposed method and Gurobi based solution are quite close.Meanwhile,based on the proposed method,the calculation speed of the trained airport gate assignment policy is much faster than that of the Gurobi based solution.This confirms that the proposed method can be used to solve the dynamic assignment problem in practice.The main contribution of this research is summarized as follows.

    1) The MDP model of the dynamic AGAP is proposed with constrains of gate resources and time interval modeled in the states.

    2) The DRL framework is adopted as the solution of the dynamic AGAP.Simulation results validate that the performance of the proposed DRL-AGAP method is close to that of the Gurobi optimization solver with extremely low computing complexity.

    The rest of this paper is organized as follows.System model is presented in Section 1.In Section 2,an optimization problem for AGAP is formulated into an MDP.Then the problem solution via deep reinforcement learning is proposed in Section 3.Section 4 discusses the simulation results.Finally,the proposed method is concluded with future work in Section 5.

    1 System model

    In this work,the research is based on the AGAP with DRL.Before landing of a certain flight,the gate assignment agent will assign it to an available gate.The system model is described in Fig.1.The total number of flights and gates are denoted byNandM,respectively.The landing time of all the flights is based on a fixed schedule with random time fluctuation.Two types of gates are considered:fixed gates next to the terminal and remote gates located on apron.As an initial research on the AGAP with DRL,the simplest scenario is taken into consideration.Assuming the size of all the gates is the same and all the aircraft types are the same.And also assuming runway assignment has been made prior to the assignment of gates[14].Besides,the runway conflict is not considered in this initial research.

    Fig.1 Example of airport gates assignment

    1.1 Objective

    Airport gate assignment is closely related to passenger’s QoE.In general,people prefer a shorter waiting time and a shorter walking distance after landing.Therefore,when the flight is assigned to the fixed gates,the passengers on the flight are satisfied the most.When the flight is assigned to the remote gates,passengers have to go back to the terminal by shuttle bus.Thus,the passengers are not satisfied and the operation cost is higher.Taking the QoE of passengers as the key operational indicators,the optimization object can be designed as maximizing the rate of flights assigned to fixed gates.The objective function can be calculated as

    (1)

    where,Y=[yij]N×Mis the assignment matrix and the optimization variable,g=(g1,g2,…,gj,…,gM) is the gate type vector.If flightiis assigned to the gatej,yij=1,otherwiseyij=0.If the gatejis a fixed gate,gj=1,otherwisegj=0.

    1.2 Constraints

    As the ideal scenario for AGAP is considered in this work,only the basic constraints are taken into account.

    1) Each flight must be assigned to one and only one gate

    (2)

    where,Mis the total number of gates.This formula ensuresi-th could only be assigned to one gate.

    2) Time interval constraint

    Normally,flight takes some time in the process of entering and leaving the gate.When two or more flights are assigned in the same gate,there must be a certain interval constraint between the departure time of the previous flight and landing time of the next flight[15].According to Ref.[15],the next flight should be assigned to the gate at least 300 s after the departure of the previous one.It can be mathematically expressed as

    yikyjk(lj-di-300)≥0

    (3)

    where,flightfiandfjare assigned to the same gatek,fiis earlier one,yik=yjk=1.ljis landing time offj,diis departure time offj.

    3) 0/1 variable constraint

    yij∈{0,1}

    (4)

    yijis the decision variable of AGAP.AGAP is modeled as a 0-1 programming problem.

    1.3 AGAP optimization mode

    According to the optimization objective and constraints given above,the optimization problem in this paper can be expressed as

    yikyjk(lj-di-300)≥0

    (5)

    yij∈{0,1}

    2 Markov decision process formulation of AGAP

    The Markov decision process provides a mathematical architecture model for how to make decisions in a state where part of the randomness is partially controlled by the decision maker[16].As the arriving flow of flight is a sequence in time,the assignment decision for these flights must be made as discrete time series.Therefore,the real-time (dynamic) AGAP can be naturally modeled as a Markov decision process.In the rest of this section,the MDP formulation of AGAP and the definition of state space,action space,and rewards will be described.

    2.1 State space

    The state is formulated as a resource images denoted bySt.The resource view can be divided into two parts:gate image and fight image.

    As shown in Fig.2,the horizontal axis represents the gate resources,and the vertical axis represents the time steps.The gate image represents the gate resource occupancy and the available gate resources from current time to aTmaxtime steps in the future.The flight image represents duration between the landing time and the departure time of the next several flights.To simplify the expression of the state,the safety interval is also included in the duration.Besides,the occupied gates are marked in the flight image by colored,which will be updated at each time step after gate assignment.

    Fig.2 An example of state representation,with gate resources and 3 pending flights

    More specifically,taking the colorful image in Fig.2 as an example,the different colors in these images represent different flights.At the current time step,gates 1,4 and 7 are occupied.The blue colored flight is assigned to gate 1 and will still stay in gate 1 for two more time steps.The flight images represent the gate resource requirements of awaiting flights.Flight 1 requires to stay for 5 time steps and there are 7 available gates for it currently.

    2.2 Action space

    Theoretically,if the flights information remains unchanged during the experimental time,the agent can decide the gates assignment for theNflights at the same time and the size of action space isMN.However,due to the random time fluctuation of the landing time,only the predicted landing time of the upcoming flight is accurate.Thus the dynamic AGAP discussed in this paper is more accurate and practical.As a result,for dynamic AGAP,agent only assigns the first awaiting flight to an available gate,and the size of the action space is greatly reduced toM.The action can be denoted byAt∈{1,2,3,…,M-1,M},whereAt=imeans to assign the first awaiting flight to thei-th gate.

    2.3 State transition

    When an action has been selected,it will changes the state.Fig.3 shows how the state is affected by the action.In the first row of images,Flight 1 is the first awaiting flight to be assigned.According to the gate image,Fight 1 can not be assigned to gate 1,4 and 7.Flight 2 is the second awaiting flight.According to the image,it will arrive 2 time steps later and leave 7 time steps later.During that time,only gate 1,7 are occupied,gate 4 will be available.If the agent selects gate 2 for Flight 1,the state will change to the images in the second row.Gate 2 will be occupied and the flight images are changed accordingly.

    Fig.3 An example of state transition

    2.4 Rewards

    As mentioned in Section 1,the proposed objective is to maximize the rate of flights assigned to fixed gates.The rewardRis set asRbwhen the flight is assigned to the fixed gate;Rgwhen assigned to the remote gate.The relationship thatRb>Rgis satisfied,which ensures that the agent tends to assigned more flights to the fixed gates.

    According to the definition of value of state in reinforcement learning and the definition of the equivalent optimization objective in Eq.(5),the value of the initial stateS0can be calculated as

    (7)

    where,γis the discount factor andRtis the reward of time stept.Ifγ=1,the reward at time steptcan be calculated as

    Rt=gAt

    (8)

    whereAtis the selected action at time stept.

    3 Deep reinforcement learning based problem solution

    As the state transition of the MDP of AGAP is quite complicated,a DRL based method is proposed to solve the dynamic AGAP called DRL-AGAP.The policy which the agent uses to make gate assignment decision is designed as a deep neural network (DNN) denoted by πθ.The AGAP is transformed to a DRL problem and the major work is to train the policy network.The framework of DRL-AGAP is given in Fig.4.

    Fig.4 The framework of DRL-AGAP

    More detailed information of the policy gradient method used can be found in Ref.[13].In Algorithm 1,the process of one time step in the episodic simulation is detailed.

    Algorithm1 Simulationprocessofk-thepisodeofj-thflightschedulesStart: Step1:InitializestateSk0. Step2:t=0 Step3:Whetherthereisaflightinthestateoftheflight,ifthereisaflightgotoStep4,otherwisegotoStep8. Step4:InputSktintopolicynetworkπθ,calculatetheas-signmentprobabilityp=(p1,p2,…,pM)ofthefirstflightintheflightstate. Step5:Deleteunavailablegatesinp.Availablegatesprobabilityisp′. Step6:Chooseagateaccordingtop′asAkt. Step7:CalculateRkt,andstore(Skt,Akt,Rkt). Step8:t=t+1 Step9:Ift

    4 Simulation results

    In this section,the simulation results of the proposed DRL-AGAP method will be introduced.As an initial research,a neural network is built with a fully connected hidden layer with 200 neurons,which is the policy network.The maximum visible time steps of the resource view isTmax=20.Each time step is set to be 5 min in practice,so that resource view can cover 2 h.It is definedM=202 gates at an experimental airport and about 35% of them are fixed gates.The number of flights in each flight schedule isN=400 which lasts for one day.And 100 flight schedules are generated as training sample which means 40 000 flights data would be trained as an sample.The maximum simulation time of each episode is set asT=400 time steps.In each training literation,four episodic simulations are run for each flight schedule.Due to the characteristics of the problem,the step size of the gradient drop cannot be too large.If it is too large,the gradient will disappear.Learning rate is set as 0.0001.After the simulation of 500 iterations,the algorithm converges.

    We compared the performance of the proposed DRL-AGAP with 2 methods:Greedy and Guorbi.Greedy is the traditional algorithm to solve AGAP.For the current flights,it is assigned to the available fixed gate,regardless of its impact on subsequent flights.Gurobi is a new generation of large-scale mathematical programming optimizer developed by Gurobi Corporation[17].This software is an advanced method to solve AGAP at present.When using this optimization software,the AGAP is modelled as a mathematical programming problem.It is also worth noting that the AGAP solved by software Gurobi is a pre-assignment problem which assigns 400 flights at one time.

    Firstly,the performance of convergence is evaluated.The sample flight schedules the software Gurobi used to test the optimization performance are the same as the training samples of DRL-AGAP.

    As shown in Fig.5,the lines indicate the trend of the training process of the policy network.The original policy is a random assignment policy,with which only 35% of flights are assigned to the fixed gates (FG) and the rest 65% of flights are assigned to the remote gates (RG).After about 300 iterations of training,the ratio of flights assigned to FG has increased to more than 70%,which means the improvement of optimization objective is more than 100%.And the RG assignment rate reaches 28%.The range of discount is about 50% which means DRL-AGAP agent can converge the assignment result to the ideal result after training.

    From Fig.5,the optimized ratio of flights assigned to FG of optimization software Gurobi is slightly higher than the DRL-AGAP method.However,this is a very normal difference between the optimized performance of the global optimization method and the local dynamic optimization method.Generally,the global optimization result can be considered as an upper bound,and the proposed DRL-AGAP method can closely achieve this upper bound.The advantage of the DRL-AGAP is that it can adapt to different flight schedules when the policy network has been trained.

    Fig.5 Training performance comparation

    There are 100 new flight schedules generated to test the effectiveness of the policy network.As shown in Fig.6,the average FG assignment rate of DRL-AGAP is about 75%,and the average RG assignment rate of DRL-AGAP is about 28%.Compared with Greedy,the FG assignment rate is increased by nearly 10%,and the RG assignment rate is decreased by nearly 7%.Results also show that the ratio of flights assigned to FG using DRL-AGAP method is only 2% less than that using Gurobi.This is the same as the converged training performance.Meanwhile,the efficiency of using the proposed DRL-AGAP method is much better than using the software Gurobi.During the test,the assignment of 400 flights only needs to cost 29.29 s.But with Gurobi,for each new flight schedule,the global re-initialization is required which takes 1 910.49 s.The calculation speed has been increased by more than 65 times using DRL-AGAP.Calculation speed of the algorithm is improved.

    Fig.6 Test performance comparation

    Furthermore,the optimization software Gurobi models the problem as an optimization problem,which can only be pre-assigned according to the flight schedule.The DRL-AGAP proposed in this work models the problem as a Markov decision process,which can assign the gate dynamically according to current flights.

    5 Conclusions

    In this work,a DRL-AGAP method is proposed to deal with the dynamic AGAP in practice.The dynamic AGAP is modelled as an MDP to ensure real-time decision making.The optimization objective is designed as maximizing the rate of flights assigned to the fixed gates.Simulation results confirm that the proposed DRL-AGAP can significantly increase the optimization objective.Compared with the optimization software Gurobi,the optimization results are close.Meanwhile,the computational cost of the proposed DRL-AGAP is much less than that of Gurobi,which can be used as a real-time dynamic assignment method.As an initial research on AGAP with DRL,the ideal scenario is taken into consideration to validate the feasibility and the effectiveness of the method.In the future,more accurate constraints in practice will be considered to make this kind of DRL based methods actually usable and meet the real operational needs of airports.

    猜你喜歡
    趙家
    中國有色金屬工業(yè)協(xié)會原副會長、本刊理事會原理事長趙家生致賀本刊
    趙家祥教授
    觀察與思考(2022年3期)2022-04-22 10:32:20
    婚姻像海上行船 萬般努力只為去看更遠(yuǎn)更美的風(fēng)景
    中國商人(2018年9期)2018-09-14 08:05:36
    虱子
    喊門
    遼河(2015年11期)2015-11-28 03:36:15
    Development and Prospectives of Ultra-High-Speed Grinding Technology
    趙家寨煤礦二3煤層巷道掘進(jìn)新型支護(hù)研究探索
    河南科技(2014年18期)2014-02-27 14:14:46
    出版名家:趙家壁
    塵緣落幕:昨日的白馬,今天載了他人遠(yuǎn)行
    這一次的勇敢,是放棄你
    意林(2011年19期)2011-02-11 11:09:16
    国产精品一区二区三区四区久久 | 免费看美女性在线毛片视频| 国产免费av片在线观看野外av| 日本免费一区二区三区高清不卡| 又紧又爽又黄一区二区| 天堂影院成人在线观看| 十分钟在线观看高清视频www| aaaaa片日本免费| av片东京热男人的天堂| 亚洲第一欧美日韩一区二区三区| svipshipincom国产片| 国产成人欧美在线观看| 成熟少妇高潮喷水视频| 正在播放国产对白刺激| 国产在线精品亚洲第一网站| 成人一区二区视频在线观看| 久久 成人 亚洲| 成人永久免费在线观看视频| av免费在线观看网站| 一区二区三区激情视频| 亚洲第一av免费看| 精品国产乱码久久久久久男人| 操出白浆在线播放| 国产激情欧美一区二区| 身体一侧抽搐| 在线免费观看的www视频| 18美女黄网站色大片免费观看| 午夜福利成人在线免费观看| 99国产精品一区二区三区| 女人高潮潮喷娇喘18禁视频| a级毛片a级免费在线| 女警被强在线播放| 脱女人内裤的视频| 黄色 视频免费看| www.999成人在线观看| 搡老熟女国产l中国老女人| 精品少妇一区二区三区视频日本电影| 成人三级黄色视频| 久久中文字幕一级| 国内精品久久久久久久电影| 亚洲中文日韩欧美视频| 日韩欧美 国产精品| 久久99热这里只有精品18| 亚洲黑人精品在线| 色av中文字幕| 久久久久国产精品人妻aⅴ院| 国产精品久久久久久亚洲av鲁大| 视频区欧美日本亚洲| 久久久国产成人免费| 日本一本二区三区精品| 成人特级黄色片久久久久久久| 国产成人系列免费观看| 一进一出好大好爽视频| 亚洲精品中文字幕一二三四区| 99国产精品一区二区蜜桃av| 国产精品久久视频播放| 欧美黑人巨大hd| 黄片大片在线免费观看| 在线观看午夜福利视频| 午夜免费激情av| 亚洲五月婷婷丁香| 可以在线观看的亚洲视频| 久久久久精品国产欧美久久久| 日本熟妇午夜| 久久狼人影院| 老鸭窝网址在线观看| 欧美另类亚洲清纯唯美| 亚洲国产精品合色在线| 97碰自拍视频| 久久精品夜夜夜夜夜久久蜜豆 | 女同久久另类99精品国产91| 亚洲 欧美 日韩 在线 免费| 精品不卡国产一区二区三区| 悠悠久久av| 国产在线观看jvid| 欧美在线黄色| 老汉色av国产亚洲站长工具| 午夜亚洲福利在线播放| 日韩高清综合在线| 熟女少妇亚洲综合色aaa.| 色综合婷婷激情| 天天一区二区日本电影三级| 国产av在哪里看| 淫秽高清视频在线观看| 国产极品粉嫩免费观看在线| 欧美日韩亚洲国产一区二区在线观看| 手机成人av网站| 久久婷婷人人爽人人干人人爱| 欧美成人一区二区免费高清观看 | 女人被狂操c到高潮| 又黄又粗又硬又大视频| 国产不卡一卡二| av超薄肉色丝袜交足视频| 熟女少妇亚洲综合色aaa.| 国产精品98久久久久久宅男小说| 国产精品免费一区二区三区在线| 香蕉国产在线看| 日日夜夜操网爽| 嫩草影视91久久| 淫秽高清视频在线观看| 亚洲成国产人片在线观看| 久久精品亚洲精品国产色婷小说| 亚洲国产看品久久| 一边摸一边抽搐一进一小说| 精品第一国产精品| 成人av一区二区三区在线看| 露出奶头的视频| 亚洲成国产人片在线观看| 色老头精品视频在线观看| 国产av在哪里看| 男女之事视频高清在线观看| 国产精品久久电影中文字幕| 成人永久免费在线观看视频| 日韩精品免费视频一区二区三区| 日韩大码丰满熟妇| 看免费av毛片| 女人高潮潮喷娇喘18禁视频| 免费人成视频x8x8入口观看| 国内久久婷婷六月综合欲色啪| 亚洲三区欧美一区| 国产av不卡久久| 国产成人一区二区三区免费视频网站| 国产精品亚洲美女久久久| 男人舔女人下体高潮全视频| 久久精品91无色码中文字幕| 久久久久久久久免费视频了| 两性午夜刺激爽爽歪歪视频在线观看 | 国产97色在线日韩免费| 国产成人av激情在线播放| 波多野结衣巨乳人妻| 国产熟女xx| 国语自产精品视频在线第100页| 婷婷精品国产亚洲av在线| 黄色视频不卡| 国产视频内射| 美女国产高潮福利片在线看| 久久精品aⅴ一区二区三区四区| 嫁个100分男人电影在线观看| 日日爽夜夜爽网站| 国产三级在线视频| 色播亚洲综合网| 制服人妻中文乱码| 日韩免费av在线播放| 听说在线观看完整版免费高清| 久久久国产成人精品二区| 国产精品国产高清国产av| 国产精品久久久人人做人人爽| 青草久久国产| 国产成人精品久久二区二区免费| 亚洲国产毛片av蜜桃av| 麻豆av在线久日| 国产激情欧美一区二区| 在线视频色国产色| 桃红色精品国产亚洲av| 国产激情偷乱视频一区二区| 久久国产乱子伦精品免费另类| 精品久久久久久久末码| 一级毛片女人18水好多| 日本五十路高清| 国产成人精品无人区| 黑人欧美特级aaaaaa片| 色哟哟哟哟哟哟| 亚洲精品粉嫩美女一区| 91麻豆av在线| 日本黄色视频三级网站网址| 琪琪午夜伦伦电影理论片6080| 老司机午夜福利在线观看视频| 国产精品一区二区精品视频观看| 变态另类丝袜制服| 精品无人区乱码1区二区| 国产精品国产高清国产av| 亚洲男人天堂网一区| 女警被强在线播放| 手机成人av网站| 一二三四社区在线视频社区8| 黄色片一级片一级黄色片| 欧美日韩瑟瑟在线播放| 18禁国产床啪视频网站| 很黄的视频免费| 成年版毛片免费区| 国产av不卡久久| 午夜老司机福利片| 美女国产高潮福利片在线看| 亚洲熟妇熟女久久| 一卡2卡三卡四卡精品乱码亚洲| 中文亚洲av片在线观看爽| 99在线人妻在线中文字幕| 免费看日本二区| 一边摸一边做爽爽视频免费| 亚洲无线在线观看| 亚洲国产精品久久男人天堂| 天天躁夜夜躁狠狠躁躁| 久久精品国产亚洲av高清一级| 国产单亲对白刺激| 又黄又粗又硬又大视频| 亚洲真实伦在线观看| 成人亚洲精品一区在线观看| 欧美色视频一区免费| 欧美一级毛片孕妇| 一卡2卡三卡四卡精品乱码亚洲| 国产亚洲欧美精品永久| 精品卡一卡二卡四卡免费| 少妇粗大呻吟视频| 国产精品免费一区二区三区在线| 中亚洲国语对白在线视频| 国产97色在线日韩免费| 亚洲专区字幕在线| 国产精品影院久久| 亚洲熟妇熟女久久| 亚洲成av人片免费观看| 国产人伦9x9x在线观看| 一二三四在线观看免费中文在| 国产真人三级小视频在线观看| 精品国产乱码久久久久久男人| 久久精品国产99精品国产亚洲性色| 亚洲一区二区三区不卡视频| 亚洲 国产 在线| 欧美日韩中文字幕国产精品一区二区三区| 在线观看日韩欧美| 日日爽夜夜爽网站| 国产一区二区三区视频了| av欧美777| 久久久水蜜桃国产精品网| 中国美女看黄片| 黄色丝袜av网址大全| 日本熟妇午夜| 欧美av亚洲av综合av国产av| 极品教师在线免费播放| 成年女人毛片免费观看观看9| www日本在线高清视频| 久久人人精品亚洲av| 露出奶头的视频| 日本撒尿小便嘘嘘汇集6| 欧美乱色亚洲激情| 欧美日本亚洲视频在线播放| 露出奶头的视频| 非洲黑人性xxxx精品又粗又长| 50天的宝宝边吃奶边哭怎么回事| 国产免费男女视频| 亚洲成人久久性| 日本精品一区二区三区蜜桃| 桃色一区二区三区在线观看| 在线观看www视频免费| 国产黄色小视频在线观看| 午夜免费鲁丝| 久久精品夜夜夜夜夜久久蜜豆 | 操出白浆在线播放| 欧美成人午夜精品| 久久久精品欧美日韩精品| 国产又黄又爽又无遮挡在线| 男女午夜视频在线观看| 国产成人影院久久av| 中文字幕精品免费在线观看视频| 久久香蕉国产精品| 日本在线视频免费播放| 国产三级在线视频| av福利片在线| 午夜两性在线视频| 人成视频在线观看免费观看| 久久久久免费精品人妻一区二区 | 正在播放国产对白刺激| 97人妻精品一区二区三区麻豆 | 美女国产高潮福利片在线看| 成在线人永久免费视频| 亚洲 欧美一区二区三区| 少妇被粗大的猛进出69影院| 欧美丝袜亚洲另类 | 少妇 在线观看| 叶爱在线成人免费视频播放| 日日摸夜夜添夜夜添小说| 国产成人系列免费观看| 很黄的视频免费| 亚洲一区二区三区色噜噜| 男女视频在线观看网站免费 | 国产野战对白在线观看| 亚洲av第一区精品v没综合| 中文在线观看免费www的网站 | 精品少妇一区二区三区视频日本电影| 草草在线视频免费看| 国产一区二区三区在线臀色熟女| 国产国语露脸激情在线看| 精品久久久久久久人妻蜜臀av| 此物有八面人人有两片| 精品久久久久久久毛片微露脸| 99在线视频只有这里精品首页| 怎么达到女性高潮| 老司机靠b影院| 免费高清在线观看日韩| 亚洲欧美精品综合久久99| 手机成人av网站| 欧美性猛交黑人性爽| 国产v大片淫在线免费观看| av免费在线观看网站| 亚洲熟妇中文字幕五十中出| aaaaa片日本免费| 欧美日韩精品网址| 欧美激情 高清一区二区三区| 最好的美女福利视频网| 99久久无色码亚洲精品果冻| 精品一区二区三区av网在线观看| 丝袜在线中文字幕| 亚洲狠狠婷婷综合久久图片| 91老司机精品| 丰满的人妻完整版| 性欧美人与动物交配| 自线自在国产av| 国产精华一区二区三区| 一级黄色大片毛片| 欧美成人性av电影在线观看| 精品卡一卡二卡四卡免费| 变态另类丝袜制服| 日韩大尺度精品在线看网址| 久久精品国产清高在天天线| 亚洲男人的天堂狠狠| 99国产极品粉嫩在线观看| av福利片在线| 每晚都被弄得嗷嗷叫到高潮| 桃红色精品国产亚洲av| 国产精品亚洲一级av第二区| 中亚洲国语对白在线视频| 岛国在线观看网站| 嫩草影院精品99| 国产1区2区3区精品| 免费高清视频大片| 久久国产精品人妻蜜桃| 欧美午夜高清在线| √禁漫天堂资源中文www| 精华霜和精华液先用哪个| 亚洲一区高清亚洲精品| 国产野战对白在线观看| av片东京热男人的天堂| 亚洲真实伦在线观看| 国产成人av教育| 午夜视频精品福利| 欧美乱妇无乱码| 久久香蕉精品热| 色精品久久人妻99蜜桃| 老熟妇乱子伦视频在线观看| 两个人视频免费观看高清| 国产免费男女视频| 亚洲av第一区精品v没综合| 一本综合久久免费| 亚洲无线在线观看| 亚洲国产欧美网| 国产黄a三级三级三级人| 男女那种视频在线观看| 亚洲男人天堂网一区| 久久午夜综合久久蜜桃| 久久99热这里只有精品18| 亚洲国产中文字幕在线视频| 午夜福利欧美成人| 成人亚洲精品av一区二区| 国产成人系列免费观看| 91麻豆av在线| 男人舔女人下体高潮全视频| 欧美丝袜亚洲另类 | 国产精品自产拍在线观看55亚洲| 国产亚洲欧美精品永久| 伊人久久大香线蕉亚洲五| 国产亚洲欧美在线一区二区| 国产亚洲欧美98| 一本综合久久免费| 免费电影在线观看免费观看| 久久久国产成人精品二区| 国产午夜福利久久久久久| 免费女性裸体啪啪无遮挡网站| 91老司机精品| 久久国产精品男人的天堂亚洲| 在线观看免费午夜福利视频| 色老头精品视频在线观看| av有码第一页| 日韩欧美国产在线观看| 午夜福利18| 日韩大码丰满熟妇| 国产成+人综合+亚洲专区| 国产一卡二卡三卡精品| 真人一进一出gif抽搐免费| 国产免费男女视频| 亚洲熟妇熟女久久| 亚洲国产精品久久男人天堂| 国产男靠女视频免费网站| 欧美成狂野欧美在线观看| 女性被躁到高潮视频| 十八禁网站免费在线| 亚洲avbb在线观看| 人人妻,人人澡人人爽秒播| 色精品久久人妻99蜜桃| 亚洲国产欧美网| 久久天堂一区二区三区四区| 精品一区二区三区四区五区乱码| 天天一区二区日本电影三级| www.999成人在线观看| 亚洲第一欧美日韩一区二区三区| 日本黄色视频三级网站网址| 久久久国产精品麻豆| 国产精品久久电影中文字幕| 欧美不卡视频在线免费观看 | 香蕉国产在线看| 黑人欧美特级aaaaaa片| 亚洲自拍偷在线| 欧美成人一区二区免费高清观看 | 亚洲,欧美精品.| e午夜精品久久久久久久| 亚洲精品在线美女| 免费搜索国产男女视频| 国产伦一二天堂av在线观看| 亚洲精品美女久久av网站| 亚洲国产中文字幕在线视频| av中文乱码字幕在线| 午夜福利视频1000在线观看| 国产精品国产高清国产av| 日韩视频一区二区在线观看| 精品一区二区三区av网在线观看| 欧美在线黄色| 老司机午夜十八禁免费视频| 色哟哟哟哟哟哟| 久久天躁狠狠躁夜夜2o2o| 香蕉久久夜色| 国产精品免费一区二区三区在线| 久久青草综合色| 久久久国产精品麻豆| 国产熟女xx| 日日干狠狠操夜夜爽| 久久久久国产精品人妻aⅴ院| 亚洲精华国产精华精| 国产精品免费视频内射| 国产欧美日韩一区二区精品| 国产av在哪里看| 婷婷丁香在线五月| 亚洲激情在线av| 国产在线精品亚洲第一网站| 成人亚洲精品一区在线观看| 色婷婷久久久亚洲欧美| 十八禁人妻一区二区| 侵犯人妻中文字幕一二三四区| 亚洲天堂国产精品一区在线| 中文字幕人妻丝袜一区二区| 色综合婷婷激情| 一夜夜www| 亚洲精品一卡2卡三卡4卡5卡| 91国产中文字幕| 妹子高潮喷水视频| 人人妻人人澡人人看| 久久国产乱子伦精品免费另类| 99热这里只有精品一区 | 99久久99久久久精品蜜桃| 亚洲精品粉嫩美女一区| 在线播放国产精品三级| 国产在线观看jvid| 国产成人一区二区三区免费视频网站| 国产一级毛片七仙女欲春2 | 亚洲三区欧美一区| 亚洲五月天丁香| 两性午夜刺激爽爽歪歪视频在线观看 | 男人舔女人下体高潮全视频| 欧美成狂野欧美在线观看| 老司机在亚洲福利影院| 露出奶头的视频| 日日爽夜夜爽网站| 国产精品久久久av美女十八| 热99re8久久精品国产| 99久久久亚洲精品蜜臀av| 禁无遮挡网站| 一边摸一边做爽爽视频免费| 母亲3免费完整高清在线观看| 国产免费男女视频| 精品国产乱码久久久久久男人| 色综合婷婷激情| 成人免费观看视频高清| 50天的宝宝边吃奶边哭怎么回事| 国产精品,欧美在线| 亚洲精品国产一区二区精华液| 搡老岳熟女国产| 欧美中文日本在线观看视频| 人人妻人人澡欧美一区二区| 精品无人区乱码1区二区| 91字幕亚洲| 女警被强在线播放| ponron亚洲| 操出白浆在线播放| 18禁观看日本| 久99久视频精品免费| 老熟妇仑乱视频hdxx| 两性夫妻黄色片| 久久久久免费精品人妻一区二区 | 日韩大码丰满熟妇| 成年版毛片免费区| 人成视频在线观看免费观看| 757午夜福利合集在线观看| 免费看美女性在线毛片视频| 国产成人精品久久二区二区免费| 久久久久久国产a免费观看| 可以在线观看毛片的网站| 亚洲精品粉嫩美女一区| 亚洲 欧美一区二区三区| 一二三四社区在线视频社区8| 高潮久久久久久久久久久不卡| 亚洲,欧美精品.| 黑人欧美特级aaaaaa片| 免费在线观看完整版高清| 观看免费一级毛片| 亚洲av美国av| 国产aⅴ精品一区二区三区波| 久久婷婷人人爽人人干人人爱| 国产精品永久免费网站| 色综合亚洲欧美另类图片| 国产高清videossex| 国产一卡二卡三卡精品| 九色国产91popny在线| 亚洲av美国av| 人妻丰满熟妇av一区二区三区| 日韩欧美三级三区| 亚洲欧美日韩高清在线视频| 国产欧美日韩精品亚洲av| 国产伦一二天堂av在线观看| 亚洲五月婷婷丁香| 无遮挡黄片免费观看| 久久午夜综合久久蜜桃| av在线天堂中文字幕| 国产视频一区二区在线看| 很黄的视频免费| 在线永久观看黄色视频| 波多野结衣巨乳人妻| 男女视频在线观看网站免费 | 欧美性猛交黑人性爽| 精品福利观看| 美女 人体艺术 gogo| 国产精品久久久av美女十八| 淫妇啪啪啪对白视频| 国内毛片毛片毛片毛片毛片| 动漫黄色视频在线观看| 国产精品亚洲av一区麻豆| 亚洲男人天堂网一区| 欧美乱色亚洲激情| 一本一本综合久久| 美女大奶头视频| 国产在线精品亚洲第一网站| 搡老妇女老女人老熟妇| 999久久久精品免费观看国产| 国产成人欧美在线观看| 天堂动漫精品| 免费在线观看视频国产中文字幕亚洲| 大香蕉久久成人网| 国产在线观看jvid| 99精品在免费线老司机午夜| 母亲3免费完整高清在线观看| 伊人久久大香线蕉亚洲五| 国产精品永久免费网站| 国产精品自产拍在线观看55亚洲| 亚洲三区欧美一区| 老司机午夜福利在线观看视频| 一二三四社区在线视频社区8| 女性被躁到高潮视频| 精品熟女少妇八av免费久了| 午夜福利欧美成人| 男男h啪啪无遮挡| 久久精品成人免费网站| 欧美日本亚洲视频在线播放| 淫妇啪啪啪对白视频| 欧美成人性av电影在线观看| 精品电影一区二区在线| 国产片内射在线| 亚洲九九香蕉| 身体一侧抽搐| 国产精品久久电影中文字幕| 一个人免费在线观看的高清视频| 亚洲熟女毛片儿| 国产区一区二久久| 国产精品二区激情视频| 亚洲自偷自拍图片 自拍| 成年人黄色毛片网站| 午夜福利欧美成人| 热re99久久国产66热| 十八禁人妻一区二区| 亚洲国产精品sss在线观看| 亚洲 欧美一区二区三区| 一个人观看的视频www高清免费观看 | 桃色一区二区三区在线观看| 亚洲成av片中文字幕在线观看| 欧美激情久久久久久爽电影| 日韩有码中文字幕| 亚洲美女黄片视频| 国产av一区在线观看免费| 天堂√8在线中文| 国产成人精品久久二区二区免费| 中文字幕另类日韩欧美亚洲嫩草| 白带黄色成豆腐渣| www.www免费av| 每晚都被弄得嗷嗷叫到高潮| 夜夜夜夜夜久久久久| 国产成人系列免费观看| 色综合欧美亚洲国产小说| 黄色女人牲交| 国产精品一区二区免费欧美| 啦啦啦 在线观看视频| 欧美日本视频| 欧美在线一区亚洲| 动漫黄色视频在线观看| 精品一区二区三区四区五区乱码| 一级a爱片免费观看的视频| 欧美最黄视频在线播放免费| 欧美成人一区二区免费高清观看 | 18禁观看日本| 久久热在线av| 亚洲aⅴ乱码一区二区在线播放 | 欧美日韩乱码在线| 日韩免费av在线播放| 最新美女视频免费是黄的| 国产成人精品无人区| 午夜免费成人在线视频| 免费在线观看日本一区| 视频在线观看一区二区三区| 一级a爱片免费观看的视频| 在线国产一区二区在线| 精品欧美国产一区二区三|