• <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 ?

    Hierarchical Reinforcement Learning With Automatic Sub-Goal Identification

    2021-10-25 01:41:22ChenghaoLiuFeiZhuQuanLiuandYuchenFu
    IEEE/CAA Journal of Automatica Sinica 2021年10期

    Chenghao Liu,Fei Zhu,Quan Liu,and Yuchen Fu

    Abstract—In reinforcement learning an agent may explore ineffectively when dealing with sparse reward tasks where finding a reward point is difficult.To solve the problem,we propose an algorithm called hierarchical deep reinforcement learning with automatic sub-goal identification via computer vision (HADS)which takes advantage of hierarchical reinforcement learning to alleviate the sparse reward problem and improve efficiency of exploration by utilizing a sub-goal mechanism.HADS uses a computer vision method to identify sub-goals automatically for hierarchical deep reinforcement learning.Due to the fact that not all sub-goal points are reachable,a mechanism is proposed to remove unreachable sub-goal points so as to further improve the performance of the algorithm.HADS involves contour recognition to identify sub-goals from the state image where some salient states in the state image may be recognized as sub-goals,while those that are not will be removed based on prior knowledge.Our experiments verified the effect of the algorithm.

    I.INTRODUCTION

    DEEP reinforcement learning,controlled by input information,similar to the decision mode of humans,is a combination of deep learning and reinforcement learning [1].Since Manih et al.introduced the deep Q network (DQN) [2],applications of deep reinforcement learning have continued to increase and many improved DQN algorithms have been proposed,such as double Q-learning [3],the Priority Replay Buffer [4] and the Dueling Deep Q Network (DuDQN) [5],which solve the control problem in the continuous action space with a human-level performance.

    However,these algorithms perform poorly in tasks with sparse rewards,e.g.,the Montezuma game where agents barely find reward points.In order to solve these problems,hierarchical reinforcement learning is used to deal with such problems,such as with the abstract method proposed by Suttonet al.[6],where abstraction was used to decompose large-scale problems into smaller ones and complex problems into simpler ones.After the agent solves all the decomposed problems,the complex large problem is then correspondingly resolved.Hierarchical deep reinforcement learning,which takes advantage of the concept of hierarchy in deep reinforcement learning,consists of hierarchical reinforcement learning and deep learning.Kulkarniet al.proposed hierarchical deep reinforcement learning (hDQN) [7],which alleviates the sparse reward problem by constructing hierarchical decision models and setting up small goals.Andrychowiczet al.proposed hindsight experience replay(HER) [8],which made the agents’ learning relatively smooth by establishing relatively feasible goals.Florensaet al.introduced a stochastic network for hierarchical reinforcement learning [9],which divided tasks into two layers and learned useful skills in a pretrained environment.Leet al.put forth hierarchical imitation reinforcement learning to solve the problem of sparse rewards by imitating the behavior of human experts [10].

    In hierarchical deep reinforcement learning,sub-goal identification is a crucial issue,where a well chosen sub-goal speeds up learning [11].Methodsto find sub-goals in hierarchical reinforcement learning include determining subgoals based on the state access frequency,recognizing subgoals based on the state transition graph or common subspace,etc.Rafati and Noelleproposed a model-free algorithm for detecting sub-goals associated with the intrinsic motivation learning mechanism,learning sub-goals and skills based on the experience from the environment [12].Imani and Braga-Neto used a partially observable boolean dynamic system(POBDS) signal model to measure the time sequence of noise expression in Boolean GRN and developed the Bayes inverse reinforcement learning (BIRL) [13].Dilokthanakulet al.proposed an algorithm with the advantage of learning representations from intrinsic rewards to improve the performance of HER [14].However,the majority of approaches require defining sub-goals manually,which not only causes poor performance in complex problems,but also results in poor generalization ability,difficulty in determining sub-goals in large space,and large deviations between manual methods and other methods.

    To address the problem,we propose a hierarchical deep reinforcement learning algorithm with automatic sub-goal identification via computer vision (HADS),which obtains the sub-goals from pixels.One of the goals of automatic control of hierarchical deep reinforcement learning is to detect subgoals without any human intervention.However it is very hard to achieve at present.Standard approaches for detecting subgoals require heavy computation while the performance is not as good as expected.Some points are in fact unreachable,such as points behind the walls,as well as fringe points.As for obstacles,if they are identified as sub-goal points,the agent gets a positive reward when reaching these points,while according to most settings the agent gets a negative reward,which is a contradiction.Correspondingly,obstacle points are also removed.Removing these points helps to improve the performance.Although our ultimate goals are to provide automatic control,it is worth adding a little bit of human knowledge to enhance behavior of the algorithm.The paper is organized as follows.In Section II,we introduce preliminaries,including reinforcement learning,hierarchical reinforcement learning and hierarchical deep reinforcement learning.In Section III,we expand the model and the processing workflow of the HADS algorithm.In Section IV,we explain the experiments and analyze the results of experiment.In Section V,we conclude our work.

    II.PRELIMINARIES

    A.Reinforcement Learning

    Reinforcement learning,which is different from supervised learning and unsupervised learning,is a trial-and-error method in which the agent receives a reward signal in exploration process rather than receiving timely feedback [15].Reinforcement learning consists of four elements,including the state,action,reward,and environment,where for the environment,the agent interacts with it to obtain a higher reward;the state refers to the single state in the state space that the agent is in for each time step;the action refers to the single action in the action space that the agent adopts for each time step;the reward refers to the reward the agent receives for adopting a certain action.

    A reinforcement learning problem can be modelled with a Markov decision process (MDP) by a tuple〈S,A,P,R,γ〉whereSis a state set,Ais an action set,Pis the state transition probability,Ris a reward function,and γ ∈[0,1] is a discount factor.A policy π is a distribution over actions given states,π(a|s)=P[At=a|St=s].Q-learning [16],a classical algorithm of reinforcement learning,where temporaldifference (TD) [1] is utilized,uses aQvalue to evaluate the action under some certain state.Q(st,at),the expected cumulative reward of taking actionatin statest,is calculated as follows:

    where α is the step size parameter,Rt+1is the immediate reward and γ is the discount rate.In this case,theQ-value function that is near the optimal state action-value function is denoted asQ?.

    The partially observable Markov decision process(POMDP) is a variation of MDP in which the POMDP model only has part of the information.POMDP is represented by a tuple(S,A,P,R,Ω,O,γ),where Ω is a group of result sets which contain the environment data collected by the sensor,andOis the situation-observation probability,that is,the probability the agent is under stateswhen it observes the environment datao.In POMDP,since the information is inadequate,the history sequence information is utilized to estimate the state,referred to as the belief state.

    Recently,deep reinforcement learning (DRL),which combines deep learning with reinforcement learning,has achieved much success.Deep reinforcement learning algorithms usually learn control strategies from highdimensional raw data directly.In particular,as a widely used deep reinforcement learning algorithm,DQN takes advantage of a convolutional neural network (CNN) [17] to get state information from pixel data and makes a decision by using Qlearning.

    B.Hierarchical Reinforcement Learning

    Hierarchical reinforcement learning is an extended framework of common reinforcement learning [18],e.g.,HAM [19],Options [20],MAXQ [21],and HER [8],which decompose the original tasks into smaller and simpler subtasks so as to form a hierarchical structure.Learning tasks are assigned to each sub-task,which,according to reinforcement learning methods,explores policy in the constrained task hierarchical space [22].As soon as the policy exploration of each sub-task is completed,the policy of the original task is also achieved.

    Abstraction [20] is a key issue in hierarchical reinforcement learning,by which some irrelevant or redundant information from different sub-tasks are removed so as to simplify the problem.There are various abstractions in hierarchical reinforcement learning,including state abstraction (space abstraction) [23],task abstraction [24],and time abstraction(action abstraction) [25].Space abstraction can be divided into two cases:one is to encapsulate several states so as to form a single abstract state;the other is to simplify the state representation by deleting state variables independent of substates in the case of a decomposable state.Task abstraction decomposes the main learning task into several sub-tasks by some evaluation index,solves the sub-tasks separately by which the whole problem can be solved.Time abstraction encapsulates several actions as a single abstract action,of which the purpose is to make full use of the hierarchical structure of the problem.Each abstract action corresponds to a closed-loop local policy defined on a subset of environmental states with termination conditions,which is called Option or Temporally-extended actions that is a framework based on time abstraction [20].

    Abstractaction is an important concept in the Option framework,which uses prior knowledge to speed up learning.It is known that the reuse of a learned model can speed up learning.In practice,it is of very little value to have two tasks that are exactly the same.Therefore,it is usually unable to achieve reuse at the entire task level.In fact,it is more likely that two tasks are partially similar,making it is possible to apply reuse at a more subtle level,i.e.,to reuse policy of subtasks rather than the entire task.Although the sharing policy of the sub-task has many constraints,it really provides a feasible way for the present.The abstract action includes the basic action defined in the corresponding problem domain.A reinforcement learning agent can select an abstract action in a certain state and then execute the abstract action according to the policy,until the termination condition is satisfied.

    C.DRL With Time Abstraction

    The difference between deep reinforcement learning with time abstraction [26] and standard deep reinforcement learning is that deep reinforcement learning with time abstraction appends a goalgtoQvalue.The correspondingQestimates of the controller are

    wheregis the goal at state s,πagis the policy currently adopted,andErepresents expectation.TheQestimates of the meta-controller are also derived from the previously aforementioned equations

    whereNis the time steps required for the controller to achieve its goal,g′represents the agent’s goal atst+N,and πg(shù)is the policy adopted by the meta-controller to reach the goalg[27].

    In DQN,theQvalue is constantly updated to approach the optimal [28] value,and theQvalue with targetgcan also be expressed by adding parameter θ:Q?(s;g)≈Q(s,g;θ).Similar but different from DQN,eachQ∈{Q1,Q2} needs to be trained to minimize loss functionsL1(θ1) andL2(θ2) [29].The interaction sequences(st,gt,ft,st+N) ofQ2and (st,at,gt,rt,st+1)are stored in corresponding replay buffersD1andD2.The loss function ofQvalue is defined as mean squared error.The loss function ofQ1is similar to that of DQN.

    whereirepresents the number of iterations,y1,iis determined by (8),andErepresents expectation.When the loss function is updated during the optimization stage,the parameters of the previous iterations,θ1,i?1remain unchanged.The gradient is used to optimize the para(meter θ1.

    Loss functionL2and its gradient updating parameters [30]can also be obtained in the same way.

    D.Improving Exploration by Sub-Goal Identification

    Exploration and exploitation are important issues in reinforcement learning [31],in which,given a single policy agent,during exploration,one tries to find a larger search space so as to get a better policy,bringing a higher cost but not necessarily a higher success rate;during exploitation,the agent adopts the optimal policy with the highest estimated reward at present.Performing too much exploration requires lots of computation despite the fact that it provides a higher probability of attaining a better global or near-optimum solution; although excessive exploitation speeds up convergence,it often leads to a local optimum.The efficiency of an algorithm lies on the fine balance of exploration and exploitation [32].

    In a Markov process,effective exploration is indispensable to learn a better control policy [33],which,is a major challenge.Although some methods have been proposed,searching for more effective exploration methods has been an ongoing direction in reinforcement learning.In the study of reinforcement learning,the most frequently used policy is ? ?greedy,which plays an important role in common reinforcement learning,because it chooses the best policy for the time being,taking into account the exploration to prevent falling into a local optima.However,due to the sparsity of rewards and the large state space,common methods often fail to find out the final goal [34].

    By exploring a series of internal goals set in advance,the agent achieves the goal during which it is necessary to learn the corresponding policy for that goal.An internal reward [35]is used to determine whether the agent has reached the goal point,and to optimize the corresponding policy.In deep hierarchical reinforcement learning,Kulkarniet al.[7]proposed a framework for deep hierarchical learning (hDQN),where agents adopted a two-layer hierarchical structure,which was composed of the meta-controller and the controller.The meta-controller interacts with the environment to get the statest,and then chooses a goalgt∈G,whereGis the current set of all possible goals.

    After the meta-controller chooses the goalgt,the controller chooses the actionataccording to the current statestandgt.It requires certain time steps for the agent to reachgtor the final state.As shown in Fig.1 [7],the meta-controller chooses a goalgt,and in the next period,the controller tries to reachgt.In the framework of hDQN,there is an internal critic module[36],which verifies whether the controller has reached the corresponding goalgtand returns an appropriate internal rewardrt(g) to the controller.If the controller reaches the corresponding goal,it gets a reward of 1 or a reward of 0 otherwise.The goal of the controller is to maximize the cumulative internal rewards,Similarly,the goal of the meta-controller is to maximize the cumulative external rewards,whereis the cumulative reward for interaction with the external environment in time stept,whereare the rewards after finishing a series of sub-goals.

    Fig.1.Meta-controller selects the goal g,and the controller completes the goal g.The meta-controller selects a sub-goal g that is supposed to visited by controller.At state st,the sub-goal gt is generated by the meta-controller for the controller.The controller gets to gt after N steps.The procedure will be carried on iteratively with the task is completed.

    As soon as the sub-goals are completed,the agent can successfully reach the end point,and be rewarded.In order to let the agent reach the sub-goal with a better method,the controller is optimized by incorporating the internal reward function,so the agent can reach the sub-goal quickly and finally reach the end point smoothly.

    III.ALGORITHM DESCRIPTION

    On the basis of the original hDQN algorithm,the function of automatic sub-goal identification is incorporated,which makes the original hierarchical reinforcement learning algorithm more intelligent by identifying sub-goals from the image.By capturing states from the state image,the algorithm searches for the sub-goals,and improves them to so that they are higher quality by taking advantage of prior knowledge.Then the sub-goals and the image are put into a neural network to be trained together.As there are some sub-goals that are unreachable,we need to incorporate criteria to tailor original sub-goals.First,as the fringe parts are unreachable,the fringe sub-goal points should be filtered;second,as for obstacle points,the agent will receive a negative reward when reaching obstacles,and will receive a positive one when reaching a sub-goal,which is a contradiction;third,some subgoal points that are helpless to the mission are filtered.

    Fig.2 shows the illustration of the interaction between the agent and environment,where the environment provides the reward as well as a sub-goal set.When the controller achieves the target,an internal reward is provided by the critic.The environment generates a set of sub-goals for the metacontroller according to its understanding of the environment from Fig.3.The agent receives the current state,the metacontroller chooses the goal according to the received state,and the controller generates the action according to the received state and the selected goal.Once the controller reaches the goal chosen by the meta-controller,the critic will give the controller a positive internal reward.Otherwise,if the goal is not reached,the reward given by the critic is 0,and the reward given by the external environment is negative.The final goal of the meta-controller is the end point that the agent needs to reach.The controller achieves the last goal when the agent reaches the end point.

    Fig.2.In the model framework,for the interaction between the agent and environment,the environment provides not only the outer reward but also a sub-goal set.When the controller achieves the target (as stated previously),an internal reward is provided by the critic part.

    Fig.3.The generation of sub-goals where (a) represents the raw image as the input;(b) represents the gray-scale image from the raw image;(c) represents the binary image from the gray-scale image;(d) represents the detected sub-goal based on (c);and (e) represents the processed sub-goal points.

    A.Automatic Sub-Goal Identification

    Like most standard deep reinforcement learning algorithms,our method takes images as initial input to extract sub-goals.The state image needs to be processed at first because the original image contains much redundant information.Preserving valuable pixels,the image is cut to the fixed size because the network requires pictures of uniform size.After obtaining the image,we adopt a method in which the twodimensional Gauss template is used to perform smoothing filtering to remove the noise of the image.The smaller scale template is used to ensure the accuracy of locating the contour of the object in the image on a large scale,which will blur the edge of the contour of the object and affect the subsequent detection.Then the smoothed image is processed using object edge detection,and the preliminary response image is obtained [37].At this time,a blurred location of the contour of the object’s edge has been obtained,and the final step is to locate it accurately.The non-maximum suppression method is used to refine the image processed above,in which only the points with the greatest local variation are retained,and the edge image of the single-pixel object is obtained.According to the binary image,after the contour position of the object is obtained,the coordinates are obtained.

    There are several coordinates obtained,each corresponding to a set of coordinates,but the sub-goal points used in deep learning are not a group,for which a set of coordinates need to be processed.However,there are still two problems concerning the obtained target points.The first problem is that the coordinates of the target points are a group rather than a single point,and the data types are not consistent with the ones used in the algorithm.The second problem is that these target points are not necessarily what we need.Therefore,in order to solve the problem of a set of coordinates in the algorithm,the average value of the set of coordinates is taken as its corresponding sub-goal points.For the data type problem,the obtained goal points are converted into data types compatible with the network without changing their original meanings.On the basis of prior knowledge,some sub-goal points which can not be explored by an agent as well as some meaningless sub-goal points are removed to ensure that the obtained goal points are reachable and meaningful.For example,in snapshots of game controlling,part of the image represents the life value of agents,which is unimportant to an agent.In the exploration process,there exists some places that are beyond the reach of agents,and thus would be regraded as meaningless points,and be removed from subgoals.In addition,according to the initial position of the agent in the game,we rank the distance from the target point to the initial position of the agent.Let the agent explore these subgoal points from the near ones to the distant ones in the initial stage of exploration,which is equivalent to providing a relatively good policy for the agent at the beginning.However,the algorithm still uses the ? ?greedypolicy to explore,utilizing the provided policy to attempt to achieve better result.

    The size of the adjusted image is 84×84×1 or84×84×3 and the former is more easily processed.We use the standard method to process gray-scale maps so as to convert the image to a Gray map byGRAYmap=0.3×R+0.59×G+0.11×BwhereRis red,Gis green,andBis blue.In Algorithm 1 the image is processed to obtain sub-goals where the input image for the training network is raw RGB images.We utilizefindContoursin OpenCV to get the contour that is needed in the experiment from the binary-value image.

    B.Hierarchical Deep Reinforcement Learning With Sub-Goals

    The obtained sub-goals are applied to hierarchical deep reinforcement learning described,as follows.

    The algorithm uses dual network structures,which are different from the standard network structure in that our algorithm has sub-goal points.The corresponding network structure of the controller is to store samples with goal points in the corresponding replay buffer at each time step.The network structure corresponding to the meta-controller is that when the goal selected by the meta-controller is completed,the samples obtained by the meta-controller are stored in the replay buffer corresponding to the meta-controller.In this way,a hierarchical structure is constructed.The network structure corresponding to the controller is shown as Fig.4.The network structure corresponding to the meta-controller is similar to that of Fig.4.Since the framework in this paper is based on Deep Reinforcement Learning,the complexity of the framework is equivalent with that of the framework decided by the numbers of convolution layers and full-connected layers.Like DQN network structure,our proposed network contains 3 convolutional layers and 2 full-connected layers,solving more complex problems (the input includes sub-goals)at the cost of higher expenses.The time complexity of DQN is approximatelyO(E×B×T),whereErepresents the number of plots,Brepresents the batch size,andTrepresents the time complexity of the network.On the basis of DQN,our algorithm adds the process of finding sub-goals.Since the time complexity of finding sub-goals algorithm isO(n),the time complexity of our algorithm is approximatelyO(E×B×T)+O(n).

    Fig.4.Diagram of network structure where the image and sub-goal set are taken as the input of the network of 3 hidden layers and 2 fully connected layers with the output being a Q value with the sub-goal g.

    IV.ExPERIMENTS AND ANALYSIS

    There are lots of studies on deep reinforcement learning.The DQN algorithm is a benchmark approach of deep reinforcement learning;the DuDQN algorithm is an improved algorithm based on the DQN algorithm;the hDQN algorithm is based on hierarchical reinforcement learning with a neural network for solving large-scale continuous problems.We compared performance of our proposed algorithms with DQN,DuDQN,and hDQN.In some areas,although manually selecting sub-goals does bring a better result,it depends on human efforts and has too many constraints.In discrete areas there are various approaches for automatically searching subgoals but in continuous areas there are no such approaches.Thus,we used the Computer Vision approach to search for sub-goals.But because of the complexity of the environment,it still requires a little manual intervention.

    A.Montezuma’s Revenge

    As is shown in Fig.5,in Montezuma’s Revenge,the agent needs to find the key and use it to open the door.The agent will get a reward when it gets the key,and another reward when it gets the key and opens the door.This is characterized by features of sparse reward problems.Therefore,in order to simplify the problem,without changing the final solution,the agent is regarded as completing the task when it gets a reward.In this game,we altered the reward criteria used by DQN and other algorithms with the reward being assigned 0 for failure and 1 for succeeding in executing sub-goals.The reward setting does not essentially change the problem,and better demonstrates the result.In the experiments,we stipulate that the reward is 0 for each step of the agent,–1 for failure,or 5 for getting the external reward.Since the rewards are sparsely distributed,in most cases the agent can not find the reward points;if the reward is not altered,the result often appears to be a horizontal line,inconvenient for analysis.The learning rate is 0.00025 and the discount factor is 0.99.Although there are many tools,such as AutoML [38],that enable the agent to automatically tune hyper-parameters so as to get better performance,we did not take advantage of them in order to evaluate approaches in a fair way,because we need to use the same hyper-parameter settings used in Kulkarniet al.[7].

    Fig.5.Illustration of Montezuma’s Revenge.

    The sub-goal is obtained by Algorithm 1,which is shown in red dot in Fig.6(a),where it can be seen that,when the unreachable sub-goal points are not removed,the number of original sub-goal points is too large.If agents explore based on these sub-goal points,the performance of the algorithm cannot be improved.Therefore,according to some prior knowledge,we need to remove those points,such as the number of life values on the top of the graph and the agent itself.The number of life values that is used to indicate the remaining number of lives of the agent can not be explored.Similarly,some other unexplored points are also removed,and the resulting figure is shown in Fig.6(b),where the sub-goal points are much less than those in the top graph,and they are all accessible sub-goal points.Some reachable sub-goal points are helpless to the mission and are also filtered,such as the door.

    Fig.6.Original generated sub-goal graph and sub-goal graph after removing unreachable sub-goal points.In (a),the blurred red dots are initial sub-goal points,with some of them being unattainable for the agent and some of them being non-real sub-goal points,that is,on the agent itself.In (b),the red rectangle marks the actual sub-goal points.

    As is shown in Fig.6(b),in the middle of the agent and the ladder on the right,we find the sub-goal point,that is,the first sub-goal point explored by the agent which will be much faster than the first sub-goal point explored by the hDQN.The sub-goal points found in HADS are a little more than those manually identified sub-goal points.It is undeniable that the manually defined sub-goal points are very important,but adding some sub-goal points between these important subgoal points can improve the speed of reaching these important sub-goal points.

    The number of plots of the experiment is set as 3000 and the maximum number of steps per plot is set as 5000.In a plot,if the agent takes 5000 steps and is still alive,or if the agent does not find a reward point,the plot ends.Similarly,the experiment also sets up 20 000 steps for the agent to explore randomly.In these 20 000 steps,the network is not used to train.At the end of the random exploration,we began to use the network to train,and finally obtain the cumulative reward of all plots as shown in Fig.7,the reward of each plot returned is shown in Fig.8,and the average reward of the 3000 plot is shown in Table I.

    Fig.7.Cumulative external rewards.The result trend of DQN overlapped with that of DuDQN.HADS got a better convergence rate,as it has more subgoals than hDQN.

    Fig.8.Reward for each plot.HADS is better than the other three approaches,because sub-goals alleviated the sparse problem which made DQN and DuDQN incapable of solving the task.

    TABLEI AVERAGE REWARD FOR THE LAST 5 PLOTS

    Table I demonstrates that the average rewards of hDQN and HADS for the 3000th plot are superior to those of DQN and DuDQN.Fig.7 indicates that DQN has not explored the reward point in the searching process until the end of the plot,and each time the value of life is used up,no reward point is found.It can be seen from the Figs.7 and 8,that the result trend of DQN overlapped with that of DuDQN.These two algorithms do not explore reward points.However,in the original data,it is found that DuDQN still has reward points,although very few.In hDQN,the three manually defined subgoal points are the right ladder,the left ladder and the key in Fig.5,sequentially,and the three points are set with increasingly internal rewards.The hDQN performed much better than DQN,and DuDQN,while HADS surpassed hDQN,as can be seen from Table II,where we listed the average value (Avg),maximum value (max),minimum value(min),and mean square deviation (MSD).Since Fig.7 is a graph of cumulative rewards,the trend of each algorithm is shown.Because we do not know if there are significant differences between these values,we draw the mean value chart and various tables for comparison and the MSD value is also given in the tables.Because DQN and DuDQN do not get rewards in the current environment,the MSD value of both algorithms is 0.The reason why HADS is slightly higher than hDQN is that the sub-goal points in hDQN are given manually,so hdqn will be slightly lower than HADS.

    TABLEII RESULTS IN THE MONTEzUMA’S REVENGE

    As can be seen from Fig.8,before 700 plots,hDQN and HADS received almost the same reward for each plot,and after 700 plots,the reward for each plot of HADS was higher than that of hDQN,which shows that HADS had an advantage over hDQN,DQN,and DuDQN.From Table III,we can see that although our proposed algorithm,HADS,was not as good as an expert player,it really achieved automatic control and,more importantly,showed a better trend.We believe that with more subtle training,HADS could attain even better results.In Fig.8,the dotted line is the real reward for each plot,while the solid line is the average reward for 30 episodes.Therefore,according to the effect in the figure,HADS shows the best experimental effect.

    TABLEIII RESULTS WITH ExPERT PLAYER FOR 30 PLOTS

    B.Kangaroo

    Kangaroo is also a strategy game,which aims to find rewarding points.Unlike in Montezuma’s Revenge,in Kangaroo the reward is not sparse,which makes it a bit different from reward setting in Montezuma’s Revenge:if the reward point is found,the reward point will be set to 1;otherwise,if the agent fails,the reward is set to 0.An illustration of Kangaroo is shown in Fig.9.

    Fig.9.Illustration of Kangaroo.

    In the experiment,the sub-goals are obtained by algorithm 1 first.Kangaroo’s original sub-goals are highlighted in the red dot in Fig.10(a).If the original unreachable sub-goal points are unchanged,it is more difficult to improve performance as there are many sub-goal points which requires much computational cost.Since some goal points are not explored by the agent,it is logical to eliminate them.Therefore,according to prior knowledge,some points were removed,such as the staircase-like object on the left side,the number of life values,and the score attained in the bottom.As can be seen,the sub-goal points are much less than those in left-hand graph,and they are all accessible sub-goal points.

    The experiment takes 3000 plots,where the maximum number of steps for each plot is 5000,i.e.,if an agent survives 5000 steps in a plot,or if the agent has not explored the reward point,the plot will end.Similarly,the experiment sets up 20 000 steps for algorithms to explore randomly.The algorithms do not start training the network until the end of the exploration stage.Finally,a cumulative reward for all the plots is recorded,as shown in Fig.11,a reward for each plot as shown in Fig.12,and an average reward for the 3000 plots as shown in Table IV.

    As can be seen from Table IV,the average rewards of hDQN and HADS for the 3000th plot are better than those of DQN and DuDQN.From Fig.11,we can see that DQN performs well,making progress steadily.Although DuDQN is also growing better,it is not as good as DQN in terms of training results.The improvement of DuDQN over DQN is not decisive in Kangaroo.hDQN did not perform well in earlier stage because hDQN needs two networks to be trained,which requires more time and samples.After training for a certain plots,the strong growth rate of hDQN proved its strength.HADS has a similar result to that of hDQN.Compared with hDQN,the HADS algorithm pays much attention to automatic discovery of sub-goals.In the process of sub-goal identification,only a few manual intervention was incorporated during sub-goal detection,the sub-goal identified by HADS are better than the sub-goals automatically defined by hDQN,and Fig.11 demonstrates that the final result is better than that of hDQN algorithm.As shown in Fig.11,the sub-goals identification by HADS are superior to those automatically defined by hDQN.Therefore,on this basis,the effect of HADS is naturally better than that of hDQN.As can be seen from Table V,we show the average value,maximum,minimum and mean square deviation.Fig.11 is similar to Fig.7.

    Fig.10.Original generated sub-goal graph and sub-goal graph after removing unreachable sub-goal points.In (a),the blurred red dots are initial sub-goal points,with some of them being unattainable for the agent and some of them being not real sub-goal points,that is,being on the agent itself.In(b),the red rectangle denotes the actual sub-goal points.

    It can be seen that DQN and hDQN are relatively stable,but the rewards in each plot are relatively low.The rewards of hDQN and HADS are not stable.But after 1000 plots,the rewards of the two algorithms begin to grow and soon exceed the rewards of DQN,and the rewards of HADS are the highest.As can be seen from Fig.12,the performance of HADS is better than that of hDQN because of the sub-goals found.The expert player has obvious advantages over other algorithms,because the expert player has significant experience.After training,our algorithms also get a better training result.

    Fig.11.Kangaroo cumulative external rewards.DQN was the most stable.DuDQN did not achieve a better result.After 1300 plots,hDQN and HADS got better results than DQN.Finally the maximum reward of the HADS was better than the other three algorithms.

    Fig.12.Kangaroo reward for each plot.DQN and DuDQN have more stable trend.HADS and hDQN have better results.Before the 1200 plots,because Options have not been constructed,hDQN and HADS did not perform well.However,after successfully setting up Options,HADS and hDQN showed a powerful capacity and got much better results.

    TABLEIV AVERAGE REWARD FOR THE LAST 5 PLOTS

    C.t-Test

    In this section,we utilize the paired sample t-test to evaluate the difference between average results of the experiments.Thet-test includes the single population test and double population test,where the single populationt-test is used to test whether the difference between the average of a sample and that of a known population is significant.The double populationt-test can be divided into two situations:one is an independent samplet-test,which is used to test the difference of the data obtained by two groups of unrelated samples;the other is apaired samplet-test,which is used to test the difference between the data obtained by two groups of subjects or the data obtained by the same group of subjects under different conditions.The samples following two situations are the relevant samples.We used the method of the paired sample test,and the equations [39] for obtainingt-value is as follows:

    TABLEV RESULTS IN THE KANGAROO

    DQN and DuDQN suffer from sparse problem.As a result,by using DQN and DuDQN to solve the experiment tasks,they do not attain the reward until the termination,which means they are unable to receive an immediate reward during processing.It is meaningful to make statistical analysis.Therefore,in this paper,we use paired sample test oft-test to compare the difference between the average results of the hDQN and HADS.hDQN and HADS solve the aforementioned spare reward problem,where algorithms attain meaningful immediate reward during processing.We adopted experimental results of hDQN and HADS and the performedt-test.Thet-value andp-value are showed in Table VI.

    As pointed out by some work [40],ap-value of 0.5 denotes that there is no significant difference between these two samples that take part in comparison.Based on this,we can infer that hDQN and HADS have similar performance.Since hDQN is a state-of-the-art algorithm,HADS can also be regarded as a good approach.Compared with other results(Fig.7,Fig.8,Table I,Table II,Table III,Fig.11,Fig.12,Table IV,Table V,and Table VII) indicating HADS has a better performance than hDQN,and it can be concluded that HADS is a good algorithm.

    V.CONCLUSION

    Most standard hierarchical deep reinforcement learningapproaches have to choose sub-goals manually,showing their limitations.In this work,we put forward an algorithm of hierarchical reinforcement learning with automatic sub-goal identification via a computer vision approach.Based on current hierarchical deep reinforcement learning,sub-goals are selected automatically by using the way of computer vision before they are fed into deep layered reinforcement learning to achieve intelligent control.The experiment results demonstrate that our proposed approach has considerable improvement from others.

    TABLEVI T-TEST OF HDQN AND HADS

    TABLEVII RESULTS WITH ExPERT PLAYER FOR 30 PLOTS

    综合色丁香网| 国产不卡av网站在线观看| 亚洲精品国产av成人精品| 又粗又硬又长又爽又黄的视频| 久久久午夜欧美精品| 亚洲经典国产精华液单| 日日啪夜夜爽| 美女国产高潮福利片在线看| 亚洲国产精品一区三区| 日韩中字成人| 久久国产精品大桥未久av| 在线观看人妻少妇| 婷婷色综合www| 国国产精品蜜臀av免费| 国产深夜福利视频在线观看| 大香蕉久久成人网| 一个人免费看片子| 男人爽女人下面视频在线观看| 男女边吃奶边做爰视频| 美女主播在线视频| 久久久久久久久久久免费av| av专区在线播放| 国产69精品久久久久777片| 人妻人人澡人人爽人人| 18禁裸乳无遮挡动漫免费视频| 亚洲精品视频女| 黑丝袜美女国产一区| 中文字幕久久专区| 国产69精品久久久久777片| 热re99久久国产66热| 亚洲精品国产色婷婷电影| 日韩不卡一区二区三区视频在线| 你懂的网址亚洲精品在线观看| av在线播放精品| 国产色婷婷99| 欧美丝袜亚洲另类| 午夜91福利影院| 日韩欧美精品免费久久| 国产免费一区二区三区四区乱码| 18禁裸乳无遮挡动漫免费视频| 91午夜精品亚洲一区二区三区| 久久这里有精品视频免费| 午夜激情福利司机影院| 91久久精品国产一区二区成人| 亚洲国产欧美在线一区| 午夜av观看不卡| 人妻夜夜爽99麻豆av| 亚洲,欧美,日韩| 如何舔出高潮| 简卡轻食公司| 好男人视频免费观看在线| 激情五月婷婷亚洲| 一区在线观看完整版| 久久国产精品大桥未久av| 另类亚洲欧美激情| 亚洲人成网站在线播| 夜夜骑夜夜射夜夜干| 久久久久视频综合| 成人黄色视频免费在线看| 久久人妻熟女aⅴ| 国产探花极品一区二区| tube8黄色片| 久久久精品区二区三区| 我要看黄色一级片免费的| 夫妻午夜视频| 亚洲国产av影院在线观看| a级毛片在线看网站| 国产男女超爽视频在线观看| 少妇被粗大的猛进出69影院 | 美女脱内裤让男人舔精品视频| av专区在线播放| 久久精品国产a三级三级三级| 秋霞在线观看毛片| 精品久久久精品久久久| 色视频在线一区二区三区| av一本久久久久| 中国三级夫妇交换| 少妇被粗大猛烈的视频| freevideosex欧美| 欧美三级亚洲精品| 女性生殖器流出的白浆| 高清视频免费观看一区二区| 精品一品国产午夜福利视频| av在线老鸭窝| 国产深夜福利视频在线观看| 久久韩国三级中文字幕| 黑人高潮一二区| 3wmmmm亚洲av在线观看| 日韩电影二区| 国产乱来视频区| 天美传媒精品一区二区| 国产片内射在线| 国产男女内射视频| 18在线观看网站| 黑人巨大精品欧美一区二区蜜桃 | 国产国拍精品亚洲av在线观看| 18在线观看网站| 黄色配什么色好看| 极品少妇高潮喷水抽搐| av女优亚洲男人天堂| 久久99热这里只频精品6学生| 99久久中文字幕三级久久日本| 国产精品成人在线| 精品国产一区二区三区久久久樱花| 欧美变态另类bdsm刘玥| 国产精品一二三区在线看| 国产免费一区二区三区四区乱码| 黄片播放在线免费| 欧美精品国产亚洲| 99精国产麻豆久久婷婷| 美女中出高潮动态图| 免费黄网站久久成人精品| 精品久久久精品久久久| 国产综合精华液| 欧美精品国产亚洲| 亚洲国产av影院在线观看| 欧美亚洲 丝袜 人妻 在线| 人人妻人人澡人人爽人人夜夜| 人人澡人人妻人| 制服丝袜香蕉在线| 肉色欧美久久久久久久蜜桃| 亚洲高清免费不卡视频| 99久久综合免费| 久久久久久久久大av| 最近中文字幕高清免费大全6| 免费看光身美女| 九色亚洲精品在线播放| 成人国语在线视频| 免费看光身美女| 亚洲欧美一区二区三区国产| 少妇人妻 视频| 伦精品一区二区三区| 精品国产国语对白av| 2022亚洲国产成人精品| 精品亚洲成国产av| av一本久久久久| 久久久久久久久久久免费av| 一本一本综合久久| 一级毛片电影观看| 日韩大片免费观看网站| 欧美bdsm另类| 男人操女人黄网站| 国产精品国产三级专区第一集| 日本免费在线观看一区| 亚洲在久久综合| 国产成人午夜福利电影在线观看| 在线观看一区二区三区激情| 18禁动态无遮挡网站| 久久精品国产a三级三级三级| 国产一级毛片在线| 在线免费观看不下载黄p国产| 亚洲国产色片| av电影中文网址| 欧美日韩国产mv在线观看视频| 18禁在线无遮挡免费观看视频| 在线观看免费高清a一片| 美女福利国产在线| 激情五月婷婷亚洲| 男女边吃奶边做爰视频| 国产片内射在线| a级毛片免费高清观看在线播放| 日韩强制内射视频| 久久影院123| 亚洲国产精品一区三区| 国产综合精华液| 黄片播放在线免费| 人人澡人人妻人| 日韩中字成人| 久久精品久久久久久噜噜老黄| 国产精品一国产av| 最新的欧美精品一区二区| 亚洲一区二区三区欧美精品| 久久久久久久久久久久大奶| 熟女电影av网| 免费观看av网站的网址| 日本黄大片高清| 久久99一区二区三区| 国产高清不卡午夜福利| 免费大片黄手机在线观看| 日本黄色日本黄色录像| 欧美激情极品国产一区二区三区 | 99久久人妻综合| 精品亚洲乱码少妇综合久久| 精品亚洲成a人片在线观看| 亚洲av在线观看美女高潮| videos熟女内射| 中文字幕久久专区| 精品人妻在线不人妻| 欧美变态另类bdsm刘玥| 在线免费观看不下载黄p国产| 中文字幕最新亚洲高清| 男女边摸边吃奶| 色5月婷婷丁香| 亚洲精品乱码久久久v下载方式| 一区二区三区免费毛片| 色吧在线观看| 天天躁夜夜躁狠狠久久av| h视频一区二区三区| 欧美xxxx性猛交bbbb| 久久久久人妻精品一区果冻| 69精品国产乱码久久久| 少妇猛男粗大的猛烈进出视频| 久久ye,这里只有精品| 国产精品一区二区在线不卡| xxxhd国产人妻xxx| 人妻少妇偷人精品九色| 高清在线视频一区二区三区| 国产日韩欧美在线精品| 少妇被粗大的猛进出69影院 | 中文字幕av电影在线播放| 啦啦啦中文免费视频观看日本| av国产久精品久网站免费入址| 久久国内精品自在自线图片| 欧美日韩av久久| 亚洲av综合色区一区| 国产深夜福利视频在线观看| 高清在线视频一区二区三区| 99久久人妻综合| av在线老鸭窝| 在线观看一区二区三区激情| 亚洲欧美成人精品一区二区| 国产探花极品一区二区| 久久久久久久久久人人人人人人| 国产免费一级a男人的天堂| 欧美日韩成人在线一区二区| 欧美日韩视频精品一区| av免费在线看不卡| 桃花免费在线播放| 国产国拍精品亚洲av在线观看| 少妇的逼好多水| 观看美女的网站| 王馨瑶露胸无遮挡在线观看| 日本欧美国产在线视频| 麻豆乱淫一区二区| 婷婷成人精品国产| 久久久久久人妻| 在线观看免费视频网站a站| 国产午夜精品一二区理论片| 女性生殖器流出的白浆| 波野结衣二区三区在线| 波野结衣二区三区在线| 午夜福利网站1000一区二区三区| 日韩av不卡免费在线播放| 精品国产乱码久久久久久小说| 一边亲一边摸免费视频| 国产精品免费大片| 午夜免费鲁丝| 在线观看免费高清a一片| 高清欧美精品videossex| 97精品久久久久久久久久精品| 中文字幕免费在线视频6| 一区二区日韩欧美中文字幕 | 美女内射精品一级片tv| 免费日韩欧美在线观看| 国产精品一二三区在线看| 久久久a久久爽久久v久久| h视频一区二区三区| 国产精品.久久久| 寂寞人妻少妇视频99o| 国产午夜精品久久久久久一区二区三区| 人人妻人人澡人人看| 亚洲av.av天堂| 人人澡人人妻人| 久久久久网色| 国产一级毛片在线| 日韩电影二区| 日本黄色日本黄色录像| 国产免费一级a男人的天堂| 精品一区二区免费观看| 又粗又硬又长又爽又黄的视频| 久久99热这里只频精品6学生| 老司机影院成人| 日本91视频免费播放| 午夜视频国产福利| 国产男女超爽视频在线观看| 一区二区三区精品91| 各种免费的搞黄视频| av国产精品久久久久影院| 精品一区二区三区视频在线| 国产精品久久久久久精品古装| 国产免费又黄又爽又色| 女人久久www免费人成看片| 亚洲精品日韩在线中文字幕| 亚州av有码| av在线老鸭窝| 一级毛片aaaaaa免费看小| 女性生殖器流出的白浆| 欧美亚洲 丝袜 人妻 在线| 亚洲av福利一区| 国产免费视频播放在线视频| 青春草国产在线视频| 欧美精品高潮呻吟av久久| 黄色配什么色好看| 男女边摸边吃奶| 亚州av有码| 久久国产亚洲av麻豆专区| 黄色视频在线播放观看不卡| 亚洲国产精品国产精品| 精品少妇黑人巨大在线播放| 日韩人妻高清精品专区| 国产极品粉嫩免费观看在线 | 美女内射精品一级片tv| av卡一久久| 久久久久精品性色| 亚洲不卡免费看| 人人妻人人爽人人添夜夜欢视频| 男女高潮啪啪啪动态图| 久久午夜综合久久蜜桃| 久久国内精品自在自线图片| 99热这里只有精品一区| 麻豆成人av视频| 最黄视频免费看| 黑人欧美特级aaaaaa片| 亚洲欧美中文字幕日韩二区| 在线观看人妻少妇| 欧美最新免费一区二区三区| 18在线观看网站| 男女国产视频网站| 国产精品久久久久久精品电影小说| 91成人精品电影| 九草在线视频观看| 欧美日韩视频精品一区| 久久午夜综合久久蜜桃| 免费黄色在线免费观看| 久久国产精品男人的天堂亚洲 | 亚洲国产日韩一区二区| 久久99蜜桃精品久久| 国产精品欧美亚洲77777| av女优亚洲男人天堂| 又黄又爽又刺激的免费视频.| 久久精品人人爽人人爽视色| 国产熟女午夜一区二区三区 | 国产午夜精品一二区理论片| 纵有疾风起免费观看全集完整版| 99久久中文字幕三级久久日本| 国产精品久久久久久av不卡| 久热久热在线精品观看| 蜜桃在线观看..| 18在线观看网站| 久久ye,这里只有精品| 水蜜桃什么品种好| 亚洲三级黄色毛片| 纵有疾风起免费观看全集完整版| 777米奇影视久久| 2021少妇久久久久久久久久久| 一个人免费看片子| 黑丝袜美女国产一区| 日韩三级伦理在线观看| 91久久精品国产一区二区成人| 欧美亚洲日本最大视频资源| 亚洲久久久国产精品| 蜜臀久久99精品久久宅男| 免费av中文字幕在线| 2021少妇久久久久久久久久久| 国产亚洲av片在线观看秒播厂| 久久女婷五月综合色啪小说| 欧美bdsm另类| 国精品久久久久久国模美| 丰满饥渴人妻一区二区三| 成年女人在线观看亚洲视频| 永久网站在线| 又粗又硬又长又爽又黄的视频| 亚洲欧美清纯卡通| 婷婷成人精品国产| 亚洲在久久综合| 国产精品偷伦视频观看了| 国产亚洲午夜精品一区二区久久| 美女国产高潮福利片在线看| 高清在线视频一区二区三区| 国产成人精品在线电影| 亚洲色图 男人天堂 中文字幕 | 久久女婷五月综合色啪小说| 26uuu在线亚洲综合色| 国产av国产精品国产| 飞空精品影院首页| 日本黄色日本黄色录像| 午夜av观看不卡| 啦啦啦中文免费视频观看日本| 久热久热在线精品观看| 久久亚洲国产成人精品v| 免费高清在线观看视频在线观看| 久久午夜综合久久蜜桃| 欧美亚洲日本最大视频资源| 国产在线视频一区二区| 最黄视频免费看| 啦啦啦视频在线资源免费观看| 亚洲精品456在线播放app| 久久人人爽人人片av| 久久ye,这里只有精品| 精品酒店卫生间| 精品久久久久久久久av| 人人妻人人爽人人添夜夜欢视频| 人成视频在线观看免费观看| av又黄又爽大尺度在线免费看| 在线 av 中文字幕| av视频免费观看在线观看| 视频在线观看一区二区三区| 综合色丁香网| 特大巨黑吊av在线直播| 波野结衣二区三区在线| 亚洲一级一片aⅴ在线观看| 99久国产av精品国产电影| 在线观看免费视频网站a站| 久久久久久久久久久久大奶| 中文字幕最新亚洲高清| 亚洲成人一二三区av| 97超碰精品成人国产| 最新的欧美精品一区二区| 一级二级三级毛片免费看| 如何舔出高潮| 大陆偷拍与自拍| 国产国拍精品亚洲av在线观看| 在线免费观看不下载黄p国产| 久久久久久久久久成人| 国产熟女午夜一区二区三区 | 蜜桃在线观看..| 国产一区二区三区av在线| 一区二区三区四区激情视频| 极品少妇高潮喷水抽搐| 高清黄色对白视频在线免费看| 多毛熟女@视频| 99热这里只有是精品在线观看| 国产精品.久久久| 国产精品国产三级国产av玫瑰| 国内精品宾馆在线| 乱人伦中国视频| 麻豆成人av视频| 哪个播放器可以免费观看大片| 91在线精品国自产拍蜜月| 制服丝袜香蕉在线| 亚洲av.av天堂| 亚洲av免费高清在线观看| 麻豆成人av视频| 草草在线视频免费看| av不卡在线播放| 久久久久国产网址| 久久人人爽人人爽人人片va| 国产成人aa在线观看| 久久精品国产亚洲av涩爱| 亚洲av成人精品一二三区| 中文字幕精品免费在线观看视频 | 亚洲国产精品国产精品| 欧美日韩一区二区视频在线观看视频在线| 2018国产大陆天天弄谢| 国产视频内射| 成人免费观看视频高清| 日韩三级伦理在线观看| 国产乱来视频区| 性高湖久久久久久久久免费观看| 国产精品人妻久久久影院| 天堂俺去俺来也www色官网| av视频免费观看在线观看| 亚洲无线观看免费| 亚洲伊人久久精品综合| 婷婷成人精品国产| 狠狠婷婷综合久久久久久88av| 搡老乐熟女国产| av在线老鸭窝| 中文欧美无线码| 午夜福利影视在线免费观看| 亚洲国产精品999| 十分钟在线观看高清视频www| 韩国av在线不卡| 日本免费在线观看一区| 亚洲精品久久午夜乱码| 中文字幕久久专区| 国产欧美日韩综合在线一区二区| 中文乱码字字幕精品一区二区三区| 国产欧美亚洲国产| 免费av中文字幕在线| 国产在线一区二区三区精| 视频在线观看一区二区三区| 王馨瑶露胸无遮挡在线观看| 国产淫语在线视频| 搡女人真爽免费视频火全软件| 亚洲欧美日韩另类电影网站| 免费高清在线观看日韩| 青春草亚洲视频在线观看| 少妇精品久久久久久久| 69精品国产乱码久久久| 精品人妻熟女av久视频| 国产av精品麻豆| 亚洲精品视频女| 美女国产视频在线观看| 国产精品久久久久久精品古装| 日本免费在线观看一区| 久久97久久精品| 久热这里只有精品99| 色婷婷av一区二区三区视频| 国产精品一二三区在线看| av.在线天堂| 黄片播放在线免费| 天天影视国产精品| 看非洲黑人一级黄片| 最近的中文字幕免费完整| videosex国产| 9色porny在线观看| 国产免费又黄又爽又色| 人人妻人人澡人人看| 午夜老司机福利剧场| 久久这里有精品视频免费| 国产精品秋霞免费鲁丝片| 99热全是精品| 狠狠婷婷综合久久久久久88av| 亚洲国产精品一区三区| 亚洲精品一区蜜桃| 最新的欧美精品一区二区| 亚洲欧美清纯卡通| 久久鲁丝午夜福利片| 女人精品久久久久毛片| av女优亚洲男人天堂| 亚洲,欧美,日韩| 精品久久久久久电影网| 最近的中文字幕免费完整| 亚洲成色77777| 女人久久www免费人成看片| 久久99精品国语久久久| 9色porny在线观看| 国产日韩欧美视频二区| 人体艺术视频欧美日本| 精品国产一区二区久久| 国内精品宾馆在线| 精品国产国语对白av| 日韩在线高清观看一区二区三区| 最近最新中文字幕免费大全7| 五月天丁香电影| 欧美激情 高清一区二区三区| 男女边摸边吃奶| 亚洲高清免费不卡视频| 久久久a久久爽久久v久久| 中文欧美无线码| 18禁在线播放成人免费| 狠狠精品人妻久久久久久综合| 亚洲五月色婷婷综合| 免费黄频网站在线观看国产| 亚洲国产精品一区二区三区在线| 国产高清有码在线观看视频| 国产成人精品一,二区| 3wmmmm亚洲av在线观看| 色视频在线一区二区三区| 97超碰精品成人国产| 国产精品 国内视频| 久久久久久久精品精品| 日韩一区二区视频免费看| 久久精品国产自在天天线| 国产视频内射| 免费大片黄手机在线观看| 夫妻午夜视频| 99热这里只有精品一区| 黄片无遮挡物在线观看| 国产精品女同一区二区软件| a级毛色黄片| 欧美最新免费一区二区三区| 国产黄片视频在线免费观看| 国产爽快片一区二区三区| 男女免费视频国产| 午夜久久久在线观看| 国产亚洲一区二区精品| 伊人亚洲综合成人网| 丝袜脚勾引网站| 国产高清不卡午夜福利| 国产亚洲午夜精品一区二区久久| 成年女人在线观看亚洲视频| 一级毛片 在线播放| 韩国av在线不卡| 亚洲精品自拍成人| 亚洲四区av| 成人亚洲欧美一区二区av| 亚洲,一卡二卡三卡| 大又大粗又爽又黄少妇毛片口| 久久久久视频综合| 国产日韩欧美视频二区| 美女大奶头黄色视频| 国产日韩一区二区三区精品不卡 | 99久久人妻综合| 69精品国产乱码久久久| 国产精品久久久久成人av| 你懂的网址亚洲精品在线观看| 精品少妇内射三级| 久久亚洲国产成人精品v| 91久久精品国产一区二区三区| 亚洲图色成人| 九九久久精品国产亚洲av麻豆| 一本大道久久a久久精品| av.在线天堂| av不卡在线播放| 亚洲国产日韩一区二区| 久久97久久精品| 少妇被粗大的猛进出69影院 | 天堂俺去俺来也www色官网| 亚洲精华国产精华液的使用体验| 国产高清国产精品国产三级| 亚洲人成网站在线观看播放| 久久精品国产自在天天线| 亚洲精品456在线播放app| 久久99一区二区三区| 五月天丁香电影| 亚洲av日韩在线播放| 午夜福利视频精品| 成人免费观看视频高清| 精品一区在线观看国产| 91精品一卡2卡3卡4卡| 成人影院久久| 老女人水多毛片| 国产精品国产三级国产av玫瑰| 99热这里只有是精品在线观看| 日本与韩国留学比较| a级毛色黄片| 三级国产精品欧美在线观看| 国内精品宾馆在线| 在线精品无人区一区二区三| 最近中文字幕高清免费大全6| 国产探花极品一区二区| 精品久久蜜臀av无| 在线观看一区二区三区激情|