【作者】 金星；

【导师】 李明楚；

【作者基本信息】 大连理工大学 ， 软件工程， 2019， 博士

【摘要】 近年来,得益于信息化网络的高速发展,网络协作系统已经得到了广泛的普及,并为人们生活带来了极大的便利。个体用户是组成网络协作系统的基本单元,他们的决策行为会直接影响网络协作系统的性能。然而,与其他用户进行协作交互往往需要付出时间与精力上的花费,因此为了获取最大的收益,理性用户往往会选择:(1)拒绝提供服务,这就引起了合作困境;(2)提供低质量的服务,这就引起了信任困境。为了保证网络协作系统的正常运行,我们需要设计有效的激励机制来促进理性用户的合作行为和信任行为。在已有研究工作基础上,本文使用博弈论、演化博弈论、相似度理论、图推荐、多目标优化、强化学习等方法对合作激励和信任激励机制进行了相关研究。主要研究成果如下:(1)推荐激励机制对个体合作行为影响的定性研究。首先,本文以经典的图推荐算法EigenTrust为基础构建了一个基于用户合作属性的推荐激励机制,采用该推荐激励机制的请求者能够以一个更高的概率获得服务。接下来,本文使用演化博弈论定性分析了推荐激励机制对用户合作行为的影响。为了研究推荐激励机制的鲁棒性,本文还考虑了:推荐机制失误率,单请求者多提供者交互模型,推荐背叛策略和服务数量敏感场景下推荐激励机制的性能表现。最后,大量的仿真实验与对比实验验证了本文推荐激励机制在促进用户合作方面的有效性能。(2)理性背叛对个体合作行为影响的定性研究。在考虑互惠花费这一实际场景下,无条件合作策略会抑制合作的涌现,为了促进用户的合作行为,本文决定在推荐激励机制基础上进一步引入理性背叛机制。首先,本文继续使用演化博弈论对理性背叛机制进行了建模,并且设置推荐机制使用者会以概率c与无条件合作者进行合作,然后,本文定性研究了概率c与合作涌现程度之间的关系,特别地,当c=0时,推荐机制使用者被证明可以占优整个系统,从而实现了社会收益的帕累托最优。其次,为了解决实际应用中会遇到的策略识别问题,本文设计了一种基于余弦相似度的策略识别方法。最后,大量的仿真实验与对比实验验证了本文理性背叛机制在促进用户合作方面的有效性能。(3)合作与信任感知的激励机制研究。首先,基于系统监控到的用户合作行为和请求者反馈的信任评价信息,本文采用EigenTrust算法分别计算用户的全局合作属性和全局信任属性。接着,本文将用户推荐建模成一个多目标优化问题,通过计算帕累托前沿(Pareto front)里的有效解,进而可以得到一系列非支配的用户集合,该集合里的元素就是既慷慨又可信的工人。然后,为了从理论上分析合作与信任感知激励机制的性能,本文使用演化博弈论对用户间的交互行为进行了建模。通过使用李雅普诺夫稳定性理论,本文机制的有效性和稳定性得到了定性证明。最后,大量的仿真实验和对比实验验证了本文合作与信任感知激励机制在促进用户合作行为和信任行为上的有效性能。(4)结合多轮审计与真相发现的质量保证研究。首先,为了提高多轮审计机制在质量监控上的性能,本文引入了信誉系统来描述工人提供服务的可靠性。针对一个特定的众包任务,当请求者不进行审计时,本文提出了一种无监督算法RPM(Reputation-based Participant-Mine voting)来帮助请求者进行真相答案的推断,该算法是对传统PM(Participant-Mine voting)算法在考虑工人信誉场景下的一种扩展。接下来,考虑到理性请求者与工人都会根据环境动态地改变自己的策略以获得更好的收益,本文使用强化学习和(1-∈)期望概率学习来建模理性请求者的策略更新过程,使用多臂老虎机学习和邻居学习来建模理性工人的策略更新过程。为了从理论上分析机制的有效性和稳定性,本文使用了李雅普诺夫稳定性理论,并且证明:诚实地提供有效服务是理性工人唯一的演化稳定策略。最后,大量的实际数据集实验和人工数据集实验验证了本文机制在质量保证方面的有效性能。更多还原

【Abstract】 With the rapid development of the Internet,network-based collaborative systems have been widely used in our lives.Individual users are the basic unit of network-based collaborative systems,as a result,the performance of network-based collaborative systems will be directly affected by the behaviors of individual users.However,due to the fact that it is costly to collaborate with others in terms of time and energy,therefore,rational users who aim to maximize their benefits will:(1)refuse to provide services,which leads to a cooperation dilemma;(2)provide services with low quality,which leads to a trust dilemma.In order to ensure the performance of network-based collaborative systems,efficient incentive mechanisms which can promote users to be cooperative and trustworthy should be introduced In.Based on the related work,we use game theory,evolutionary game theory,similarity theory,graph-based recommendation system,multi-objective optimization,reinforcement learning to further study the methods on solving the cooperation dilemma and trust dilemma.The main contributions of this dissertation are listed as follows:(1)The qualitative study of recommendation incentive mechanism.Firstly,we use the typical graph-based recommendation algorithm EigenTrust to build a recommendation incentive mechanism,as a result,users who adopt the proposed recommendation incentive mechanism can have higher probabilities of been served.Secondly,we use evolutionary game theory to qualitatively study the performance of the proposed recommendation incentive mechanism on promoting cooperation.To further study the robustness of the proposed mechanism,IRIM(Imperfect Recommendation Incentive Mechanism),ORMP(One Requester Multi-service Providers).FS(Four-Strategies)and QS(Quantity Sensitivity)scenarios have been researched.Finally,extensive numerical and simulation experiments are conducted to highlight the performance and validate the theoretical properties of our mechanism.(2)The qualitative study of rational defection mechanism.Under a practical scenario which considers the reciprocity cost,unconditional cooperation strategy would inhibit the emergence of cooperation.To solve this issue.we propose a rational defection mechanism.Firstly,we still use evolutionary game theory to model the interaction between users,and R strategy users who use the rational defection mechanism are assumed to cooperate with unconditional cooperation strategy with the c(c ∈[0,1))probability.Particularly,if c=0,it is proved that R strategies users can dominate the system,as a result,the system can achieve the Pareto optimality.Secondly,to distinguish the R strategy users and unconditional cooperation strategy users,we propose a strategy recognition method based on the cosine similarity.Fmally,extensive numerical and simulation experiments are conducted to highlight the performance of our mechanism on promoting cooperation.(3)The study of cooperation and trust-aware incentive mechanism.Firstly,we propose a modified EigenTrust algorithm to simultaneously calculate the global cooperation value and global trust value of each user based on the past behaviors.Secondly,we propose a cooperation and trust-aware worker recommendation mechanism by determining a Pareto front from all users.Thirdly,we use evolutionary game theory to study the acceptance and effectiveness of the proposed incentive mechanism by strategically modeling the interaction between users.And then,the Lyapunov stability theory is employed to mathematically investigate the stability of the cooperation and trust-aware incentive mechanism.Finally,both numerical simulations and simulator-driving experiments illustrate that our mechanism has an outstanding performance in promoting cooperation and inhibiting malicious activity,(4)The study of quality assurance by incorporating multi-auditing and truth inference mechanisms.Firstly,in order to improve the performance of multi-auditing mechanisms on executing crowd tasks,we introduce a task-based temporal reputation mechanism to measure the reliability of each crowd worker.Secondly,we propose an unsupervised truth inference algorithm RPM(Reputation-based Participant-Mine voting)which can be used to discover the truth when the requester does not audit the feedbacks from one certain task.Thirdly,rational requesters and rational workers who will constantly change their strategies to obtain more benefits are considered in this dissertation.Specially,we use reinforcement learning and 1-∈accuracy learning to model the updating policy of a requester’s strategy and use K-armed bandits learning and neighborhood learning to model the updating policy of rational workers.By using Lyapunov stability theory,it is qualitatively proved that the trustful provision of sensed data provides a unique evolutionary stable equilibrium for each rational worker in our proposed model.Finally,extensive simulations and real data experiments illustrate that the proposed mechanism has an outstanding performance on quality assurance.更多还原