【作者】 郭航；

【导师】 高济；

【作者基本信息】 浙江大学 ， 计算机应用技术， 2010， 博士

【摘要】 随着计算机网络和信息技术的快速发展和普及,分布式计算领域的研究和应用得到了长足的发展,分布式协同工作模式在计算机应用中占有越来越重要的地位。由Agent管理和控制资源,并组成多Agent系统(MAS)来主导分布式协同工作,被认为在未来的分布式协同工作中占有不可或缺的地位。在MAS领域的深入研究过程当中,研究者们逐渐认识到松散型协同工作模式是未来组织和个人迫切需要的协作组织方式。因此对由异构和自私性Agent组成的多Agent系统研究成为MAS领域的主要议题。但是由于这种环境下的Agent行为以各自利益最大化为主要原则,同时由于协同参与程度的差异,很容易导致多个Agent在一起的组合不但没有形成合力,反而对团体产生负效用,也即“负叠加效用”问题。另外由于Agent的黑箱和自治性,Agent行为对系统和其它Agent是不可测和不可控的,造成Agent之间互不信任,从而导致多Agent系统的“可信危机”问题。规范调控的多Agent协同系统(NMAS)被视为解决上述问题的一种有效手段。NMAS系统利用规范约束成员Agent的行为,从而实现多主体相互协同,是介于集中控制和完全P2P结构之间的一种多主体协同机制,被视为是实现MAS系统可信协作的有效方法,然而,目前NMAS领域的研究工作主要集中在某些议题上的理论和抽象模型研究,尚处于初级阶段,在该领域的许多方面上都缺少可有效支持实用系统开发的可操作性架构和模型。首先,目前许多研究没有明确所面向的背景,研究多聚焦于规范系统的共性方面,而缺乏面向具体特性环境的规范体系研究。这造成在将规范体系引入某种类型的MAS时,现有的模型和机制会在某些方面与其失配。目前NMAS研究的另一大问题是缺乏可操作性,集中体现在以下几个方面:第一、缺乏支持规范系统建立和运作的体系结构和整套方法论。第二、对规范表示方面的研究,目前集中于对规范表示要素和表示逻辑本身属性,如完备性、一致性等方面的考虑,而缺乏对规范表达方式的可操作性,如易于Agent参照决策、易于违规和处罚判定方面因素的考虑。第三、对规范实施,包括违规和处罚的判定和实施,尚缺乏深入研究和有效模型机制的支持。第四、目前提出的规范推理模型做出的一些取消不确定性的前提假设在应用中难以达到,尚缺乏不确定性的规范推理与决策模型。针对上述问题,本文以自治服务协同的多Agent系统为背景,对NMAS的体系结构、规范表示、推理和实施等多方面进行了深入研究和探索。本文的研究工作力求形成一套从宏观的规范体系定义到微观的规范约束推理的规范系统方法论,从而系统化解决当前多Agent系统存在的负叠加效用和可信危机问题。本文的研究内容和主要创新点包括:(一)提出了面向自治服务协同MAS环境的“规范调控的自治服务协同体系结构”,即NASCA体系结构。该体系结构围绕着核心规范模型,提出三层框架体系来支持各个不同层面上对规范相关问题的建模,包括由定义组织框架的宏观层、描述Agent推理和行为决策的微观层以及通过规范约束VO连接两者的连接层组成的。该体系结构为NMAS系统建立和运行提供了完整的方法论体系。(二)通过结合道义逻辑算子和时态逻辑算子,在扩展传统规范表示逻辑的基础上,提出了可操作性规范逻辑(ONL);并以此为表示基础,提出了基于结构化、分层次E机构,及E机构之间继承关系的规范定义模型--组织规范模型(ONM)。该模型有助于规范语义的清晰表达,也便于在其之上构建实施和推理模型,能够弥补目前规范表示研究对Agent行为决策和规范实施支持上的不足。(三)建立规范实施的体系模型,提出由适用规范(E机构)的判定、守违规判定、处罚判定和处罚实施四个部分组成的规范实施过程和规范实施模型(NEM)。该模型以Agent的行为(消息)为证据,以ONM模型表示的规范定义体系为判决依据,通过给出上述四个紧密联系的实施子过程的解决机制和算法,能够加强规范对自治Agent协作行为的软约束效果,保证规范的有效性。(四)提出了不确定性规范推理和决策模型NRDM。该模型针对当前规范推理模型中对Agent守违规进行了确定性假设的不足,采用根据Agent历史守违规行为记录推断其守规可信度(概率)的方法,并根据当前VO的总体可信度环境对该值进行迭代修订,最后将可信度值作为几率结点引入贝叶斯决策网络模型中进行不确定性推理和决策。该模型弥补了规范推理研究中缺乏不确定性推理模型的不足。本文通过体系化的方式形成了一套支持NMAS系统建立和有效工作的体系结构,为规范系统中的各个层次上的建模提供有效支撑;进一步,通过对规范表示、规范实施和规范推理等具体方面工作的深入研究,又反过来论证了该体系结构的可操作性,最终为自治服务协同环境提供了对NMAS建模的一整套可操作的方法论模型。更多还原

【Abstract】 With the rapid development of network and information technology,distributed computing technology has got rapid growth and been wildly applied.Distributed cooperation becomes increasingly important in applications.It is considered a promising method that resources is managed and controlled by agents,and composed to Multi-Agent Systems(MAS) for cooperative work.In the study of the MAS area, researchers have come to realize that a loosely coupled cooperation mode is the future need for organizations and individuals.Therefore,research on MAS that is composed by heterogeneous and selfish agents becomes a main topic in the area.However, because agents try to maximize their own interests and differences in the level of cooperative participation,it easily leads to the "negative collective utility" problem.In addition,as the clandestine character and autonomy of Agent,agent behavior is unpredictable and uncontrollable.It causes mistrust between agents,resulting in the "crisis of credibility" problem in MAS.Normative Multi-Agent Systems(NMAS) is considered as a promising method to solve these problems.NMAS uses norms to restrict the behavior of members in order to achieve multi-agent coordination with each other.It is mechanism between centralized control and complete P2P mode.So it is considered as an effective method to achieve credible cooperation in MAS.However,current research is mainly concentrated on theory and abstract models,still being in its infancy and lack of practical operational framework and effective model for applications in many aspects of this field.First,there is no clear background for current models.They mostly focus on commonness of norms system,rather than norms system under specific environment.It results in the mismatch between certain aspects of applications and current models.Another big problem in NMAS research is the lack of Operability. First,it is lack of systems architecture and set of methodologies to support the establishment and operation of normative systems.Second,research on norm representation is currently focused on attributes of norms and logic itself,such as completeness,consistency and other considerations,while the lack of standardized expression of operability,such as easy for decision-making,and easy to determine violations.Third,it is lack of in-depth study and effective model to support normative enforcement,including determination and implementation of violations and penalties. Fourthly,it is hard to achieve the preconditions and the assumptions proposed by the current models of normative reasoning.And it is lack of probabilistic method of normative reasoning and decision-making. To address the above problems,this thesis takes in-depth study and exploration on many aspects of NMAS architecture,norms expression,reasoning and implementation, under the background of autonomic service cooperation.Research in this thesis is focused on improving the operability of normative system models,and tries to form a set of methodology from the macro level to the micro level to support system modeling. Thus it systematically solves the "negative collective utility" and "credibility of crisis" problems.This study points the content and main innovations include:(A) Oriented to autonomic service cooperation environment,proposing a Norm-governed Autonomic Service Cooperation Architecture,namely NASCA. Focusing on the key models,this architecture proposes a three level framework, composed of macro-level of organization framework definition,micro-level of agent reasoning and behavior description and connection level that connect the above two and form VOs.This thesis takes in-depth study of those models in each level.The architecture supports a more workable NMAS system to establish and operate.(B) By combining the deontic and temporal logic operator and extending to traditional logic,this thesis presents an Operational Norm Logic(ONL).Based on this logic,it presents structured,hierarchical norm definition model,namely Organizational Normative Model(ONM).This model is helpful to the clear representation of norm semantic,and it supports the enforcement and reasoning well.It is a supplement for current study,which is lack of support to normative reasoning and enforcement.(C) It is proposed a Normative Enforcement Model(NEM).This model is composed of four parts,including E-institution determination,violation determination, punishment determination and punishment enforcement.Using agent actions (messages) as evidences and ONM as determination foundation,it presents the solutions and algorithms of those four problems.It can enhance the effect of soft constrains of norms.(D) It is proposed a probabilistic reasoning model for normative reasoning and decision-making process,namely NRDM.It aims at the deterministic assumptions of current normative reasoning models,presents a method that computing the norm-compliance trustiness by agents’ history action records. Moreover,it revises the trustiness value iteratively according to the total environment trustiness measure.Finally,it brings the trustiness value as a probabilistic node into a Bayesian decision network for uncertain reasoning and decision-making.This thesis presents a set of models for supporting the establishment and effective work of NMAS systems,and modeling in each level.Furthermore,by deeply study in the details of norm representation,enforcement and reasoning,it shows the operability of the architecture.We can conclude that it is proposed a set of effective models for NMAS modeling in autonomic service cooperation environment.更多还原