【作者】 郭庆；

【导师】 陈纯；

【作者基本信息】 浙江大学 ， 计算机科学与技术， 2003， 博士

【摘要】 agent和多agent系统(Multi-agent Systems，简称MAS)系统正在崛起为人工智能研究实用化和分布计算环境下软件智能化的重要技术。系统提供具有社会和领域知识，能依据心理状态自主工作，并具有语义互操作和合作行为协调能力的软件实体，不仅为实施紧凑一致的协同工作提供有力的支持，也为建立面向分布计算的开放性、可重构和可伸缩的新型计算环境建立了基础。 MAS技术的研究主要集中在三个方面：agent行为理论、agent体系结构和多agent协作。其中，多agent协作主要是研究自动agent系统如何能够更有效的交互。近来协商作为agent之间达成对双方都有利的决定的机制，已经成为MAS系统协作研究中的一个重要的分支。许多学者提出了一些理论和方法，包括Rosenschein等人的基于对策论的理性协商理论，R.Smith等人提出的合同网理论，Sycara的以劳资谈判为背景的协商理论，Kraus的最佳平衡协商理论，Sandholm的多agent核结盟理论以及Jenning等人基于协商中思维状态变迁的联合承诺理论等等。这些理论从协商主体数目、交互方式、协商议题数目、时间限制、环境信知、效用评价机制、协作个体的社会性等多个方面对协商框架进行了研究。总体来看，协商领域的研究范畴主要包括协商过程中的涉及到的思维状态模型、协商通讯语言、协商交互过程、协商决策框架四个方面。本文则针对这些问题作了的深入的探讨，分析了相关研究工作的特点和不足，并给出了我们的研究成果。 首先，从思维状态模型的角度来看，在MAS系统中的协商是一个多agent之间形成联合承诺的过程，而联合承诺的形成、维持、消解除了受到agent个体自身和其他agent的思维状态的影响之外，还和agent个体自身能力有关。本文扩展了Jenning等人提出的BDI模型，提出一个基于能力算子的多agent意识态度模型，并阐述了能力因子在多agent联合承诺的产生过程中的重要作用。 多agent之间的多议题协商框架也是目前协商研究领域的热点。本文针对限时条件下的两方多议题协商提出了多回合协商框架。框架首先给出了包含协商双方关注的多个与协商命题相关的因子的多议题集合；其次，框架提供了一个agent双方根J白︸丫!:据自身和对手的各种信知，交替提议的推理机制，该推理机制综合考虑时间、资源、对方底价等多方面的因子，提高了协商交互中的效率;在协商决策机制方面，该框架采用比较多议题向量整合效用值之间曼哈顿距离的算法来进行效用评估，决策是否达成一致的协定。与其他的研究成果相比，该框架提出了一个多策略整合的交互推理机制，并且在协商效用评估机制中考虑了各议题之间的相关性，提出了一个多议题整合的效用评估机制。 为提高复杂环境下的协商中的协商效用，必须解决参与者的信知不完整性和不精确性问题。本文采用BP神经网络学习算法来更新agent关于环境和其他个体的信知，从而使agent在协商交互中能够对协商对手的推理策略、保留值等要素动态预测，从而能够更有效的进行交互推理。与相关的研究成果相比，本文采用的学习算法通过考虑到历史信知、对方推理策略等多方面的因子，更加适合复杂环境下的多议题协商交互。 限时条件下的多议题协商中，由于某一个议题没有达到平衡点使得整个协商陷入僵局是影响协商效用的一个重要因素。为避免协商僵局，本文给出了一个多议题保留值置换算法，在保证协商参与者整体协商效用的前提下动态放宽某个议题的保留值，促使协商双方避免协商僵局。在等效置换值向量的选择中采用了强化学习算法进行优化，提高了置换效率。 基于本文的研究成果，我们设计并实现了一个面向金融业务的交易处理平台一一SunBridge，并在国内银行业得到应用。该系统的实际应用效果表明了本文的研究能够有效协调系统资源，高效的完成各种交易处理，满足该系统对并发性和开放性的要求。 文章最后对本文研究工作作了简短的总结，并提出了本文研究存在的不足和对将来工作的展望。更多还原

【Abstract】 In recent years, agent and Multi-Agent System(MAS) is growing up to be oneof the most important techniques in the practical research on Artificial Intelligence and intelligent software in distributed computing environment. As a kind of software entity that has social knowledge and works automatically according to self-mental state, agents operate and collaborate mutually among the MAS as well, contributing to provide a substantial support for the open and flexible distributed computation and cooperation environment.In the literature of MAS there are three main categories, known as agent behavior theory, agent architecture and collaboration among agents that is focused on how agents can effectively interact, notably negotiation, as a useful mechanism to get a beneficial result for both sides in the course of collaboration, becoming more inviting to researchers. Many researchers presented corresponding models for multi-agent negotiation framework. Rosenschein introduced a negotiation model using the idea of Game Theory, Sycara constructed a model based on the labour and capital negotiation, Kraus created a common negotiation framework considering sociality and cooperation in negotiation, Sandholm gave a commitment based negotiation mechanism. Another model given by Jennings took the changing of the mental state and joint-commitment into account. These researches concentrated on different aspects of the performance of a negotiation model including amounts of negotiation participants, negotiation protocol, the numbers of the issues involved, commitment in negotiation, time limit in negotiation.In general, the exploration for negotiation involves the following aspects:agent mental state model,ACL(agent communication language), interactive process and decision-making framework. In this paper, we will first deeply discuss these problem, then analysis the advantage and disadvantage of some related research work, our research result are presented as well.Firstly, from the viewpoint of mental state, the negotiation among agent is a process to fulfill the joint commitment, while how to fulfill, maintain and destroy the joint commitment is influenced not only by the individual mental state, but by the individual capability of the agents. In this paper, we firstly propose a individual mental state model based capability factor, where the relationship of the individual capability, individual intention and individual commitment is extended, then apply the joint-capability to present a MAS mental model. We also describe the process of the establishment of joint commitment.We also improve the multi-issues negotiation framework among multi-agent system in our work. This paper gives a bi-lateral multi-issues negotiation model under the time-limited condition, consisting of a multi-issues set, an alternation proposal reasoning frame and a utility integration decision-making mechanism. The multi-issues set is a container of some issues that relates to the negotiated proposition. The reasoning frame works out the proposal to the opposing party in terms of the integration multi-tactic, and this integration tactic comprehensively considers time, resource and opponent-tactic. We give a method of calculating the Manhattan Distance between the utility of two multi-issues vectors with the integrated-utility evaluation. Compared with related research, this model proposes an interactive reasoning mechanism integrating multi-tactic, and considers the participant’s expectation for all the factors related to the negotiated proposition, which makes the negotiation more efficient and more practical.In the complex environment, the incompleteness and incorrectness of the negotiator’s belief is essential to be solved, thus to improve the efficiency of thenegotiation. Adopting above bilateral multi-issues negotiation model, we introduce a machine to forecast the result-set of the negotiation, which employs history result of the negotiation and the first proposal of both negotiators to determine the reasoning tact更多还原