【作者】 于江涛；

【导师】 钱积新；

【作者基本信息】 浙江大学 ， 控制科学与工程， 2003， 博士

【摘要】 关于Agent理论和多Agent系统的研究是近年来分布式人工智能领域的研究热点。论文从知识表示、模型建立、学习和协作等方面对Agent技术做了全面而深入的研究，在总结了前人研究成果的基础上做了有效的改进，提出了自己的创新点和应用成果。 本文的主要研究内容和创新包括： 1．在传统的理性Agent的BDI形式化逻辑模型中作者引入新的逻辑算子BEL、ASM、DES、GOAL和INT等，表达了信念、愿望和意图三者间的动态约束与相互激发关系，补充了正规模态逻辑的KD45公理，建立Agent从信念到动作选择的的意图模型，为研究Agent与环境交互的自主行为模式提供了理性化模型。 2．Agent的推理能力被认为是衡量Agent智能性重要的指标。针对现有符号逻辑描述方法难以保证知识表达的完整性，推理过程陷于复杂的逻辑演绎的问题，作者引入了模糊因果关系的网络模型，基于模糊认知图理论构造Agent推理模型，用简单的数值计算代替复杂符号系统的表示和演绎推理过程，实现了复杂环境下的Agent智能决策。 3．Agent的学习能力是体现智能性的基础。论文的研究在现有强化学习算法的基础上，采用模糊建模的方法对于Agent的内部模型和状态的表示方法进行改进，提出一种模糊强化学习算法，降低了Agent学习对于精确模型和知识的要求，提高了算法的实用性。 4．对于Agent在协作技术方面的研究，针对传统合同网模型资源消耗大，协商过程长的缺点，在原有的合同网中定义各网元之间的关系权值，提出一种关系型合同网模型。通过对系统内的Agent进行面向任务的预分类，大大节省了通讯时间和资源占用，提高了系统的整体性能。另外，关系权值可以随着环境的变化而动态调整，具有较大的灵活性。加介次驴浙江大学博士论文 亘卑国国典硬理单以国目口口.口典 5.在工程应用研究方面，本文研究了Agent在系统优化和交通调度领域 的应用技术并取得了一定的工程成果，同时对基于Ageni的决策支持 系统做了探讨。 论文分从不同角度和层次多Agent系统理论的关键技术做了全面的探讨，既继承了前人的研究成果，同时对基于Agent的思想和方法做了深入的发掘和创新，展望了Agent技术在人工智能领域的开拓性前景。更多还原

【Abstract】 In the past few years, most of the research of DAI (Distributed Artificial Intelligence) focused on Agent and Multi-agent System theory. In this dissertation, detailed study was made on Agent technology from knowledge denotation, modeling, Agent learning to multi-agent cooperation. Improvements were made on the basis of early research, as well as new opinions and applications were proposed.This dissertation is composed of follows:a) Introducing new logic operator into the traditional Agent BDI model, including BEL, ASM, DES, GOAL and INT, in order to describe the dynamic restrictions and interactive triggering relations between BELIEF, DESIRE and INTENTION of Agent. A new intentional model was built in complementation of the KD45 regular modal logic axiom, which is the base of Agent self-control interaction with the outer environment.b) Deducing is an important property of Agent intelligence. Symbol logic method is unable to guarantee the complement of knowledge description, which leads to complicated deducing process. We introduce Fuzzy Cognitive Map into Agent modeling and deducing, substitute symbolic description and inference with simple mathematical computing, achieving Agent intelligent decision-making in complex environment.c) Learning ability is the base of Agent self-determination behavior. Reinforcement Learning is an applicable machine learning method of Agent state and knowledge. We combined RL and fuzzy logic together to make improvement on Agent inner model and state denotation method. This Fuzzy Reinforcement Learning Method decreases the requirement of modeling precision and makes the algorithm more applicable.d) On the research of Agent cooperation, on the basis of Contract Net Model,we define relationship weights between the network nodes. By means of pre-classification of the agents in a system, agents negotiation and task distribution are handled in less time and resource consumption, the whole system performance are greatly improved.e) Application of Agent technology are also studied in many fields such as system optimization, transportation schedule and decision support system, ect.更多还原