【作者】 文敦伟；

【导师】 蔡自兴；

【作者基本信息】 中南大学 ， 模式识别与智能系统， 2001， 博士

【摘要】 在智能控制领域，智能信息集成技术和智能控制算法是构成智能控制系统的两大核心技术。本论文正是从面向多智能体的信息集成和面向神经网络的控制算法两个不同的角度，展开对智能控制和智能系统的研究。通过面向多智能体的研究，提供了一种理论方法与体系结构，用于分布式智能控制系统的信息集成。而通过对神经网络的结构和学习算法的研究，结合模糊逻辑、进化计算和随机优化等方法，探讨了智能控制中的重要的分支——神经网络控制的有关理论与技术。 本论文的第一章为绪论，对智能控制、神经网络、分布式人工智能和多智能体的发展和研究概况进行了回顾和讨论，提出了本文的研究路线。 第二章的研究对象为多层前向神经网络。主要研究前向神经网络的权值学习和结构学习方法。介绍了神经网络原理和遗传BP算法。提出了样本划分启发式遗传BP算法和复合结构学习方法(CSL)，并进行了仿真实验。 第三章研究对象为递归神经网络。主要研究递归神经网络原理、结构及其学习算法。探讨了几种全局递归和局部递归神经网络的结构原理和学习算法。在模拟迟火及Alopex算法基础上，提出了用于一般递归神经网络的基于模糊逻辑的Alopex学习算法(FLA)，给出了仿真实验。在局部递归神经网络结构方面，提出了一种递归神经网络结构——自环对角递归神经网络结构(SDRNN)，给出了相应的学习算法，证明了算法的收敛性，并进行了仿真实验。 第四章研究基于神经网络的辨识和控制。在分析了神经网络控制的一些关键技术基础上，提出了预去模糊FAM原理和FAM神经控制器，给出了预去模糊FAM和一般FAM的等价性的构造性证明。提出了基于第三章提出的SDRNN的递归神经网络控制结构和实现方法，证明了算法的收敛性。分别以小车倒立摆系统和non-BIBO动态对象为控制对象对两种方法进行了仿真实验。 第五章研究分布式人工智能、智能体和多智能体的理论基础。侧重研究智能体的精神状态及其形式表示方法。首先分析探讨了智能体的基本概念、基本观点、结构分类、研究内容和应用领域。研究了智能体的精神(心智)状态的几种形式化表示。在已有的形式模型基础上，提出了多智能体系统的形式模型MASCL，该模型是一种多类的一阶分支时序模态逻辑BDI理论，综合了单智能体和多智能体的精神状态和动作规划表示，并体现了社会规范的约束作用。以MASCL为基础，定义了多智能体的协作过程的协作承诺等相关概念，提出了协作承诺的形式化模型。 第六章研究智能体和多智能体的体系结构、求解过程、实现技术及其在工业过程控制中的应用。首先探讨了基于多智能体的分布式智能控制在工业控制应用中的意义和特点。以第五章提出的MASCL模型为基础，提出了一种智能体的概念模型，设计了相应的智能体体系结构和智能体的求解过程，并 文敦伟：面向多宫能体和神经冈络的了能控饲研究以面向对象的代码形式给出了智能体信息与行为的描述。以烧结过程为对象，提出了多智能体建模与控制的框架。开发了一套用于智能控制和多智能体研究测试的实验系统，进行了实验演示。 第七章对全文进行了总结，指出了今后进一步研究的方向。更多还原

【Abstract】 In the field of intelligent control, there are two kinds of core technology: intelligent information integration and intelligent control algorithms, which construct intelligent control systems. Multi-agent oriented intelligent information integration and neural network oriented control algorithms are studied in this paper. On the one hand, theory methods and architecture are proposed for integration of information in distributed intelligent control systems. On the other hand, by combining fuzzy logic, evolutionary computation and stochastic optimal methods, structure and learning algorithms of neural networks are studied for neural control, which is one of the important branches of intelligent control.Chapter 1, based on a survey on intelligent control, neural networks, distributed artificial intelligence and multi-agent systems, presents the study tasks of this paper.Chapter 2 focuses on weight learning and structure learning of multi-layer feed-forward neural networks. The genetic BP algorithm are discussed. Then two novel algorithms, samples division based heuristic genetic BP and composition structure learning, are proposed. Simulations are performed respectively.Chapter 3 studies structure and learning algorithms of recurrent neural networks. Based on simulated annealing and alopex learning, a novel leaning algorithm, fuzzy logic alopex learning (FLA), is proposed and simulated.A new recurrent neural network structure, self-feedback diagonal recurrent neural networks (SDRNN), is also designed in this chapter. The learning algorithm of SDRNN is given and the convergence of this algorithm is proved. The simulation results show the validation of the structure and the learning algorithm.Chapter 4 deals with neural network based identification and control. After the analysis of several important aspects of neural control, two novel neural controllers are then proposed. The first controller FAM neural controller (FAMNC), based on Pre-defuzzifing FAM, is presented for bridging the gap between FAM and NN. The equivalence of pre-defuzzifing and general FAM is proved constructively. The simulation research on inverted pendulum control is performed. The simulation results show that the FAMNC is sound. The second controller is SDRNN based neural controller. The control architecture is given and the convergence of the control algorithm is proven. The simulation research on a non-BIBO dynamic plant control is performed. The simulation results show the availability of this controller.Chapter 5 turns to distributed artificial intelligence, intelligent agent and multi-agent. First, the concepts, structure categories, research content and application of intelligent agent are introduced. Several formulation of mental states of intelligent agent are discussed. Second, based on former logic theories, a improved formal model, MASCL, is proposed in this chapter. MASCL is a many-sorted first-order branching-time BDI logic, which can capture the requirements for representation mental states, acts, plans and social laws of multi-agent systems. With the use of MASCL, the concept of cooperation commitment of cooperation process in multi-agent systems is defined, and the formal model of cooperation commitment is proposed.Chapter 6 researches into the architecture, solving process and implementation of intelligent agent andmulti-agent systems, and the application to industry process control. First, the significance and characteristics of the application of multi-agent based distributed intelligent control to industry control are discussed. Second, a concept model of intelligent agent is proposed, the architecture and solving process are designed and the information and acts of intelligent agent are described by object-oriented code. Third, a multi-agent modeling and controlling framework are designed for agglomeration process. An experiment system is developed for intelligent control and multi-agent systems studies, and tests are performed on this system.The conclusions of this paper are drawn and the futur更多还原