【作者】 李勇平；

【导师】 李荣钧；

【作者基本信息】 华南理工大学 ， 管理科学与工程， 2009， 博士

【摘要】 电信业务的传统预测模型多为统计回归模型和时间序列模型。前者基于输入变量和输出变量之间的因果关系,要求变量满足某些特定的统计假设;后者基于时间序列的惯性推演,必须确知或假定序列的变化规律。由于实际情况很难满足上述条件,所以传统预测模型的误差偏大、使用效果不佳。近年来,以神经网络为代表的智能预测系统开始在电信业务的预测中得到应用,但是单一的智能预测技术都或多或少地存在着这样那样的缺陷与问题。为此,不同智能技术之间的相互促进与补充便成为一种自然的考虑和首要的选择。虽然智能技术具有某些共同的机制和原理,但不同的智能技术表现出不同的行为特征。神经网络是模仿人脑结构及功能的非线性信息处理系统,具有大规模的并行计算与分布式存储能力,且在处理信息的同时,通过对信息的有监督和无监督学习,实现对任意复杂函数的实值映射。但是,普通神经网络的BP学习算法受初始权值的影响较大,不仅收敛速度缓慢、且容易陷入局部极值,故在实际应用中受到诸多限制。而基于人工生命和演化计算理论的粒子群优化算法将生物的优胜劣汰过程类比为可行解优化的迭代过程,形成一种以“生成+检验”为特征的自适应人工智能技术。由于粒子群优化算法对于参数搜索空间没有苛刻的条件,故在许多工程优化的实际问题中得到了成功的应用。但迄今为止,智能优化和智能预测技术基本上停留在仿真模拟阶段,还缺乏能够完全阐明智能技术运算特性的理论基础。本研究旨在对BP神经网络和标准粒子群优化算法进行综合的理论及应用分析,试图结合混沌变异技术与小生境进化策略,改进粒子群算法的优化机制,使之具备学习适应与协调进化的双重智能,以达到提高算法搜索速度和精度的目的。然后,将改进的粒子群算法植入神经网络的拓扑结构,用以替换网络的BP学习算法,建立新的粒子群神经网络系统,并最终在电信业务样本的基础上,构建基于改进粒子群神经网络的电信业务预测模型。论文的研究内容主要包括:(1)电信业务的经营现状与发展趋势,影响电信业务的主要因素及预测要求,现行预测模型的性能及存在的主要问题;(2)粒子群算法的基本原理与优化机制,现行粒子群算法存在的问题及原因,改进粒子群搜索性能的理论基础与现实途径;(3)粒子群算法与遗传算法、混沌算法之间的差别与联系,混沌变异技术及混沌初始化程序和小生境进化策略对粒子群算法的作用机制与结合方式,改进粒子群算法的参数设计与计算程序;(4)标准测试函数的特性与选择,改进粒子群算法与标准粒子群算法的比较实验与结果分析;(5)粒子群算法与神经网络的结合原理与集成方式,结合网络的拓扑结构与学习算法,改进粒子群神经网络的模型设计与算法程序;(6)电信业务的预测指标及影响因素,样本数据的采集与统计分析,基于改进粒子群神经网络的电信业务预测模型,预测系统的结构参数与训练结果;(7)六种电信业务预测模型实验结果的比较分析与初步结论;(8)所有智能预测模型在MATLAB7.0平台基础上的设计与开发程序。研究的成果及创新性主要表现为:(1)将混沌优化技术和小生境进化策略融入粒子群算法结构,并通过适应度函数变换和惯性权重自适应调整,提出了一种具备学习适应与协调进化双重智能的改进粒子群优化算法,显著提高了算法的搜索速度和精度;(2)将改进的粒子群优化算法植入神经网络的拓扑结构,用以替换网络的BP学习算法,集成了新的粒子群神经网络系统,显著改进了系统的学习进化能力和预测效果;(3)在MATLAB7.0软件平台的基础上设计和开发了所有智能优化算法和智能预测模型的计算机应用程序,顺利完成了所有智能优化算法和智能预测模型的实现过程;(4)确定了电信业务预测的指标体系和影响因素,并基于中国电信和中国移动的样本资料,结合样本数据的统计分析,构建了基于改进粒子群神经网络的电信业务预测模型;证实了改进粒子群神经网络预测系统的显著成效。更多还原

【Abstract】 Telecommunications services over the traditional forecasting models for statistical regression model and time series model. The former based on the input variables and the causal relationship between output variables require variables to meet certain statistical assumptions; time series based on the inertia of the latter deduction, you must really know, or assume that the sequence variation. As the actual situation is difficult to meet the above conditions, so the traditional prediction model error is too large to use ineffective. In recent years, neural networks, represented by intelligent forecasting system began in the telecommunications business, has been applied to forecast, but a single intelligent predictive technologies are more or less the existence of such kinds of defects and problems. To this end, among the different intelligent technology to promote and complement has become a natural consideration and inevitable choice.Although the intelligent technology has some common mechanisms and principles, but different intelligent techniques show different behavioral characteristics. Neural network is to imitate human brain structure and function of non-linear information processing system, with large-scale parallel computing and distributed storage capacity, and in processing information at the same time, through information, supervised and non-supervised learning to achieve for any complex functions of real-valued mapping. Thus based on the theory of artificial life and evolutionary computation, particle swarm optimization process of biological survival of the fittest feasible solution for the optimization of analog iterative process, forming a kind of a "Build + test" is characterized by adaptive artificial intelligence techniques. As the particle swarm optimization algorithm for the parameter search space is not harsh conditions, so in many practical problems in engineering optimization has been applied successfully. But so far, intelligent optimization and forecasting techniques are basically remain in the simulation stage, but also a lack of intelligence technology to fully clarify the theoretical basis for computing features.This study was designed on the BP neural network and the standard particle swarm optimization theory and application of comprehensive analysis, try to combine niche technology with chaotic mutation evolutionary strategy to improve particle swarm optimization mechanism, so as to learn to adapt to and coordination with the evolution of dual intelligent search algorithm in order to achieve increased speed and accuracy purposes. Then, the improved particle swarm algorithm embedded neural network topology, to replace the network BP learning algorithm to create a new particle swarm neural network system and, ultimately, the sample in the telecommunications business, based on the building of improved particle swarm neural telecommunications network prediction model.The main contents of this study include:1) Telecommunications business operation status and development trend of the main factors affecting the telecommunications business and forecast demand, the current forecast model performance and major problems.2) The basic principles of particle swarm optimization and optimization of the mechanism, the current problems in particle swarm optimization algorithms. The theoretical basis and practical way to improve the search performance of PSO.3) The difference and relation between Particle swarm optimization,genetic algorithm and chaos algorithm, Chaotic mutation techniques and chaotic initialization and niche evolution strategy on the role of particle swarm optimization mechanism and binding mode, improved particle swarm algorithm design and calculation of parameters procedures;4) The features and options of Standard test functions, The comparative experiment and the results analyzing of the improved standard particle swarm algorithm and particle swarm optimization ;5) The principle of Particle Swarm Optimization and Neural Networks combination and the mode of their integrated approach, combining the network topology and learning algorithm to improve the neural network model of particle swarm algorithm for the design and procedures;6) Predictor of telecommunication services and influencing factors, the sample data collection and statistical analysis, neural network based on improved particle swarm telecom business forecast models to predict the structure of the system parameters and training results;7) Six kinds of telecommunications forecasting model comparative analysis of experimental results and preliminary conclusions;8) All intelligent prediction model based on MATLAB7.0 platform, the design and development process.The research achievement and the innovative main performance are:1) Using Chaotic optimization technology and niche evolution strategy into the particle swarm algorithm structure, and through the fitness function transform and adaptive inertia weight adjustment, proposed a study to adapt to and coordination with the dual evolution of intelligence, improved particle swarm optimization algorithm, significantly improve the speed and accuracy of search algorithms 2) An improved particle swarm optimization algorithm embedded neural network topology, to replace the network BP learning algorithm, integrating a new particle swarm neural network system, significantly improved the system’s learning ability and prediction of evolutionary effects;3) Design and development of all the intelligent optimization algorithms and intelligent predictive model of computer applications base on the MATLAB7.0 software platform , the successful completion of all the intelligent optimization algorithms and intelligent predictive model realization process;4) Identified the telecommunications business forecast indicators and influencing factors, and based on China Telecom and China Mobile, the sample data, combined with statistical analysis of sample data to construct a neural network based on improved particle swarm telecom business forecast model;5) Predictive models for various telecommunications services of the experimental results are necessary empirical examination and comparative analysis, confirmed the improved particle swarm neural network prediction system significant results.更多还原