【作者】 李静宜；

【导师】 田雨波；

【作者基本信息】 江苏科技大学 ， 信号与信息处理， 2010， 硕士

【摘要】 多层前馈神经网络是应用最广泛的一种神经网络,然而它的理论和应用都还存在一些困难,如易陷入局部极小点和泛化能力差等问题。蚁群优化(Ant Colony Optimization,简称ACO)算法是一种良好的群智能算法,它具有良好的全局性和并行性,因此用蚁群优化算法训练前馈神经网络的权值和阈值,能够使神经网络脱离局部极小,提高网络的泛化能力。近年来,随着人们对带宽需求的不断增加以及通信技术的不断发展,微带天线的带宽技术和小型化等已成为目前研究的一个热门话题。微带天线的许多优点,例如体积小、重量轻、剖面薄、易集成以及低成本等,使微带天线得到了广泛的应用,因此研究高性能的微带天线很有意义。本文首先简要介绍了神经网络及蚁群优化算法的基本知识,然后对蚁群优化算法的更新方法进行了研究,其次又构建了三种用蚁群优化算法或蚁群优化算法和其他算法相结合训练神经网络权值和阈值的模型(以下简称蚁群神经网络),并通过函数拟合、LED分类和广义异或问题验证了这三种模型的性能,最后用效果最好的一种蚁群神经网络模型进行了矩形微带天线谐振频率的计算和一种I型微带天线的优化设计。论文的主要研究成果可归纳如下:(1)为克服蚁群优化算法在寻优时仍有可能陷入局部极小的现象,给出了基于代间差分和混沌变异等更新算法。由实验结果可知,更新算法可以有效克服标准蚁群优化算法的早熟现象,并且能够加快收敛速度,达到跳出局部极小点,获得全局最优的目的。(2)构建了三种蚁群神经网络(ACONN)模型。这三种模型分别为用蚁群优化算法训练前馈神经网络权值和阈值的ACO_NN模型;先用蚁群优化算法对前馈神经网络的权值和阈值进行训练之后再用BP算法训练前馈神经网络权值和阈值的ACO_BP_NN模型和先用粒子群优化(Particle Swarm Optimization,简称PSO)算法对蚁群优化算法的初始值进行调整,然后再用蚁群优化算法对前馈神经网络的权值和阈值进行训练的PSO_ACO_NN模型。(3)通过函数拟合、LED分类和广义异或问题,验证三种ACONN模型的性能,实验结果证明PSO_ACO_NN模型在处理这些问题时效果均是最好的。(4)将PSO_ACO_NN模型应用于矩形微带天线谐振频率的计算和一种I型微带天线的优化设计上,取得了满意的结果。更多还原

【Abstract】 Multi-layer feed-forward neural network is the most widely used network, but it still has some difficulties in theory and applications, such as the local minimum, the learning and generalization ability. Ant Colony Optimization (ACO) algorithm is a typical intelligent algorithm that has good global optimization and parallelism. Therefore, using the ACO algorithm to train the weight value and threshold value of network can make it possible to escape from the local minimum, and improve its generalization capability.With the development of communication technology and people’s increasing demand for bandwidth, the microstrip antenna’s bandwidth technology and miniaturization have become a hot topic in the current study. Micorsrtip antennas have been used in many fields because of its advantages, such as lightweight, thin profile, packaging, installing, low-cost and so on. Therefore, it is a significant task to research high performance micorsrtip antennas.In this thesis, the basic concepts of neural network and ACO algorithm will be introduced firstly, and then the parameters selection and updating methods of ACO algorithm are studied. Three network models based on ACO algorithm or ACO algorithm connect with other algorithms are constructed (called ACONN in this thesis).The performance of these three models are verified through function fitting, LED classification and general XOR problem. At last use the ACONN model which has the best performance to calculate the resonant frequency of microsrtip antenna and also to design and optimization of an I-shaped microsrtip antenna.The main research results in this thesis can be summarized as follows:(1) To overcome the local minimum of ACO algorithm, this paper adopted generation-difference, chaos and variation of particles, and other updating algorithms. Many stimulation experiments are made. The experimental results show that the methods may effectively overcome premature of ACO algorithm, speed up the convergent rate, escape from local minimum, and approach to global optimum.(2) Three network models are constructed which are called ACONN in this thesis. The ACO algorithm can make the network escape from local minimum limitation, and improve its generalization capability further. ACO_NN is a model which neural network’s weight value and threshold value are trained by ACO algorithm. ACO_BP_NN is a model which neural network’s weight value and threshold value are trained firstly by ACO algorithm and then trained by BP algorithm. PSO_ACO_NN is a model which firstly adjusted ACO algorithm’s initial value by Particle Swarm Optimization (PSO) algorithm, then trained neural network’s weight value and threshold value by ACO algorithm.(3) To verify the performance of three ACONN models through function fitting, LED classification and general XOR problem. The computing results show that PSO_ACO_NN model has the best effects in dealing with these issues.(4) PSO_ACO_NN model is applied to calculate the resonant frequency of micorsrtip antenna. PSO_ACO_NN model is also used to optimize and design the structure of an I–shaped microstrip patch antenna. The simulating results on the two problems are satisfactory.更多还原