基于进化计算的神经网络设计方法

【作者】 俞健；

【导师】 蒋静坪；

【作者基本信息】 浙江大学 ， 工业自动化， 1998， 博士

【摘要】 神经网络是在现代神经科学研究成果的基础上提出的用来“模拟”人脑功能基本特征的网络模型，它具有很好的并行处理、学习、非线性映射和泛化能力，并在模式识别、信息处理、设计、规划、诊断和控制等许多领域都得到了研究人员的亲赖。近十年来，基于神经网络的控制方法的研究取得了很大的进展，相应的硬件产品也已经上市。随着应用研究的不断深入和扩展，人们也碰到了1系列急需解决的问题，例如网络类型、结构、参数集的选择，学习算法的选择，基于神经网络的控制系统稳定性、鲁棒性分析等等。本文主要研究的是如何选择网络结构和参数集这个问题。 第1章是绪论。首先简单地回顾了控制理论的发展，介绍了智能控制研究的主要内容。通过这些介绍，主要是为了说明控制理论研究和应用中面临的挑战和基于神经网络的控制方法具有的巨大潜力。接着比较详细地论述了神经网络在控制系统中所起的作用，给出了具体的控制系统结构，并且指出了神经网络研究中需要澄清的3个模糊认识和需要解决的5个典型问题。最后简单地介绍了神经网络硬件研究的现状。 第2章提出了1种基于进化算法的神经网络结构设计方法。首先阐述了神经网络结构设计的意义，然后总结了神经网络结构设计的常用方法。神经网络的结构设计可以看成是1个寻优过程，而基于进化思想的进化算法正是1类很好的优化方法，所以利用进化算法获得较佳的网络结构就成为我的研究方向。利用进化算法设计网络的结构，首先要确定网络结构的编码方式，然后根据编码方式来决定选用何种进化算法。按照我提出的网络结构的编码方式，只有使用进化规划的方法才能完成结构寻优的任务。结构寻优的过程是和网络训练同步进行的，由于网络结构不是很规则，所以无法使用“梯度法”等优化手段，为此，网络权值训练采用的是遗传算法。通过解异或问题和逼近1个非线性函数的仿真实验，证明了这种基于进化算法的神经网络结构设计方法的可行性。 第3章提出了1种利用遗传算法设计BP网络的方法。通过引入正交试验设计的思想，重新定义了遗传算法中的交叉操作。新的交叉操作产生的子代个体多于2个，这些个体通过内部竞争优选出2个作为交叉操作的最终结果进入遗传算法的下1个步骤。在精心安排下，双亲中适应性强的1个也自动参与了内部竞争，这就保证了子代个体中至少有1个个体的性能不差于父辈。故障诊断的仿真实验初步证明了该方法的有效性。神经网络的训练过程和网络结构的寻优过程都是十分耗费时间的，为了寻找解决这个难题的方法，我尝试利用DSP／PC机组成的典型的主从式系统来实现自己提出的网络设计方法。通过对CSTR系统中状态估计问题的仿真研究，一方面进1步证明了设计方法的可行性，另一方面也在寻找缩短寻优时间的途径上进行了1次成功的尝试。更多还原

【Abstract】 Artificial neural networks try to mimic the nerve system in a mammalian brain into a mathematical model. Therefore, neural networks have some desirable characteristics and capabilities similar to the brain system, such as parallel processing, learning, non-linear mapping, and generalization. Many researchers have developed neural networks as new tools in many fields such as pattern recognition, information processing, design, planning, diagnosis, and control. The past decade have witnessed a great deal of progress in both the theory and the practice of control using neural networks. After a long period of experimentation and research, neural network-based controllers are now being realized in a wide variety of fields. The practice applications are also calling for a better understanding of the theoretical principles involved. There are some problems which always make users puzzle, such as deciding the suitable types, structures and parameters of neural networks, selecting a good method for training, and analyzing the stability and robust of neural network-based control systems. In this paper, I focus my effort on how to obtain the suitable structures and parameters of neural networks.In Chapter One, I give the complete introduction of neural networks used in the field of control. Through the review of the development of control theory and the research directions of intelligent control, I point out the challenges which the control theory is facing, and I also point out the great potential of the control methods based on neural networks. After describing the structures of control systems based on neural networks carefully, I clarify some specious thoughts and summarize the problems needed to be solved urgently. In the end, I introduce the research situation of the hardware implementation of neural networks simply.In Chapter Two, I put forward a structural design method based on evolutionary algorithms. At first, I elaborate the significance of structural design, and then set out all kinds of the structural design methods which are familiar to me. Designing suitable structures for neural networks belongs to a kind of optimization questions, so it is natural to look for good optimization techniques. In recent years, evolutionary algorithms are popularly used for their population-based optimization mechanism. Genetic algorithm is well-known among them, and evolutionary programming employs the same Darwinian evolutionary principles as genetic algorithm and is implemented by the interaction of individuals in a population. These two algorithms are somewhat similar, but genetic algorithm relies on genetic operators, while evolutionary programming emphasizes the performance change from the population level. Furthermore, when evolutionary programming is used for optimization, mutation is the unique recombination operator. After taking the encoding method into consideration, I think evolutionary programming is most suitable for my research work. In general, structural evolution is accompanied by the training of connection weights. For sake of the irregularity of structures, the optimization methods based on gradient-descent can’t be used to optimize the connection weights in a fixed structure, so I select genetic algorithm to finish this task. Finally, simulation results are given to illustrate the efficiency of the proposed method.更多还原