节点文献

人工鱼群智能优化算法的改进及应用研究

Research on the Modified Artificial Fish Swarm Optimization Algorithm and Its Applications

【作者】 张梅凤

【导师】 邵诚;

【作者基本信息】 大连理工大学 , 控制理论与控制工程, 2008, 博士

【摘要】 为能更有效地解决工业生产过程中大量存在的优化问题,自20世纪80年代以来,涌现出了一些智能优化算法,它们通过模拟某一自然现象或过程而发展起来,为解决复杂系统的优化问题提供了新的思路和手段,自诞生就引起了国内外学者的广泛关注并被应用于许多领域。人工鱼群算法(Artificial Fish Swarm Algorithm,AFSA)是源于对鱼群觅食行为研究而提出的一种新型群体智能优化算法。该算法具有对初值和参数选择不敏感、鲁棒性强、简单、易于实现,且具备并行处理能力和全局搜索能力等方面的特点。但AFSA在应用过程中还有很多不完善的地方,如:算法后期收敛速度慢,搜索精度不高,在多峰问题寻优时难以找到全部最优解等等。并且,AFSA的应用还不够深入。为此,本文着重从AFSA的改进和应用方面进行了研究。主要研究工作如下:(1)针对AFSA在较大或变化平坦的区域寻优时,收敛于全局最优解的速度减慢、搜索性能劣化,特别是在优化后期往往收敛较慢的问题,提出了一种基于变异算子与模拟退火混合的人工鱼群优化算法。该算法保持了AFSA简单、易实现的特点,同时克服了人工鱼漫无目的随机游动或在非全局极值点大量聚集的局限性,显著提高了运行效率和求解质量,为解决复杂寻优问题提供了有效方法。通过函数和实例测试验证,表明该算法是可行和有效的。(2)针对AFSA在多峰问题寻优时难以找到全部最优解及精度不高的问题,提出了一种基于生境人工鱼群算法的多峰问题优化算法。该算法融合了模拟退火、小生境技术的思想,并加入了变异算子和自动生成合适小生境半径机制。通过对几种典型多峰函数的测试,表明该算法不仅能有效、精确找出多峰问题的全局和局部所有最优解,而且无需预先设置小生境半径,实现了真正的自适应搜索,较好地解决了复杂多峰优化问题。(3)针对连续属性样本分类挖掘时需离散化预处理,可能导致原始信息的缺失问题,提出了基于人工鱼群算法的分类规则挖掘算法,给出了适用于AFSA的分类规则编码方案、构造了新的准确提取规则集的分类规则适应值函数。该算法从优化的角度来解决分类问题,自动实现连续属性样本分类规则的挖掘,从而为连续属性样本提供了一个不需要离散化处理而直接进行数据挖掘的新方法。实验结果表明,该算法不仅能够挖掘出简洁、易于理解的规则集,而且具有较强的鲁棒性和较高的准确率,是一种可行和有效的分类规则优化算法。(4)针对神经网络需要依靠经验确定网络结构及其优化问题,设计了一种基于人工鱼群算法的网络分类器。该方法把输入属性选取和网络结构设计结合,通过人工鱼群算法寻优,同时实现了输入属性选择、神经网络结构和参数的优化。实验表明,该算法能够获得一个具有性能可靠、较好泛化能力的简单分类器,避免了一般神经网络依靠经验确定网络结构的困难,拓宽了AFSA的应用领域。(5)在对AFSA研究和改进的基础上,结合国家863项目“太阳能生物制氢技术研究”,在部分实验所获得的样本数据基础上,引入全局寻优人工鱼群优化算法,通过AFSA优化神经网络结构,获得影响生物制氢的最相关因素,建立了基于优化神经网络的光合细菌制氢过程模型;再用AFSA对已确定的主要工艺条件进行优化,获得了最大制氢量的最佳工艺条件。实验结果表明所提出的优化计算方案可行,此项研究为太阳能光合细菌制氢工艺技术优化探索了一条新的途径。本论文是在国家“十五”863计划项目“太阳能生物制氢技术研究”(编号:2004AA515010)和国家自然科学基金项目“光合生物制氢体系的热效应及其产氢机理研究”(编号:50676029)资助下开展的科学研究。

【Abstract】 In order to solve the optimization problems extensively existing in the industrial processes, intelligent optimization algorithms have been developed by simulating the certain nature and social processes since the 1980s, which provide new approaches to get the optimization of some complex systems. Intelligent optimization algorithms have attracted a lot of attentions from researchers around the world and have been applied in many areas. Artificial Fish Swarm Algorithm (AFSA) is a new kind of swarm intelligent bionic algorithm based on the "looking for food" behaviour of fish swarm. AFSA has been proved to have many advantages, such as insensitivity to the initial values and parameters varying, easy implementation, the abilities of parallel processing and global search, and so on. However, some shortcomings exist in AFSA, such as slower convergence speed, hard to local all optima for multimodal problem. So it is very significant to improve basic AFSA to solve concrete engineering problems. The main contents of the dissertation are as follows:(1) When using AFSA to search optimization in a larger and smoother region, the algorithm has the problems of slower speed of convergence to the global optimum and weaker search ability, especially near to the optimum. This paper proposes a hybrid artificial fish swarm optimization algorithm based on the mutation operator and the simulated annealing. The implementation of the hybrid algorithm is as simple as that of AFSA, and the algorithm can also overcome the limitations of artificial fish stochastic moving without a definite purpose or gathering around the local optimum solution. The operation efficiency and searching ability of the hybrid algorithm are greatly improved, which gives an effective method to solve the problem of complex searching optimization. The feasibility and effectiveness of the hybrid algorithm are verified by the test to function and practical problem.(2) It is difficult to find all of the optimum when AFSA is used in multimodal optimization, so a niche artificial fish swarm algorithm (NAFSA) based on basic AFSA is proposed. NAFSA combines the niche technique and the simulated annealing method with AFSA. Moreover, the ideas of mutation operator and automatic calculating the niche radius are used in NAFSA. NAFSA is applied to the optimizations of some typical multimodal functions. The experimental results show that NAFSA can locate all of the optimal solutions including the global ones and local ones effectively and accurately. Furthermore, NAFSA not only has the good performance, but also can realize self-adapting searching.(3) When mining continuous attributes classification rules, the discrete pre-process is needed, which will cause the decrease of the accuracy of original information data. A classification rules mining algorithm based on AFSA is proposed. To make it suitable to AFSA algorithm, a new classification rule coding is designed and a function is defined to evaluate the classification rule. The algorithm solves the classification problem from the perspective of optimization and implements the classification rules mining of continuous attribute samples automatically, which presents a new approach to mine continuous data directly. The simulation results show that the proposed algorithm can mine better classification rules, including rule sets with higher accuracy, stronger robustness, the smaller number of rules, and simpler rule with fewer terms.(4) Aiming at the problem of determining the neural network architecture by experience, a network classifier is proposed based on AFSA. The algorithm combines the selection of input attributes and design of network architecture. The choice of input attributes, network architecture and parameters optimization are realized by AFSA, simultaneously. The experimental results demonstrate that the algorithm can achieve a simpler classifier which has a more reliable performance and better generalization ability. The difficulty of determining neural network architecture by experience is overcome. The application areas of AFSA are also extended.(5) On the basis of the research and improvement of AFSA, some sample data are obtained from the experiments of the project on the biological hydrogen production technology with solar energy supported by national "863" plan, AFSA is employed to optimize the topology structure of neural network. The main parameters, which influence the hydrogen production quantity, are obtained. The process is modelled based on optimization neural network. Finally, AFSA is used to optimize the main process conditions of hydrogen production, which can ensure to obtain the optimal conditions in which the maximum hydrogen production quality can be obtained. The experimental results show that the proposed optimization computation project is feasible. The research provides a new way for the technology optimization of biological hydrogen production by solar energy.The dissertation is a series research supported by national "863" plan (No. 2004AA515010) and the National Natural Science Foundation (No. 50676029).

节点文献中: 

本文链接的文献网络图示:

本文的引文网络