【作者】 赵亮；

【导师】 杨煜普；

【作者基本信息】 上海交通大学 ， 控制理论与控制工程， 2009， 博士

【摘要】 粒子群优化(Particle swarm optimization,PSO)算法是一种基于群体智能的算法,它模拟鸟群、鱼群和蜂群等动物群体的觅食行为,通过个体之间的相互协作使群体达到最优化目的。同遗传算法类似,PSO算法也是一种基于种群的优化技术,它初始化为一组随机解,粒子群在搜索空间中追随种群中的最优粒子进行协同搜索。PSO算法具有操作简单、需要调节的参数少、收敛速度快等特点,因而引起越来越多的关注,成为计算智能及经济、社会、生物等交叉学科的研究热点和前沿。协同搜索的主要思想是使用多个模块同时搜索问题空间,这些模块之间相互交换信息,以提高算法的效率,它在研究大规模复杂优化问题中得到了广泛的应用。本文主要研究了结合协同搜索和PSO算法的协同PSO算法,以及其在模糊辨识和神经网络学习中的应用。本文首先介绍了PSO算法的产生背景、研究的主要内容和开放问题等;接着给出了协同PSO算法的基本框架和参数分析;然后提出了协同随机PSO算法来改进标准PSO算法的性能;最后将提出的改进算法和协同进化PSO算法应用在模糊辨识和神经网络学习等领域。本文研究的主要内容和创新点可概括如下:(1)提出了一种协同随机PSO算法,使用多个子种群同时搜索问题空间,在迭代的过程中,不同的子种群间以随机方式相互交换信息。粒子通过学习不同子种群的最优历史信息来更新自己的速度和位置,保持了种群的多样性;同时,使用多个子种群的有用信息也保证了算法的收敛速度。从而算法的全局和局部搜索能力达到了很好的平衡。(2)提出了一种基于减法聚类和协同随机PSO算法的二阶段模糊辨识方法:减法聚类用来辨识模糊模型的结构,协同随机PSO算法用来优化模型的参数,同时使用减法聚类的结果来初始化算法的种群。该辨识方法能有效地获得紧凑而精确的模糊模型。(3)针对一类用于时间序列预测的单乘法神经元模型,引入协同随机PSO算法来加强其学习能力。单乘法神经元模型可以看作是结构简单,参数较少的神经网络,代替多层神经网络来完成函数逼近等任务。协同随机PSO算法作为该模型的训练算法,提高了模型的学习效率和鲁棒性。(4)针对模糊模型辨识中模型结构难以确定的问题,提出了一种基于协同进化PSO算法的自动模糊模型辨识方法。首先预定义一个最大的模糊规则数,每条规则都有一个标签来决定其是否属于模糊模型;然后将标签、模糊模型的前件参数和后件参数编码成不同的粒子,使用三个PSO算法协同地搜索;通过标签的进化可以得到模糊模型的结构,在模型性能达到最优时,也得到了最优的模型参数。该方法能直接从输入输出数据抽取精确的模糊模型。(5)针对全连接神经网络的结构冗余问题,提出了一种基于协同进化PSO算法的同时调节神经网络结构和权值的方法。首先定义了一种带开关权值的神经网络模型,开关权值由离散值0和1表示,用来决定神经网络的结构;然后使用协同进化的二进制PSO算法和实值PSO算法来优化网络结构和模型参数。该方法能有效地获得精度高而结构精简的神经网络。更多还原

【Abstract】 Particle swarm optimization (PSO) algorithm is a novel swarm intelligence-based algo-rithm which mimics the movement of birds ?ocking or fish schooling looking for food. Thealgorithm finds optimal solution by the cooperation among particles. Similar to the geneticalgorithms, PSO algorithm is an optimal technology based on population, which generatesa set of random solutions and searches the space cooperatively through pursuing the bestexperiences of the swarm. PSO algorithm draws more and more attention and becomes aresearch focus because it has such features as easy operation, fewer tuning parameters, highconvergence speed, etc.Cooperative search is one of the many areas that have been extensively studied in thepast decade to solve many large size optimization problems. The main idea involves havingmore than one search module running and exchanging information among each other inorder to explore the search space more efficiently and reach better solutions. Together withtheir applications in fuzzy identification and neural networks learning, the cooperative PSOalgorithms, which combine the cooperative search and PSO algorithms, are mainly studied.First, the social background, research contents and open problems of PSO algorithmare introduced; next, the basic framework and analysis of parameters are provided; thenthe cooperative random learning PSO algorithm is presented to enhance the performanceof the original PSO algorithm; finally, the improved algorithm and the coevolutionary PSOalgorithm are used in fuzzy modeling and neural network learning. The main contributionsof this dissertation can be summarized as follows:1. A cooperative random learning PSO, which employs several sub-swarms to searchthe space simultaneously, is proposed. In the evolving process, the information among dif-ferent sub-swarms exchanges randomly. The particles learn the best history information toupdate the velocities and positions, thus the diversity of the swarm is maintained. At thesame time, the convergence speed of the algorithm is kept by employing more useful infor-mation during the iteration. The cooperative random learning strategy balances the globaland local search abilities very well. 2. A two-stage fuzzy identification method based on subtractive clustering and coop-erative random learning PSO algorithm is proposed. In the proposed method, subtractiveclustering is utilized to partition the input space and extract a set of fuzzy rules and cooper-ative random learning PSO algorithm is used to find the optimal membership functions andconsequent parameters of the rule base. Furthermore, the subtractive clustering is used toinitialize the swarm of cooperative random learning PSO algorithm. The proposed methodcan extract compact and accurate fuzzy model effectively.3. The cooperative random learning PSO algorithm is introduced into the single mul-tiplicative neuron (SMN) model to enhance its learning ability for time series prediction.The SMN model can be considered as a neural network with simple structure and fewer pa-rameters to take the place of the multiple layers neural network for function approximation.The learning efficiency and robustness of the SMN model are improved through training bycooperative random learning PSO algorithm.4. A new approach for constructing the fuzzy systems automatically by coevolutionaryPSO algorithm is proposed. At first, a maximum rule number is predefined according to priorknowledge. Every rule has a label to determine whether it belongs to the fuzzy inferencesystem. The labels, antecedent and consequent parameters of the model are encoded intothree particles and searched by three PSO algorithms cooperatively. The model structure isobtained through evolving of the labels; meanwhile, the optimal parameters are reached. Theproposed method can construct the accurate fuzzy model from the input-output data directly.5. A coevolutionary PSO-based method to tune the structure and parameters of a neu-ral network for tackling the connections redundant problems is proposed. A neural networkwith switches is introduced firstly, and the switches have two values 0 and 1. The struc-ture of a neural network is decided by the switches. Then, coevolutionary PSO algorithm,which uses a binary PSO algorithm and a standard PSO algorithm, is employed to optimizethe structure and parameters of the proposed model. This method can obtain an accuratepartially-connected neural network effectively.更多还原