【作者】 王明兴；

【导师】 钱积新； 陈曦；

【作者基本信息】 浙江大学 ， 系统工程， 2005， 硕士

【摘要】 本文对连续禁忌搜索算法作出改进，提出了两种改进的算法ITS和TS_SQP，并将改进的算法应用于系统辨识以及反馈神经网络的训练。本文主要的研究成果和创新点包括： 1) 对连续禁忌搜索算法进行改进，提出了改进的算法ITS。在ITS算法中，既考虑到多样性搜索策略，将当前点的邻域空间用一组同心超矩形进行划分，在每个外围同心超矩形中随机选取一个点组成部分邻域；又通过改进引入了特赦规则，在中心超矩形内也随机选取一定数量的点，与外围同心超矩形内选取的点共同组成当前点的邻域。通过对一些典型测试函数的仿真结果表明改进算法有助于更快、更精确的搜索到全局最优点。 2) 针对禁忌搜索算法局部搜索的随机性，首次提出了一种与SQP算法结合的禁忌搜索算法，TS_SQP。利用禁忌搜索算法的全局收敛性，结合SQP局部搜索快速收敛的能力，改善传统禁忌搜索算法的搜索能力，使禁忌搜索算法可以获得精确的最优点。在TS_SQP算法中，首先产生当前点的邻域，然后以邻域内的每个点为初始点运行SQP算法，所有收敛点构成新的邻域，最后运用TS规则更新当前点。仿真结果表明，与ITS比较，TS_SQP全局收敛的速度更快，获得的最优点更精确。 3) 将改进禁忌搜索算法(ITS、TS_SQP)应用于系统辨识，以改善传统的辨识办法存在局部极小等缺点，并实现了对包括滞后在内的所有参数同时辨识。通过对液位储罐模型、离散、连续以及高阶系统的仿真实验表明了算法的可行性及有效性。 4) 将改进的禁忌搜索算法(ITS、TS_SQP)应用于反馈神经网络的训练。其方法实质是将神经网络的训练问题转化为优化问题，利用禁忌搜索算法的全局寻优能力获得最优的神经网络权值与阈值。仿真实验表明了该方法具有很好的性能，并且简单易实现。更多还原

【Abstract】 Two improved continuous tabu search algorithms, ITS and TSSQP, are developed for global optimization problems. Their applications on system identification and training of recurrent neural network are also studied. The main results can be summarized as follows:(1) An improved continuous tabu search algorithm, ITS, is developed. In the ITS, the neighbor space of current solution is partitioned by a set of concentric hyperrectangles. Inside each concentric hyperrectangle, one point is randomly selected as a neighbor of the current solution. Considering the intensification strategy an improvement is made by selecting a certen number of points also inside the central hyperrectangle following an aspiration criterion. All selected points generate the neighborhood of the current solution. Numerical simulation results prove that the extra selection inside the central hyperrectangle enables the ITS to obtain more precise global miminum with less time cost.(2) Further improvement of the tabu search is proposed by combine with the SQP method. As a kind of random search algorithm, the performance of TS is limited. The SQP, good at fast convergencey in local search, is introduced to propose a new algorithm, TSSQP. In TS_SQP, the SQP algorithm starts from each point in the neighborhood of current solution and converges a local miminum; All these local minima consititute the new neighborhood of the current solution; then the program continues by exacting ITS. Compared with ITS, the simulation results shuw that TS_SQP can obtain more precise global miminum and costs less time.(3) The improved tabu search algorithms ITS and TS_SQP are applied on system identification. The problems of system identification are transformed to optimization problems in parameter space, and then the ITS and TS_SQP algorithms are used to obtain the global optimal estimation of the system parameters. The results of simulations on a tank model, a discrete and a continuous two-order system, and a highorder system demonstrate the feasibility and effectiveness of ITS and TS_SQP.(4) The improved tabu search algorithms, ITS and TS_SQP, are applied on the training of recurrent neural network. The training problem is cast as optimization problem in parameter space. The numerical simulation results show that they are simple yet effective.更多还原