【作者】 季伟东；

【导师】 王克奇；

【作者基本信息】 东北林业大学 ， 机械设计及理论， 2013， 博士

【摘要】 系统仿真领域的研究难点和研究热点之一是对复杂系统进行仿真及建模研究。人工神经网络通过大量互连神经元之间连接权值的分布存储来表示网络的解题知识,具有并行处理、自学习和能以任意精度逼近任意非线性函数等特性,神经网络具有的这些优点使其适宜于复杂系统建模中。但神经网络在应用中还存在一些问题,譬如网络学习局部极小,收敛速度缓慢,网络结构设计复杂及泛化性能弱等缺点,神经网络的这些缺陷阻碍了其应用。神经网络学习从本质上说是网络结构与权值的优化过程,神经网络这些局限性都不可避免的会牵扯到网络学习中的优化问题。进化计算是一类模拟生物进化过程与机制求解问题的人工智能技术,进化计算具有智能性和本质并行性两大特点,它不受对象目标函数连续可微条件约束,在一些离散、多态、含噪声的优化问题中更显示出该方法独特的优势,进化计算具有的这些优点为解决神经网络自身缺陷提供了一条全新可行途径。本文利用群体智能中的进化计算来解决神经网络在复杂系统建模中存在的主要问题,并将其应用于木材表面缺陷识别。主要研究内容如下：首先,针对网络结构固定情况下传统BP学习方法对网络学习易陷入局部极值、学习精度不高等缺陷,提出一种基于BP算子的粒子群(PSO)优化神经网络权值的学习方法。该方法有效结合粒子群算法和BP算法优点,在PSO全局勘探和BP局部开发之间能给出一个较好的折中；更重要的是该结合方式能给PSO进化和BP搜索之间提供一个相互充分作用的机会,即在PSO每代寻优解的基础上执行BP操作可以获得精度更高的优化解,同时该优化解又返回到粒子群优化过程并共享其位置信息引导群体的快速进化,两种算法之间相辅相成,相互促进来达到共同的寻优目的。实验结果表明该学习方法学习精度高,收敛速度快,在性能上优于常规神经网络的学习方法。其次,针对单一进化算法对复杂优化问题存在早熟收敛、优化效率低的问题,提出一种实数编码遗传算法与粒子群优化方法相结合的改进混合遗传算法(IHGA),该方法引入生态学中生态位构建思想,使得个体具有一定程度上的学习能力。采用遗传算法和粒子群算法的不同搜索机制共同作用于父代个体生成子代个体,一方面可以保持种群个体的多样性,同时也在一定程度上避免了算法早熟收敛。理论证明IHGA具有概率为1的全局收敛性,实验结果也表明该方法显著提高了算法优化性能及其优化解的可靠性。在此基础上,引入带连接开关的神经网络,提出一种基于IHGA的网络结构与权值并行调整优选方法,仿真结果表明该方法设计的网络学习精度高,结构节俭。然后,研究了径向基网络(RBF)的设计问题,提出了一种基于PSO的RBF网络新设计方法,该方法由正则化正交最小二乘法与D-最优试验设计结合算法自动构建结构节俭的RBF网络模型；通过改进粒子群优化算法优选结合算法中影响网络泛化性能的三个学习参数,即基宽度参数、正则化系数和D-最优代价系数的最佳参数组合。仿真实例表明了该方法是一种较好的RBF网络设计方法。在木材缺陷检测问题的研究工作中,表面缺陷检测技术一直占有重要的地位。木材表面缺陷模式识别通常采用人工神经网络技术,与其它检测技术相比,神经网络能够提高缺陷检测的准确率和检测效果,但同时仍然不能摆脱神经网络自身缺陷限制,使得其识别率不高。最后,将本文所提出的学习方法应用于木材表面缺陷识别中,与其它常见的网络学习方法做比较,并分析本文方法优缺点。结果表明采用本文方法构建的网络模型具有较高识别精度,且结构节俭,表明了本文提出的神经网络学习方法适宜应用于复杂网络模型建模中。更多还原

【Abstract】 The research of simulation and modeling for the Complex System is one of a difficulty and a hot point in the System simulation research. The Artificial neural network, represents the network problem-solving knowledge through a large number of neurons which are connected to each other of using weights,has the characteristics of self-learning, parallel processing and also it can approximate any nonlinear function with arbitrary precision. The neural network is suitable for modeling complex systems because of its these advantages.However,there are some problems in the application of the neural network, for example, the minimum of network learning, the slow convergence speed,the compelx of network structure design and the weak of the generalization performance, all of these limitations prevent the neural network from applying in application fields. Neural network learning is essentially the optimization of network structure and weights, these limitations of the neural network will involve the optimal problems of network learning inevitably. The evolutionary computation is one of the artificial intelligence technology to simulate the evolution process and mechanism. The evolutionary computation is intelligent and parallelism of these two characteristics, and it is not affected by the target function which is continuously differentiable constrains,and it will show the unique advantage of this method in the problems of discrete,polymorphism and noise.The advantages of the evolutionary computation provide a new feasible way of solving the defects of neural network.In this paper, we use evolutionary computation of the swarm intelligence to solve the main problems of neural networks in the modeling of complex systems, and also applied to the wood surface defect recognition. Main contents are as follows:First of all, aiming at the problem that back-propagation (BP) learning algorithm is trapped easily in local minima and the BP learning precision is not high, a method for optimizing weights of neural network based on BP and particle swarm optimization is proposed. This method combines particle swarm optimization with back-propagation, and gives a good tradeoff between PSO global exploration and BP local exploitation. More importantly, this hybrid provides an opportunity to interact fully between PSO and BP, that is, BP operation can obtain better solution based on the result of PSO in each generation, meanwhile this solution returns back the PSO swarm and guides the swarm evolution rapidly through sharing its position information for swarm. The two algorithms supplement and promote each other to achieve the optimization purposes. Some simulation experiments show that this method has merits of high precision and rapid convergence speed, and its learning performance outperforms some prevail algorithms as to the neural network with fixed structure. Single evolutionary algorithm has premature convergence, the low optimization efficiency, in complex optimization problems, so a improved hybrid genetic algorithm (IHGA), which is a combination of real-coded genetic algorithm and particle swarm optimization method, has been put forward.IHGA Introduces the idea of ecology niche construction, and makes individuals have the ability to learn. Different search mechanisms of genetic algorithm and particle swarm optimization to generate offspring, can maintain the diversity of the individuals,and to some extent avoid the algorithm premature convergence. Theory proved IHGA to have a probability of a global convergence.Experimental results also show that this method significantly improves the reliability of the algorithm to optimize performance and its optimal solution. On this basis, the neural network with the connection switch has been introduced and a method that simultaneously tunes structure and parameters of the neural network based on IHGA is proposed. The simulation results show that the network learning designed by the method has high precision and structure thrift.Study on the radial basis function network (RBF) design problems, a new design method of RBF network based on PSO has been put forward. The method,which consists of regularized orthogonal least squares method and D-optimal experimental design combined algorithm automatically build the structure frugal RBF network model. Simulation examples show that the method is a better RBF network design method. Improved three learning parameters, base width parameter,regularization parameter,and D-optimal cost of coefficients,which affect generalization performance.The particle swarm optimization is used to search the optimal combination of three important learning parameters, i.e., the RBF width, the regularized parameter and D-optimality weight parameter, which influence the network’s generalization ability.Simulation examples show that the method is a better RBF network design method.Research in the wood defect detection problem, surface defects detection technology has always played an important role. The pattern recognition of surface defection commonly used artificial neural network technology. Compared with other detection technologies, the neural network can improve the accuracy and efficiency of detection, But still can not get rid of the neural network defect limits, which let the recognition rate be low. Finally, the proposed learning method is applied to the wood surface defect recognition and compared with other common network learning methods and analyze the advantages and disadvantages of this method. The results show that the network models constructed using the proposed method has higher recognition accuracy and frugality structure,and in this paper the neural network learning is a suitable method applied to a complex network modeling.更多还原