【作者】 包芳；

【导师】 须文波；

【作者基本信息】 江南大学 ， 轻工信息技术与工程， 2008， 博士

【摘要】 随着人工神经网络越来越广泛地被应用于各类科研领域,对其优化算法和应用技术的研究成了人工神经网络领域的重点方向。目前,在利用智能算法优化神经网络、应用神经网络实现模糊推理控制这2方面,已经取得了一定进展,但存在着训练数据成几何形增大,应用性能不稳定等缺陷。在充分研究神经网络、群体智能算法、模糊逻辑基本概念的基础上,本文利用新的群体智能算法优化神经网络结构和学习规则,实现有监督模糊聚类、自适应神经-模糊控制,做了以下工作:1、提出了一种基于QPSO(Quantum Particle Swarm Optimization)的神经网络训练算法,并对神经网络进行训练,将其结果与传统神经网络训练算法相比较,该算法具有收敛速度快,对数据离散程度高的问题更易实现全局收敛的特点。2、提出了基于群体智能集成的神经网络结构优化算法,采用间接编码方案表达网络结构,利用二维细胞自动机中的元细胞表示网络连接,对元细胞的坐标和值分别演化,实现网络结构的生长和剪枝,其中包括:利用二进制量子化粒子群算法BQPSO(Binary Quantum Particle Swarm Optimization),导出特定的适应度函数,生成并进化元细胞的坐标;利用元细胞邻域演化规则,演化元细胞的值;利用浮点量子化粒子群算法QPSO训练当前网络;最后得到最终稳定的网络结构及其参数。本算法采用元细胞的坐标和值分别演化,突破了目前常见的网络结构设计算法中,结构编码方案的长度成几何形膨胀、且不易实现的局限性。3、在研究有监督模糊聚类SFCM(Supervised Fuzzy C-Clustering)神经网络的基础上,提出了一种新的结合输入空间聚类特性和输出空间实时逼近特性的模糊聚类目标函数,在模糊聚类中引入监督因素,构造并实现一种新型的模糊聚类神经网络。该方法在初值敏感性、有效收敛稳定性、收敛速度等方面相对于传统模糊聚类都有明显改善。4、针对机器人在未知、复杂环境下从源到目标之间,自主避开各种类型的障碍的问题,设计了神经-模糊控制算法实现动态路径规划,包括:模糊控制体系,实现输入模糊化、模糊推理规则库、输出去模糊化控制;根据规则库设计实现模糊控制的神经网络,简化了网络结构和参数集;基于QPSO的网络训练;机器人状态变量的记忆、环境信息的设置和管理策略,解决了“U”型障碍物内的死循环路径问题。本算法有效的解决了常规神经-模糊控制系统中,有可能造成机器人死循环路径的“U”型障碍物跨越问题,以及因网络规模大和采用传统的梯度下降训练法训练网络而带来的性能问题。更多还原

【Abstract】 Along with wider application of artificial neural network, the optimization algorithm and the application technique of it have become the most important research orientation in the field. At present, certain progress has been made in the fields of neural network optimization using intelligent algorithm and fuzzy logic control using neural network, but there still exists the limitation of geometrically growing training data and the unstable performances. Based on sufficient research in the basic conceptions of neural network, colony intelligent algorithm and fuzzy logic, the optimization of the structure design and learning rule of neural network by using novel colony intelligent algorithm, the realization of supervised fuzzy clustering and adaptive neural-fuzzy control by using optimized neural network, have been discussed in the thesis.1. A training algorithm of neural network based on QPSO is proposed, which train the neural network by means of the quantum particle swarm optimization. The algorithm show better convergent speed and better global convergent characteristic when dealing with more dispersed issue compared with the traditional neural network training algorithm.2. A novel algorithm of neural network structure design based on BQPSO and 2-dimension cellular automate system is proposed. The algorithm introduces unique indirect encoding schema representing the structure of neural network, using the cell in the 2-dimension cellular automate system representing the existence of connection in neural network, by separately evolving the coordinates and value of the cell, the growing and pruning of the network structure is achieved. Create and evolve the coordinates of the cell by virtue of BQPSO with specially-designed fitness function, evolve the value of the cell using properly-designed neighboring evolving rule of cellular system, train current network with float-point QPSO, the final stable structure is found step by step.By separately evolving the coordinates and value of the cell, the proposed algorithm can solve the problem of geometrically growing encoding length and the difficult realization of commonly used structure design algorithm.3. The research of the supervised fuzzy C-clustering neural network. A novel objective function of fuzzy clustering that integrates the clustering characteristic of input space and the real-time approximate characteristic of output space is proposed, thus importing the supervise factor into the former fuzzy clustering. An extraordinary neural network handling the SFCM is also proposed. SFCM has better performance in the stable convergent rate, convergent speed, as well as in the initial condition sensitivity compared with traditional fuzzy clustering.4. The research of the neural-fuzzy control. According to the issue of dynamic path plan of mobile robot in unknown environments from the start to the destination with obstacle avoidance, a systemic neural-fuzzy control algorithm is proposed. Fuzzy logic control is designed to do the input fuzzification, fuzzy reasoning rule base, output defuzzification. The simplified structure of neural network handling the fuzzy control based on the rule base and the corresponding simplified network parameter set is also designed. Train the network using QPSO. Solve the "dead cycle" problem in U-shaped obstacle through the storage and management strategy of state variable of robot.The algorithm solve the problem of big network scale, low network performance via using grade-descend training method in the conventional neural-fuzzy system, and the problem of getting into the dead cycle when striding across the U-shaped obstacle is also solved.更多还原