首页> 外文OA文献 >Fuzzy neural networks with genetic algorithm-based learning method
【2h】

Fuzzy neural networks with genetic algorithm-based learning method

机译:基于遗传算法的模糊神经网络学习方法

代理获取
本网站仅为用户提供外文OA文献查询和代理获取服务,本网站没有原文。下单后我们将采用程序或人工为您竭诚获取高质量的原文,但由于OA文献来源多样且变更频繁,仍可能出现获取不到、文献不完整或与标题不符等情况,如果获取不到我们将提供退款服务。请知悉。

摘要

This thesis is on the reasoning of artificial neural networks based on granules for both crisp and uncertain data. However, understanding the data in this way is difficult when the data is so complex. Reducing the complexity of the problems that these networks are attempting to learn as well as decreasing the cost of the learning processes are desired for a better prediction. A suitable prediction in artificial neural networks depends on an in-depth understanding of data and fine tracking of relations between data points. Inaccuracies of the prediction are caused by complexity of data set and the complexity is caused by uncertainty and quantity of data. Uncertainties can be represented in granules, and the reasoning based on granules is known as granular computing. This thesis proposed an improvement of granular neural networks to reach an outcome from uncertain and crisp data. Two methods based on genetic algorithms (GAs) are proposed. Firstly, GA-based fuzzy granular neural networks are improved by GA-based fuzzy artificial neural networks. They consist of two parts: granulation using fuzzy c-mean clustering (FCM), and reasoning by GAbased fuzzy artificial neural networks. In order to extract granular rules, a granulation method is proposed. The method has three stages: construction of all possible granular rules, pruning the repetition, and crossing out granular rules. Secondly, the two-phase GA-based fuzzy artificial neural networks are improved by GA-based fuzzy artificial neural networks. They are designed in two phases. In this case, the improvement is based on alpha cuts of fuzzy weight in the network connections. In the first phase, the optimal values of alpha cuts zero and one are obtained to define the place of a fuzzy weight for a network connection. Then, in the second phase, the optimal values of middle alpha cuts are obtained to define the shape of a fuzzy weight. The experiments for the two improved networks are performed in terms of generated error and execution time. The results tested were based on available rule/data sets in University of California Irvine (UCI) machine learning repository. Data sets were used for GA-based fuzzy granular neural networks, and rule sets were used for GA-based fuzzy artificial neural networks. The rule sets used were customer satisfaction, uranium, and the datasets used were wine, iris, servo, concrete compressive strength, and uranium. The results for the two-phase networks revealed the improvements of these methods over the conventional onephase networks. The two-phase GA-based fuzzy artificial neural networks improved 35% and 98% for execution time, and 27% and 26% for the generated error. The results for GA-based granular neural networks were revealed in comparison with GA-based crisp artificial neural networks. The comparison with other related granular computing methods were done using the iris benchmark data set. The results for these networks showed an average performance of 82.1%. The results from the proposed methods were analyzed in terms of statistical measurements for rule strengths and classifier performance using benchmark medical datasets. Therefore, this thesis has shown GA-based fuzzy granular neural networks, and GA-based fuzzy artificial neural networks are capable of reasoning based on granules for both crisp and uncertain data in artificial neural networks
机译:本文是基于颗粒的人工神经网络对明快数据和不确定数据的推理。但是,当数据非常复杂时,很难以这种方式理解数据。为了更好的预测,期望降低这些网络试图学习的问题的复杂性以及降低学习过程的成本。人工神经网络中的合适预测取决于对数据的深入了解以及对数据点之间关系的精确跟踪。预测的不准确性是由数据集的复杂性引起的,而复杂性是由数据的不确定性和数量引起的。不确定性可以用颗粒表示,基于颗粒的推理称为颗粒计算。本文提出了一种改进的粒状神经网络,以实现不确定和清晰数据的结果。提出了两种基于遗传算法的方法。首先,基于遗传算法的模糊人工神经网络对基于遗传算法的模糊颗粒神经网络进行了改进。它们包括两个部分:使用模糊c均值聚类(FCM)进行粒化,以及基于GA的模糊人工神经网络进行推理。为了提取粒度规则,提出了一种造粒方法。该方法分为三个阶段:构造所有可能的细化规则,修剪重复以及删除细化规则。其次,基于遗传算法的模糊人工神经网络改进了基于遗传算法的两阶段遗传算法。它们分两个阶段进行设计。在这种情况下,改进基于网络连接中模糊权重的alpha削减。在第一阶段中,获得的alpha最优值分别为零和一个,以定义网络连接的模糊权重的位置。然后,在第二阶段,获得中间alpha切口的最佳值以定义模糊权重的形状。针对两个改进的网络的实验是根据生成的错误和执行时间进行的。测试的结果基于加利福尼亚大学欧文分校(UCI)机器学习存储库中的可用规则/数据集。数据集用于基于GA的模糊颗粒神经网络,规则集用于基于GA的模糊人工神经网络。使用的规则集是客户满意度,铀,使用的数据集是酒,虹膜,伺服,混凝土抗压强度和铀。两相网络的结果表明,与传统的单相网络相比,这些方法得到了改进。基于两阶段GA的模糊人工神经网络的执行时间提高了35%和98%,产生的错误提高了27%和26%。与基于GA的清晰人工神经网络相比,揭示了基于GA的颗粒神经网络的结果。使用虹膜基准数据集与其他相关的粒度计算方法进行了比较。这些网络的结果显示平均性能为82.1%。使用基准医学数据集,根据规则强度和分类器性能的统计度量分析了所提出方法的结果。因此,本文证明了基于GA的模糊颗粒神经网络,基于GA的模糊人工神经网络能够基于颗粒对人工神经网络中的明晰数据和不确定数据进行推理。

著录项

  • 作者

    Mashinchi M. Reza;

  • 作者单位
  • 年度 2013
  • 总页数
  • 原文格式 PDF
  • 正文语种 en
  • 中图分类

相似文献

  • 外文文献
  • 中文文献
  • 专利

客服邮箱:kefu@zhangqiaokeyan.com

京公网安备:11010802029741号 ICP备案号:京ICP备15016152号-6 六维联合信息科技 (北京) 有限公司©版权所有
  • 客服微信

  • 服务号