基于决策树算法确定一个最小神经网络

神经网络隐层神经元的个数对于网络的性能有着重要的影响,通常情况下,对于一个特定问题来说,没有一个确定的方法来决定隐含层到底应该有多少个神经元,一般采用试探的方法通过多次实验来达到理想效果.在分类问题中,决策树和神经网络的结构有着一定的关联性,通过把决策树映射到神经网络结构中来确定隐层神经元的个数的方法能够有效地设计神经网络的结构,从而提高训练的效率并达到良好的分类效果.实验结果表明,该方法能够得到一个有着良好识别率的最小神经网络.方法简单有效,直观且易于操作.     

基于熵的自组织神经网络树

神经网络由于优越的学习和分类能力已被用于许多模式识别的问题,并取得了很好的结果。但是对于识别大样本集和复杂模式的问题,绝大多数常规的神经网络在决定网络的结构和规模以及应付庞大的计算量等方面有着种种困难。为了克服这些困难,文中提出一种基于条件类别熵的结构自适应的神经网络树;这种神经网络树由具有拓扑有序特性的子网络组成,而树的规模由条件类别熵决定。它的主要优点是对于识别大样本集和复杂模式的问题能够通过结构自适应自动地确定网络的结构和规模。实验显示这种神经网络树对于识别大样本集和复杂模式是非常有效的。     

基于神经网络的结构损伤状态模拟

识别一个结构在震动状态下的变化,在结构监测中是非常重要的,神经网络就非常适用于这种目的。本文研究了使用可分析的学习样本来训练神经网络的可行性问题,神经网络从损伤状态中训练产生,然后用于诊断一个五层钢框架在一系列震动模拟中的状态。结果表明,使用神经网络可使在线结构诊断更加可行。     

基于参数调整的动态模糊神经网络算法

模糊逻辑与神经网络结合形成的模糊神经网络同时具有易于表达人类知识、存储与学习分布信息的优点,基于此,提出一种基于参数调整的动态模糊神经网络算法。采用扩展卡尔曼滤波器法将全局算法划分为线性和非线性部分,线性参数由最小二乘法和滤波器法决定,非线性参数由训练样本和启发式法直接决定,线性和非线性参数可进行实时更新。仿真结果表明,该算法能保证更简洁的结构和更短的学习时间。     

配煤过程基于神经网络和数学模型的专家控制

介绍一种专家控制策略,它解决配煤过程控制中精确地决定和跟踪每种煤的目标配比这两个关键问题。首先基于统计数据和经验知识,建立神经网络,数学模型和规则模型。然后提出一种决定目标配比的专家推理方法,它结合了神经网络,数学模型和规则模型。最后使用一种分布式PI控制方案,跟踪决定的目标配比。这种专家控制策略由一个专家控制器和一个分布式控制器实现,实际的运行结果证明了这种策略的有效性。     

基于神经网络的结构损伤状态模拟

识别一个结构在震动状态下的变化在结构监测中是十分重要的，神经网络能用于这种目的．本文研究了使用可分析的学习样本来训练神经网络的可行性问题。神经网络从损伤状态中训练产生，然后用于诊断一个5层钢框架在一系列模拟震动测试中的损伤状态。结果表明，使用神经网络可使在线结构诊继更为可行。     

进化神经网络的编码表示机制研究

神经网络学习包括权值学习和拓朴结构学习,目前大部分妍究集巾于权值学习,结构的选择和确定一般是靠经验和专家知识。设计一个好的神经网络招扑结构是非常困难的。一般地,学习,连同选择一个优化网络招扑结构,是一个NP宇全问题。神绎     

基于深度学习的八类蛋白质二级结构预测算法

蛋白质二级结构预测是结构生物学中的一个重要问题。针对八类蛋白质二级结构预测,提出了一种基于递归神经网络和前馈神经网络的深度学习预测算法。该算法通过双向递归神经网络建模氨基酸间的局部和长程相互作用,递归神经网络的隐层输出进一步送入到三层的前馈神经网络以便进行八类蛋白质二级结构预测。实验结果表明,提出的算法在CB513数据集上达到了67.9%的Q₈预测精度,显著地优于SSpro8和SC-GSN。     

基于自适应并联结构神经网络的交通流量预测

由于现有优化算法在全局优化方面的局限性,导致神经网络需要多次训练才能获得满意的结果。为了解决神经网络训练中的一致性问题,文章提出了一种自适应并联结构神经网络（Adaptive Parallel Structure Neural Network, APSNN）。APSNN由多个神经网络单元并联组成,在训练过程中,采用常规优化算法对各神经网络单元进行训练。神经网络单元的训练样本由上一级神经网络单元的训练残差构成,通过训练残差在各神经网络单元中的单向传递,实现训练残差的逐级减少。神经网络根据训练残差,决定是否进行神经网络单元级联和结构扩张,从而保证训练结果的一致性。文章对5种非线性函数进行了神经网络逼近测试。与BP神经网络相比较,APSNN在50次不同初始条件下,训练精度十分稳定,具有很好的一致性。为了实现对交通流量预测,文章将APSNN与BP神经网络和小波神经网络进行了对比研究,结果表明：APSNN的预测总体标准差均小于BP神经网络和小波神经网络,交通流量的预测偏差较BP神经网络和小波神经网络分别降低2.7%和9.7%。     

神经网络规则抽取研究

尽管神经网络已经在很广泛的领域得到应用,但由于训练好的神经网络中的知识不易于理解,神经网络被视为一个典型的“黑箱”结构。从神经网络中抽取规则来表示其中隐含的知识是解决个问题的一个有效的手段,将对一些具有代表性的神经网络规则抽取算法进行综述分析,并提出一些未来的研究重点。     

实现无线传感器网络规模可扩展性的一种算法

无线传感器网络是由许多分布的智能传感器节点组成的多跳无线网络。针对无线传感器网络的特点,在能量有效的MAC协议下,提出了适应网络规模扩展性的分布式算法,并利用网络仿真工具NS 2(networksimulator2. 27版)实现仿真。该算法在无线传感器网络MAC层实现了网络规模的可扩展性,增强了网络自适应网络规模变化的能力。     

Investigating the fault tolerance of neural networks

Particular levels of partial fault tolerance (PFT) in feedforward artificial neural networks of a given size can be obtained by redundancy (replicating a smaller normally trained network), by design (training specifically to increase PFT), and by a combination of the two (replicating a smaller PFT-trained network). This letter investigates the method of achieving the highest PFT per network size (total number of units and connections) for classification problems. It concludes that for non-toy problems, there exists a normally trained network of optimal size that produces the smallest fully fault-tolerant network when replicated. In addition, it shows that for particular network sizes, the best level of PFT is achieved by training a network of that size for fault tolerance. The results and discussion demonstrate how the outcome depends on the levels of saturation of the network nodes when classifying data points. With simple training tasks, where the complexity of the problem and the size of the network are well within the ability of the training method, the hidden-layer nodes operate close to their saturation points, and classification is clean. Under such circumstances, replicating the smallest normally trained correct network yields the highest PFT for any given network size. For hard training tasks (difficult classification problems or network sizes close to the minimum), normal training obtains networks that do not operate close to their saturation points, and outputs are not as close to their targets. In this case, training a larger network for fault tolerance yields better PFT than replicating a smaller, normally trained network. However, since fault-tolerant training on its own produces networks that operate closer to their linear areas than normal training, replicating normally trained networks ultimately leads to better PFT than replicating fault-tolerant networks of the same initial size.     

基于随机行走的自适应规模估计方法

在大规模动态网络环境中,节点需要系统规模信息以支持高效可靠的数据分发,针对该问题,提出一种基于随机行走的自适应规模估计方法——RASE。采用反熵聚集方法,在系统规模变化时发起随机行走以增强对动态环境的自适应性。结果证明,与已有方法相比,RASE方法在节点频繁波动的网络环境中仍然能准确快速地提供规模估计信息。     

一种基于语义吸引的节点规模估计方法

节点规模是各种分布式应用的基础信息,节点波动的大规模网络环境要求节点规模估计方法具有较高的估计精度和较强的鲁棒性,已有的节点规模估计方法多侧重于某个方面的优化而未能充分权衡计算精度和鲁棒性.提出一种基于语义吸引的节点规模估计方法——SEBSA(a network size estimation method based semantic attraction).SEBSA将每个节点标识所对应的实数区间上的哈希值作为语义信息,节点通过与哈希值临近的节点周期性地交换哈希空间上的邻居信息,以快速吸引与自己哈希值最近的一组节点,测量该组节点哈希值的平均间距以估计节点规模.理论分析和实验结果表明,相对于已有方法,SEBSA在节点频繁波动的网络环境中仍然能够快速地提供准确的节点规模信息.     

无线传感器网络自适应规模扩展的算法设计及实现*

无线传感器网络是由大量智能传感器节点组成的多跳无线网络.针对无线传感器网络的特点,在能量有效的传感器网络MAC协议下,提出并实现了网络自适应规模扩展的算法,并利用网络仿真工具NS-2实现仿真测试.该算法在无线传感器网络MAC层实现了网络规模的可扩展性,增强了网络自适应网络规模变化的能力.     

Examining network externalities and network structure for new product introduction

Much of the existing work on network externalities has assumed that network effects are dependent on network size. Very little consideration has been given to the view that marginal benefits from joining the network may not increase with network size, if consumer benefits come from direct interactions with neighbors in a local network. We use agent-based simulation to examine the effectiveness of the traditional network market strategy in the presence of a local network. Our results show that it is appropriate to establishing a sustainable network by leveraging the properties of local networks.     

Emotional disclosure on social networking sites: The role of network structure and psychological needs

We conducted three studies to understand how online emotional disclosure is influenced by social network structure on Facebook. Results showed that emotional disclosure was associated with both the density and size of users’ personal networks. Facebook users with denser networks disclosed more positive and negative emotions, and the relation between network density and emotional disclosure was mediated by stronger need for emotional expression. Facebook users with larger networks on Facebook disclosed more positive emotions, and the relation between network size and emotional disclosure was mediated by a stronger need for impression management. Our study extends past research by revealing the psychological mechanisms through which personal social network structure influences emotional disclosure. It suggests that social network size and density are associated with different psychological needs, which in turn lead to different patterns of emotional disclosure.     

ICN中基于节点权重的缓存大小分配方案研究

网内缓存功能是信息中心网络ICN（Information-Centric Networking）最重要的特性之一,大大减小了信息请求的响应时间和网内流量。合理地分配每个路由器的缓存空间大小,对网络性能有较大影响,也可以节约网络成本。为了使路由器的缓存大小配置合理,首先综合考虑路由器的度数权重、紧密度、网络的中心度、请求影响度等度量指标,定义了一个新的度量指标,称为节点权重;然后,提出一种基于节点权重的缓存大小分配方案,将网络所需的容量按比例分配给路由器。仿真结果表明,与均匀分配相比,路由器的缓存空间利用率至少提升了8%,命中率至少提高了6%;与基于请求影响度的分配方案相比,路由器的缓存空间利用率至少提升3%,命中率至少提高了3%。     

The Impact of Network Structure on the Stability of Greedy Protocols

A packet-switching network is stable if the number of packets in the network remains bounded at all times. A very natural question that arises in the context of stability properties of such networks is how network structure precisely affects these properties. In this work we embark on a systematic study of this question in the context of Adversarial Queueing Theory, which assumes that packets are adversarially injected into the network. We consider size, diameter, maximum vertex degree, minimum number of disjoint paths that cover all edges of the network and network subgraphs as crucial structural parameters of the network, and we present a comprehensive collection of structural results, in the form of stability and instability bounds on injection rate of the adversary for various greedy protocols: —Increasing the size of a network may result in dropping its instability bound. This is shown through a novel, yet simple and natural, combinatorial construction of a size-parameterized network on which certain compositions of greedy protocols are running. The convergence of the drop to 0.5 is found to be fast with and proportional to the increase in size. —Maintaining the size of a network small may already suffice to drop its instability bound to a substantially low value. This is shown through a construction of a FIFO network with size 22, which becomes unstable at rate 0.704. This represents the current state-of-the-art trade-off between network size and instability bound. —The diameter, maximum vertex degree and minimum number of edge-disjoint paths that cover a network may be used as control parameters for the stability bound of the network. This is shown through an improved analysis of the stability bound of any arbitrary FIFO network, which takes these parameters into account. —How much can network subgraphs that are forbidden for stability affect the instability bound? Through improved combinatorial constructions of networks and executions, we improve the state-of-the-art instability bound induced by certain known forbidden subgraphs on networks running a certain greedy protocol. —Our results shed more light and contribute significantly to a finer understanding of the impact of structural parameters on stability and instability properties of networks.     

双输入流深度反卷积的插值神经网络

在实际工作中深度学习方法通常不具备大量的训练样本，因此提出了双输入流深度反卷积生成神经网络的构架，依据给定的条件产生新的目标图像，从而扩充训练样本集。该神经网络的整体架构由双输入的卷积网络和一个反卷积网络输出构成，其中双输入卷积网络接收目标物体不同视角的两张图片并提取抽象特征，而反卷积网络则利用抽象特征和设定的参数产生新的插值目标图像。在ShapeNetCore数据集上的实验结果显示，在相同数量的训练样本空间中，与未扩展数据集的卷积网络相比，双输入流深度反卷积生成神经网络的识别率提高了20%左右。结果表明，双输入流深度反卷积生成神经网络无需输入目标物类别，可生成新参数条件下的目标图像，扩充训练样本空间，从而提高识别率，可用于少样本的目标物多角度识别。