基于m×2正则化交叉验证的神经网络超参数调优方法

超参数调优是神经网络建模的关键问题。针对传统的超参数调优方法存在的问题,该文提出了一种基于m×2正则化交叉验证的超参数调优方法。目的是给出一种适用于复杂模型、大数据集背景下的计算开销较小且稳健的超参数调优方法。该方法的思想是从完整的数据集上选取少部分数据进行调优,避免模型在数据集较大时非常耗时的超参数调优难题;在m×2交叉验证的基础上设置正则化条件均衡训练集与验证集之间的分布差异,从而减少分布不一致带来的性能波动;使用信噪比作为调优的优化目标,从而可以综合考虑模型性能评价指标的均值和方差;并采用正交设计选择相关性较低的超参数组合以提高调优效率。以命名实体任务为例进行实验,在CoNLL 2003数据集上的实验结果显示,提出的调优方法能够选到和网格搜索性能上没有显著差异的超参数组合,且调优时间可显著降低约66%。     

外点不敏感的快速图像配准算法

提出一种图像配准算法,用于视频监控中将预存的背景帧与当前视频帧配准检测出运动对象.由于外点的存在,运动参数方程求解只适于使用稳健回归算法,但稳健回归需要较长的计算时间,因此提出一种称为屏蔽外点的图像配准算法,综合利用线性回归和稳健回归,求解运动参数方程.实验表明,该方法在计算速度和稳健性两方面均有较好的综合性能.     

一种基于统计分析的存储系统性能调优方法

计算机系统参数的合理配置能有效提升应用程序的性能。以NFS网络存储系统为例,提出了一种基于统计分析的存储系统性能调优方法,该方法分为关键系统参数识别和关键参数性能优化两个子阶段。阶段一采用方差分析(ANOVA)来建模系统参数的性能灵敏度,识别出对应用性能有显著影响的关键系统参数;然后在此基础上,阶段二采用响应面分析(RSM)来考察各关键参数对性能响应的影响,并综合前两个子阶段给出了性能调优算法,通过该算法找出系统的最优配置,从而最终达到性能调优的目的。最后,用实验评价了文中方法在Web, E-mail, Fileserver,Linux实用程序以及微基准测试等多种重要应用场景下的性能调优结果,实验结果证实了该调优方法的有效性和实用性。     

一种快速稳健的超分辨率图像复原算法

提供了一种快速稳健的正则化超分辨率图像复原算法。该算法从Roberts交叉梯度算子的角度构造正则项,从自适应性的角度确定正则参数。用1-范数度量数据逼近项和正则项。算法分析和实验结果表明该算法具有好的稳健性,并且计算速度较快。     

基于粒子群优化算法的模型超参调优方法应用研究

机器学习模型的效果很大程度上取决于模型的超参数调优,如何实现自动化的超参数优化是机器学习领域的重点研究方向之一。提出一种改进粒子群优化算法的惯性系数动态更新的方法,实现基于粒子群优化算法的模型超参数的调优方法,并以信用卡核心业务场景建模的自动调参过程予以实际验证。实验结果表明,改进后的自动超参数优化方法在通过多轮迭代逐步逼近最优解的参数择优方面更容易达到预期模型的效果目标,能够提升机器学习模型的预测精度和泛化能力。     

基于耦合模拟退火优化最小二乘支持向量机的车轮踏面磨耗量预测

针对最小二乘支持向量机(LS-SVM)超参数优化问题,提出采用改进耦合模拟退火(CSA)算法优化LSSVM超参数。首先,耦合模拟退火算法通过并行处理多个独立模拟退火(SA)寻优过程,提高LS-SVM模型超参数优化效率;然后通过调整接受温度控制耦合项超参数的接受概率方差,降低CSA算法初始设置对LS-SVM最优超参数确定过程稳健性的影响;最后结合既有线轮轨现场的实际检测数据,开展了基于改进耦合模拟退火优化的最小二乘支持向量机(CSA LS-SVM)回归模型性能对比实验。结果表明,CSA LS-SVM回归模型达到了模型精度、算法快速性、算法鲁棒性的有效折中,所建立的LS-SVM优化模型用于现场的车轮踏面磨耗量的预测是有效的。     

基于自适应池化的神经网络的服装图像识别

针对传统池化方式不能提取有效特征值的问题,提出根据池化域的尺寸、池化域内的元素值和网络的训练轮数调整池化结果的自适应池化方法,该算法依据插值原理与最大值池化模型构建函数,以特定函数值作为池化结果,然后利用交叉验证进行模型对比实验。同时提出了小样本调优法以解决目前依靠经验值在全部数据集上验证选取超参数效率较低的问题。在原始数据集上,按照分层抽样的规则抽取小样本,并基于小样本数据集对已编码的超参数组合循环训练并测试,通过对识别率最高的组合解码确定最优超参数。选用DeepFashion数据库进行相关实验,结果显示自适应池化模型的识别率达到83%左右,与最大值池化模型相比提高约2.5%。通过小样本选定超参数,并与随机组合超参数在原始数据集上进行对比实验,结果显示小样本调优法选择的超参数在经验值范围内最优,识别结果为86.98%,与随机组合超参数的平均识别率相比提高了约41.4%。自适应池化方法可以扩展到其他的神经网络中,小样本调优法对高效选取神经网络的超参数提供了依据。     

影像鉴定中的超分辨率重建稳定算法

提出一种稳健的超分辨率重建算法,并在影像鉴定中加以具体运用。通过对边缘图像进行建模,构建基于洛伦兹分布的最大后验概率重建算法,并自适应地调整洛伦兹宽度参数来逐步修正迭代结果。实验结果证明,该算法能获得理想的图像画面重建效果,具有较强的稳健性。     

基于时空优化深度神经网络的AQI等级预测

针对现有空气质量预测方法精度偏低、对噪声敏感等问题,提出一种基于堆栈降噪自编码（Stacked Denoising Auto-Encoders,SDAE）模型的空气质量等级预测方法。首先以武汉市历史空气质量和气象监测数据为研究对象,建立SDAE模型逐层学习原始数据的特征表达,并将最后一层特征与分类器连接完成预测模型的调优。同时改进多参数网格搜索法,选取了最优的超参数组合。然后在测试集上进行预测,并用预测值与实际值之间的平均绝对误差和均方误差等指标作为预测性能评价标准。通过与其他网络模型的实验对比,证明了SDAE模型对于空气质量等级具有较优的预测性能。最后从时间、空间、时空三个角度对该模型输入进行优化,实验结果表明基于空间优化的SDAE模型预测性能提升最为明显,能够得到比传统方法更加精确的预测结果。     

近邻感知的标签噪声过滤算法

基于k近邻的标签噪声过滤对近邻参数k的选取较敏感.针对此问题,文中提出近邻感知的标签噪声过滤算法,可有效解决二分类数据集的类内标签噪声的问题.算法分开考虑正类样本和负类样本,使分类问题中的标签噪声检测问题转化为两个单类别数据的离群点检测问题.首先通过近邻感知策略自动确定每个样本的个性化近邻参数,避免近邻参数敏感的问题.然后根据噪声因子将样本分为核心样本与非核心样本,并把非核心样本作为标签噪声候选集.最后结合候选样本的近邻标签信息,进行噪声的识别与过滤.实验表明,文中方法的噪声过滤效果和分类预测性能均较优.     

Model-reference robust tuning of 2DoF PI controllers for first- and second-order plus dead-time controlled processes

The aim of this paper is to present a robust tuning method for two-degree-of-freedom (2DoF) proportional integral (PI) controllers. This is based on the use of a model reference optimization procedure with servo and regulatory target closed-loop transfer functions for first- and second-order plus dead-time (FOPDT, SOPDT) controlled process models. The designer is allowed to deal with the performance/robustness trade-off of the closed-loop control system by specifying the desired robustness level by selecting a maximum sensitivity in the range from 1.4 to 2.0. In addition, a smooth servo/regulatory performance combination is obtained by forcing both closed-loop transfer functions to perform as closely as possible to non-oscillatory dynamic targets. A unified set of controller tuning equations is provided for FOPDT and SOPDT models with normalized dead-times from 0.1 to 2.0 that guarantees the achievement of the design robustness level. The robustness of the control system is analyzed as well as the robustness–fragility and performance–fragility of the optimized controllers. Comparative examples show the effectiveness of the proposed tuning method. The exact achievement of the control system robustness target for all the controlled process models considered (first- and second-order) is one of the distinctive characteristics of the proposed model reference robust tuning (MoReRT) method.     

Evaluation and simple tuning of PID controllers with high-frequency robustness

An extensive study of robust and optimal tuning of PID controllers for stable non-oscillating plants is presented. It is built on a set of well defined criteria related to output performance, stability margins and control activity. Different interesting properties of the closed loop systems are observed. A set of simple tuning rules is based on these observations. These rules are compared to a couple of well established tuning methods and are shown to give well competitive results, especially when simplicity, low control activity and high-frequency robustness are emphasized. Derivative action is shown to improve performance significantly compared to PI control, with equal stability margin and a moderate increase of control activity, for most plants, including those with significant time delay.     

不稳定时滞对象的最优鲁棒二自由度控制器解析设计

针对不稳定时滞对象提出一种二自由度控制器解析设计方法．通过内模控制规律进行控制器设计,改进内模控制的滤波器提高系统的鲁棒性及其性能．同时对系统的稳定性、鲁棒性及性能进行分析,提出了严格基于模型信息的鲁棒指标选择规则,保证系统获得良好的鲁棒/性能平衡．最后,将控制器的参数整定转化为一个约束优化问题,提出控制器解析整定方法,并通过仿真实例验证该方法的有效性．     

支持向量回归超参数的混沌文化优化选择方法

支持向量回归超参数的选择会影响模型性能,常用的梯度下降选择方法要求核函数或估计函数近似可微,且对迭代初值具有较强依赖性.对此,给出一种两阶段参数优化选择方法.第1阶段根据问题实际需求,确定超参数的变化区域;第2阶段在确定的参数变化范围内,采用自适应混沌文化算法,寻找具有最优性能的超参数组合.面向Mackey-Glass时间序列预测的仿真结果表明,该参数选择方法对函数结构不具有依赖性,所得超参数对应的SVR模型具有较好的泛化性能.     

基于阶跃跟踪响应的内模控制器性能评价、诊断与调节

对阶跃跟踪扰动下内模控制器(Internal Model Controller, IMC) 的性能评价、诊断与调节方法进行 了研究．采用绝对值误差积分(Integral of Absolute Error, IAE) 和灵敏度(Ms) 作为衡量系统跟踪性能和鲁棒性的 指标,提出一种对IMC 系统跟踪—鲁棒综合性能进行评价的方法．当系统性能不佳时,给出一种性能诊断方 法,对参数调节不当和模型失配两种影响控制器性能的原因进行诊断．针对模型失配的情况,给出一种利用 常规阶跃跟踪响应数据辨识模型参数的方法．     

Robust reliability method for non-fragile guaranteed cost control of parametric uncertain systems

The problem of non-fragile guaranteed cost control of uncertain systems is studied from a new point of view of reliability against uncertainties. An efficient robust reliability method for the analysis and design of non-fragile guaranteed cost controller of parametric uncertain systems is presented systematically. By the method, a robust reliability measure of an uncertain control system satisfying required robust performance can be obtained, and the robustness bounds of uncertain parameters such that the control cost of a system is guaranteed can be provided. The optimal non-fragile guaranteed cost controller obtained in the paper may possess optimal guaranteed cost performance satisfying the precondition that the system is robustly reliable with respect to uncertainties occurring in both the controlled plant and controller gain. The presented formulations are in the framework of linear matrix inequality and thus can be carried out conveniently. The presented method provides an essential basis for the tradeoff between reliability and control cost in controller design of uncertain systems. Two numerical examples are provided to demonstrate the efficiency and feasibility of the presented method. It is shown that the coordination and simultaneous realization of the system performance, control cost, and robust reliability in control design of uncertain systems are significant.     

Robust Controller Design And Pid Tuning For Multivariable Processes

In this paper, a robust controller design method is first formulated to deal with both performance and robust stability specifications for multivariable processes. The optimum problem is then dealt with using a loop‐shaping H_∞ approach, which gives a sub‐optimal solution. Then a PID approximation method is proposed to reduce a high‐order controller. The whole procedure involves selecting several parameters and the computation is simple, so it serves as a PID tuning method for multivariable processes. Examples show that the method is easy to use and the resulting PID settings have good time‐domain performance and robustness.     

Variable structure controllers for unstable processes

A variable structure control (VSC) method for unstable industrial processes is proposed. The proposed control method is able to provide a highly satisfactory system performance and to tackle with robustness issues of the processes in the presence of uncertainties. An ITAE-based numerical tuning algorithm for acquiring optimal control parameters, and a direct auto-tuning mechanism for the proposed controller are also provided. The performance of the proposed VSC method is illustrated on some unstable process models including a continuous stirred tank reactor (CSTR), in order to show its effectiveness, validity and feasibility.     

气动人工肌肉关节的迭代反馈整定控制及优化

气动人工肌肉关节（PMA）具有低成本、柔顺性和与生物肌肉类似的力学特性等优点，在医疗设备、仿生机器人等领域得到广泛应用.本文针对气动人工肌肉充气变形过程中存在的强非线性、时变性和控制参数难以确定的问题，提出了一种基于迭代反馈整定（IFT）算法的数据驱动优化控制策略，直接基于系统的输入输出数据，定义跟踪性能准则函数并采用Gauss-Newton估计算法实现对PID控制器参数的迭代整定，并通过引入辅助因子获取性能准则函数加权因子的最优值进一步加快了IFT算法的收敛速度.仿真结果表明，该方法相对于Ziegler-Nichols等传统PID参数整定方法可以有效提高控制系统的跟踪性能和鲁棒性.     

Robust IMC–PID tuning for cascade control systems with gain and phase margin specifications

In this article, an internal model control plus proportional-integral-derivative (IMC–PID) tuning procedure for cascade control systems is proposed based on the gain and phase margin specifications of the inner and outer loop. The internal model control parameters are adjusted according to the desired frequency response of each loop with a minimum interaction between the inner and outer PID controllers, obtaining a fine tuning and the desired gain and phase margins specifications due to an appropriate selection of the PID controller gains and constants. Given the design specifications for the inner and outer loop, this tuning procedure adjusts the IMC parameter of each controller independently, with no interference between the inner and outer loop obtaining a robust method for cascade controllers with better performance than sequential tuning or other frequency domain-based methods. This technique is accurate and simple, providing a convenient technique for the PID tuning of cascade control systems in different applications such as mechanical, electrical or chemical systems. The proposed tuning method explained in this article provides a flexible tuning procedure in comparison with other tuning procedures because each loop is tuned simultaneously without modifying the robustness characteristics of the inner and outer loop. Several experiments are shown to compare and validate the effectiveness of the proposed tuning procedure over other sequential or cascade tuning methods; some experiments under different conditions are done to test the performance of the proposed tuning technique. For these reasons, a robustness analysis based on sensitivity is shown in this article to analyze the disturbance rejection properties and the relations of the IMC parameters.