Full waveform inversion based on hybrid gradient

The low-wavenumber components in the gradient of full waveform inversion(FWI)play a vital role in the stability of the inversion.However,when FWI is implemented in some high frequencies and current models are not far away from the real velocity model,an excessive number of low-wavenumber components in the gradient will also reduce the convergence rate and inversion accuracy.To solve this problem,this paper firstly derives a formula of scattering angle weighted gradient in FWI,then proposes a hybrid gradient.The hybrid gradient combines the conventional gradient of FWI with the scattering angle weighted gradient in each inversion frequency band based on an empirical formula derived herein.Using weighted hybrid mode,we can retain some low-wavenumber components in the initial lowfrequency inversion to ensure the stability of the inversion,and use more high-wavenumber components in the high-frequency inversion to improve the convergence rate.The results of synthetic data experiment demonstrate that compared to the conventional FWI,the FWI based on the proposed hybrid gradient can effectively reduce the low-wavenumber components in the gradient under the premise of ensuring inversion stability.It also greatly enhances the convergence rate and inversion accuracy,especially in the deep part of the model.And the field marine seismic data experiment also illustrates that the FWI based on hybrid gradient(HGFWI)has good stability and adaptability.

Hybrid gradient vector fields for path-following guidance

Guidance path-planning and following are two core technologies used for controlling un-manned aerial vehicles(UAVs)in both military and civilian applications.However,only a few approaches treat both the technologies simultaneously.In this study,an innovative hybrid gradient vector fields for path-following guidance(HGVFs-PFG)algorithm is proposed to control fixed-wing UAVs to follow a generated guidance path and oriented target curves in three-dimensional space,which can be any combination of straight lines,arcs,and helixes as motion primitives.The algorithm aids the creation of vector fields(VFs)for these motion primitives as well as the design of an effective switching strategy to ensure that only one VF is activated at any time to ensure that the complex paths are followed completely.The strategies designed in earlier studies have flaws that prevent the UAV from following arcs that make its turning angle too large.The proposed switching strategy solves this problem by introducing the concept of the virtual way-points.Finally,the performance of the HGVFs-PFG algorithm is verified using a reducedorder autopilot and four representative simulation scenarios.The simulation considers the constraints of the aircraft,and its results indicate that the algorithm performs well in following both lateral and longitudinal control,particularly for curved paths.In general,the proposed technical method is practical and competitive.

HYBRID MULTI-OBJECTIVE GRADIENT ALGORITHM FOR INVERSE PLANNING OF IMRT

The intelligent optimization of a multi-objective evolutionary algorithm is combined with a gradient algorithm. The hybrid multi-objective gradient algorithm is framed by the real number. Test functions are used to analyze the efficiency of the algorithm. In the simulation case of the water phantom, the algorithm is applied to an inverse planning process of intensity modulated radiation treatment （IMRT）. The objective functions of planning target volume （PTV） and normal tissue （NT） are based on the average dose distribution. The obtained intensity profile shows that the hybrid multi-objective gradient algorithm saves the computational time and has good accuracy, thus meeting the requirements of practical applications.

Several Variants of the Primal-Dual Hybrid Gradient Algorithm with Applications

By reviewing the primal-dual hybrid gradient algorithm(PDHG)pro-posed by He,You and Yuan(SIAM J.Image Sci.,7(4)(2014),pp.2526–2537),in this paper we introduce four improved schemes for solving a class of saddle-point problems.Convergence properties of the proposed algorithms are ensured based on weak assumptions,where none of the objective functions are assumed to be strongly convex but the step-sizes in the primal-dual updates are more flexible than the pre-vious.By making use of variational analysis,the global convergence and sublinear convergence rate in the ergodic/nonergodic sense are established,and the numer-ical efficiency of our algorithms is verified by testing an image deblurring problem compared with several existing algorithms.

NEW HYBRID CONJUGATE GRADIENT METHOD AS A CONVEX COMBINATION OF LS AND FR METHODS

In this paper, we present a new hybrid conjugate gradient algorithm for unconstrained optimization. This method is a convex combination of Liu-Storey conjugate gradient method and Fletcher-Reeves conjugate gradient method. We also prove that the search direction of any hybrid conjugate gradient method, which is a convex combination of two conjugate gradient methods, satisfies the famous D-L conjugacy condition and in the same time accords with the Newton direction with the suitable condition. Furthermore, this property doesn't depend on any line search. Next, we also prove that, moduling the value of the parameter t,the Newton direction condition is equivalent to Dai-Liao conjugacy condition.The strong Wolfe line search conditions are used.The global convergence of this new method is proved.Numerical comparisons show that the present hybrid conjugate gradient algorithm is the efficient one.

APPLICATION OF HYBRID GENETIC ALGORITHM IN AEROELASTIC MULTIDISCIPLINARY DESIGN OPTIMIZATION OF LARGE AIRCRAFT

The genetic/gradient-based hybrid algorithm is introduced and used in the design studies of aeroelastic optimization of large aircraft wings to attain skin distribution,stiffness distribution and design sensitivity.The program of genetic algorithm is developed by the authors while the gradient-based algorithm borrows from the modified method for feasible direction in MSC/NASTRAN software.In the hybrid algorithm,the genetic algorithm is used to perform global search to avoid to fall into local optima,and then the excellent individuals of every generation optimized by the genetic algorithm are further fine-tuned by the modified method for feasible direction to attain the local optima and hence to get global optima.Moreover,the application effects of hybrid genetic algorithm in aeroelastic multidisciplinary design optimization of large aircraft wing are discussed,which satisfy multiple constraints of strength,displacement,aileron efficiency,and flutter speed.The application results show that the genetic/gradient-based hybrid algorithm is available for aeroelastic optimization of large aircraft wings in initial design phase as well as detailed design phase,and the optimization results are very consistent.Therefore,the design modifications can be decreased using the genetic/gradient-based hybrid algorithm.

A New Hybrid Vertical Coordinate Ocean Model and Its Application in the Simulation of the Changjiang Diluted Water

Based on the analysis of the advantages and disadvantages of some vertical coordinates applied in the calculation of the Changjiang diluted water （CDW）, a new hybrid vertical coordinate is designed, which uses σ coordinate for current and σ-z coordinate for salinity. To combine the current and salinity, the Eulerian-Lagrangian method is used for the salinity calculation, and the baroclinic pressure gradient （BPG） is calculated on the salinity sited layers. The new hybrid vertical coordinate is introduced to the widely used model of POM （Princeton Ocean Model） to make a new model of POM-σ-z. The BPG calculations of an ideal case show that POM-σ-z model brings smaller error than POM model does. The simulations of CDW also show that POM-σ-z model is better than POM model on simulating the salinity and its front.

HYBRID FINITE ANALYTIC SOLUTIONS OF SHALLOW WATER CIRCULATION

The hybrid finite analytic(HFA) method is a kind of numerical scheme in rectangular element. In order to simulate the shallow circulation in irregular bathymetry by HFA scheme, the model in sigma coordinate system was obtained. The model has been tested against three cases: 1) Wind induced circulation; 2) Density driven circulation and 3) Seiche oscillation. The results obtained in the present study compare well with those obtained from the corresponding analytical solutions under idealized for the above three cases. The hybrid finite analytic method and the circulation model in sigma coordinate system can be used calculate the flow and water quality in estuaries and coastal waters.

Study and Design of Hybrid Triplet Lens

This study presents a new method for designing algorithm for a triplet lens with one or two elements that are made of a gradient index medium (GRIN). This method is based upon considering a well-known designed triplet lens (Cooke triplet lens) as a target lens for designing of the Hybrid Triplet Lens (HTL). Our design was based upon keeping the total optical path length for the axial ray fixed for each case of design. The results showed that several designs for the HTL have the same total powers of the target lens. These designs depend on the variation of the GRIN element parameter values and the order of the GRIN element position in the system. These HTL designs have been evaluated by considering several optical merit functions, i.e., the root mean square (RMS) spot radius, wave front error and the spherical aberration. To achieve the optimal design, these functions are compared for the target lens and the HTL designs through a wide range of field angles.

Study on Electromagnetic Environment of AC-DC Hybrid Transmission Lines

With the expansion of electricity demand,transmission corridors are becoming scarce.AC and DC circuits running parallel to each other and sharing the same right-of-way or even the same tower become a possible option.Due to the existence of the adjacent line,space electromagnetic field and corona of another line may be changed.Different characteristics of two line types make the electromagnetic field of transmission corridors become more complex.Hybrid line is viewed as a whole.The calculation contains surface gradient,ground level electric field,radio interference and audible noise.Interaction between the two line types is considered.The calculation results show that the interaction is mainly concentrated in the inner corridor.In the role of DC electric field,AC electric field is no longer symmetrical and ground level electric field is significantly enhanced.Under the negative DC voltage,the positive corona of the waveform is significantly strengthened,and it is inhibited under the positive DC voltage.It is better to erect the positive DC line near AC line.

基于差分隐私的联邦学习方案

联邦学习的特点之一是进行训练的服务器并不直接接触数据,因此联邦学习本身就具有保护数据安全的特性。但是研究表明,联邦学习在本地数据训练和中心模型聚合等方面均存在隐私泄露的问题。差分隐私是一种加噪技术,通过加入适当噪声达到攻击者区分不出用户信息的目的。文中研究了一种基于本地和中心差分隐私的混合加噪算法(LCDP-FL),该算法能根据各个客户端不同权重、不同隐私需求,为这些客户端提供本地或混合差分隐私保护。而且我们证明该算法能够在尽可能减少计算开支的同时,为用户提供他们所需的隐私保障。在MNIST数据集和CIFAR-10数据集上对该算法进行了测试,并与本地差分隐私(LDP-FL)和中心差分隐私(CDP-FL)等算法进行对比,结果显示该混合算法在精确度、损失率和隐私安全方面均有改进,其算法性能最优。

Coherent beam combining of hybrid phase control in master oscillator-power amplifier configuration

A novel scalable architecture for coherent beam combining with hybrid phase control involving passive phasing and active phasing in master oscillator-power amplifier configuration is presented. Wide-linewidth mutually injected passive phasing fibre laser arrays serve as master oscillators for the power amplifiers, and the active phasing using stochastic parallel gradient descent algorithm is induced. Wide-linewidth seed laser can suppress the stimulated Brillouin scattering effectively and improve the output power of the fibre laser amplifier, while hybrid phase control provides a robust way for in-phase mode coherent beam combining simultaneously. Experiment is performed by active phasing fibre laser amplifiers with passive phasing fibre ring laser array seed lasers. Power encircled in the main-lobe increases1.57 times and long-exposure fringe contrast is obtained to be 78% when the system evolves from passive phasing to hybrid phasing.

小型无人有缆遥控水下机器人智能控制方法

针对深度确定性策略梯度(DDPG)算法应用于无人有缆遥控水下机器人(ROV)运动控制时存在的坏样本影响学习稳定性、缺少环境探索能力以及学习时间长难收敛等问题,从神经网络结构、噪声引入和融合监督学习3个方面对DDPG算法进行改进,并提出了基于混合神经网络结构和参数噪声的监督式DDPG算法。仿真结果表明,监督式DDPG算法比常规DDPG算法和传统比例-积分-微分(PID)算法更加有效。

IMPROVED GRADIENT METHOD FOR MONOTONE AND LIPSCHITZ CONTINUOUS MAPPINGS IN BANACH SPACES

Let C be a nonempty closed convex subset of a 2-uniformly convex and uniformly smooth Banach space E and {An}n∈N be a family of monotone and Lipschitz continuos mappings of C into E＊. In this article, we consider the improved gradient method by the hybrid method in mathematical programming [i0] for solving the variational inequality problem for {AN} and prove strong convergence theorems. And we get several results which improve the well-known results in a real 2-uniformly convex and uniformly smooth Banach space and a real Hilbert space.

TPMS点阵结构的密度梯度杂交优化设计

三周期极小曲面(triply periodic minimal surface,TPMS)点阵结构因其优异的综合性能受到中外学者的广泛关注。在点阵结构实际应用过程中,常常需要对其进行优化设计以兼顾轻量化与承载性能两方面的要求。目前,对TPMS点阵结构的优化设计主要集中于密度梯度层面,未综合考虑载荷方向对其力学性能的影响。为此,首先研究了TPMS点阵结构的各向异性特征。基于平均场均匀化方法求解了不同类型TPMS点阵结构的等效弹性矩阵,通过Matlab插值计算,绘制了其在三维空间范围内的杨氏模量图。发现不同类型的TPMS点阵结构呈现出不同的各向异性特征,其中W点阵结构在[100]等轴线方向上性能较强,在[111]等斜向对角方向上性能较弱,而P点阵结构则刚好相反。根据TPMS点阵结构的各向异性,同时考虑主应力方向以及相对密度分布对其性能的影响,提出了TPMS点阵结构的密度梯度杂交优化设计方法。以悬臂梁模型为基础,基于载荷边界条件对其进行拓扑优化设计,并将拓扑优化密度云映射为点阵结构的相对密度分布,从而实现密度梯度设计。根据TPMS点阵结构的各向异性特征以及单元主应力方向分别选择W和P点阵单胞填充悬臂梁,使主应力方向位于点阵结构性能较强的方向,避免点阵结构在性能薄弱的方向承受较大的应力。将不同类型的TPMS点阵单元合理分布后,利用激活函数将它们进行杂交连接,实现结构梯度设计。综合相对密度分布和单元结构分布,生成密度梯度杂交点阵结构。采用有限元仿真方法对比分析优化设计前后点阵结构的承载性能,结果表明密度梯度W和P点阵结构的刚度与对应的均质点阵结构相比都有明显提高,而由W和P两种点阵单胞组成的密度梯度杂交点阵结构刚度最大,比密度梯度W和P点阵结构分别提高4.63%和33.63%。该结果表明在密度优化的基础上,根据承载时单元主应力方向将不同类型的点阵结构进行合理分布以及混合杂交设计能够进一步提高结构的整体刚度。建立的TPMS点阵结构密度梯度杂交优化方法为其在轻量化设计等方面的应用提供了一定的指导。

基于TD3-PER的氢燃料电池混合动力汽车能量管理策略研究

为优化氢燃料电池混合动力汽车的燃料经济性及辅助动力电池性能,提出了一种基于优先经验采样的双延迟深度确定性策略梯度(TD3-PER)能量管理策略。采用双延迟深度确定性策略梯度(TD3)算法,在防止训练过优估计的同时实现了更精准的连续控制;同时结合优先经验采样(PER)算法,在获得更好优化性能的基础上加速了策略的训练。仿真结果表明:相较于深度确定性策略梯度(DDPG)算法,所提出的TD3-PER能量管理策略的百公里氢耗量降低了7.56%,平均功率波动降低了6.49%。

基于梯度搜索与进化机制的多目标混合算法

多目标进化算法(MOEA)因其良好的全局探索能力备受关注,但其在最优值附近的局部搜索能力却相对较弱,且对于具有大规模决策变量的优化问题,MOEA所需的种群数量与迭代次数都十分庞大,优化效率较低。基于梯度的优化算法能够很好地克服这些问题,但梯度搜索算法很难应用于多目标问题(MOPs)。在加权平均梯度的基础上引入随机权函数,发展多目标梯度算子,将其与基于参考点的第三代非支配排序遗传算法(NSGA-Ⅲ)结合,发展了多目标梯度优化算法(MOGBA)和多目标混合进化算法(HMOEA)。HMOEA在保留NSGA-Ⅲ良好的全局探索能力的同时,极大地增强了局部搜索能力。数值实验表明:HMOEA对于各种Pareto阵面都具有优秀的捕获能力,与典型的多目标算法相比效率提升了5~10倍。进一步将HMOEA应用于RAE2822翼型的多目标气动优化问题中,得到了理想的Pareto前沿,表明HMOEA是一种高效的优化算法,在气动优化设计中具有潜在的应用价值。

基于混合并行的分布式训练优化研究

大型神经网络训练是深度学习领域的一个热点话题,而分布式训练是基于多节点实现大型神经网络训练的最佳方法之一。分布式训练通常包含数据并行、层间并行和层内并行3种并行方法。然而现有的框架在层间并行时只能对模型进行手动切分,增加了模型设计的抽象复杂度,对此提出了节点约束关系搜索算法,实现了模型的自动切分。另外,在传统的数据并行和层间并行中,由于模型的复杂约束关系和通信操作的需要,计算和通信往往受到严格的序列化限制,为此引入了同步优化算法,实现了计算和通信的重叠,有效提高了整体训练的效率。实验对不同规模的GPT-2,AlexNet,VGG16和ResNet50模型进行训练,使用同步优化算法在6节点条件下可以将GPT2-XL,GPT2-LARGE和GPT2-MEDIUM模型的训练性能分别提升1.14倍、1.18倍和1.23倍,在1节点条件下将AlexNet,VGG16和ResNet50模型的训练性能分别提升1.31倍、1.14倍和1.03倍。实验结果表明,同步优化算法能够提升混合并行中的训练效率。

基于贝叶斯推断的高斯反卷积信号恢复

针对高斯卷积和白噪声干扰的降质信号,提出一种基于贝叶斯推断的信号恢复模型和数值算法。在模型中引入模型差异项来描述卷积核估计误差,并用高斯分布来描述其统计特性。基于贝叶斯推断,将后验分布解耦为多个推断问题,对信号、模型误差和卷积核参数分别利用原始对偶混合梯度方法、高斯共轭先验法、随机游走的Metropolis算法进行交替更新,有效恢复信号的同时对卷积核参数进行不确定性量化并避免误差传播。数值实验表明,本方法可以同时估计卷积核和恢复信号,性能优于传统的信号恢复方法。

混合动作空间下的多设备边缘计算卸载方法

为降低多设备多边缘服务器场景中设备层级的总成本,并解决现有深度强化学习(deep reinforcement learning,DRL)只支持单一动作空间的算法局限性,提出基于混合决策的多智能体深度确定性策略梯度方法(hybrid-based multi-agent deep determination policy gradient,H-MADDPG)。首先考虑物联网设备/服务器计算能力随负载的动态变化、时变的无线传输信道增益、能量收集的未知性、任务量不确定性多种复杂的环境条件,建立MEC系统模型;其次以一段连续时隙内综合时延、能耗的总成本最小作为优化目标建立问题模型;最后将问题以马尔科夫决策过程(Markov decision procession,MDP)的形式交付给H-MADDPG,在价值网络的辅助下训练并行的两个策略网络,为设备输出离散的服务器选择及连续的任务卸载率。实验结果表明,H-MADDPG方法具有良好的收敛性和稳定性,从计算任务是否密集、延迟是否敏感等不同角度进行观察,H-MADDPG系统整体回报优于Local、OffLoad和DDPG,在计算密集型的任务需求下也能保持更大的系统吞吐量。