注意力机制量化剪枝优化方法

面向基于注意力机制模型的巨大计算和访存开销问题,研究量化和剪枝协同优化的模型压缩技术,提出针对注意力机制中查询、键、值、概率共四个激活值矩阵的对称线性定点量化方法。同时,提出概率矩阵剪枝方法和渐进式剪枝策略,有效降低剪枝精度损失。在不同数据集上的实验结果表明,针对典型基于注意力机制模型BERT,在较低或者没有精度损失的情况下该优化方法可达到4位或8位定点量化、0.93～0.98的稀疏度,大幅度降低模型计算量,为加速量化稀疏模型的推理奠定良好的基础。     

自主水下航行器的自适应反演变结构控制器设计

由于水下航行器是高度耦合、复杂的多输入多输出不确定性非线性系统。对水下航行器进行精确作业时的运动控制一直是其实用化过程中困扰人们的问题。将自适应反演控制技术和变结构控制策略相结合,提出了一种基于反演技术的自适应变结构控制器设计方法,实现AUV水平侧移和航向控制,并利用自主水下航行器模型进行了仿真研究。仿真结果表明,该控制方法对模型不确定性和不确定环境干扰表现出良好的鲁棒性和自适应性。     

一种新的雷达恒虚警(CFAR) 处理器

如何维持雷达检测系统恒定的虚警率是现代雷达技术和信号处理技术中重要的问题之一。为使雷达检测系统在复杂的杂波环境中获得恒定的虚警率,就需要有高性能的恒虚警(CFAR)处理器。本文提出的新型CFAR 处理器选择一个与被检测距离单元的值相接近的均值来作为CFAR 的门限,从而使这种处理器具有能够及时反映外部杂波环境变化的自适应性;它硬件实现简单,具有很好的抗多目标效应、抗杂波边缘效应、抗大目标“吃”小目标效应以及降低(CFAR)损失的性能。     

Sojourn times in G/M/1 fork‐join networks

We consider a processing network in which jobs arrive at a fork‐node according to a renewal process. Each job requires the completion of m tasks, which are instantaneously assigned by the fork‐node to m task‐processing nodes that operate like G/M/1 queueing stations. The job is completed when all of its m tasks are finished. The sojourn time (or response time) of a job in this G/M/1 fork‐join network is the total time it takes to complete the m tasks. Our main result is a closed‐form approximation of the sojourn‐time distribution of a job that arrives in equilibrium. This is obtained by the use of bounds, properties of D/M/1 and M/M/1 fork‐join networks, and exploratory simulations. Statistical tests show that our approximation distributions are good fits for the sojourn‐time distributions obtained from simulations. © 2008 Wiley Periodicals, Inc. Naval Research Logistics, 2008     

The stochastic single resource service‐provision problem

The service‐provision problem described in this paper comes from an application of distributed processing in telecommunications networks. The objective is to maximize a service provider's profit from offering computational‐based services to customers. The service provider has limited capacity and must choose which of a set of software applications he would like to offer. This can be done dynamically, taking into consideration that demand for the different services is uncertain. The problem is examined in the framework of stochastic integer programming. Approximations and complexity are examined for the case when demand is described by a discrete probability distribution. For the deterministic counterpart, a fully polynomial approximation scheme is known 2 . We show that introduction of stochasticity makes the problem strongly NP‐hard, implying that the existence of such a scheme for the stochastic problem is highly unlikely. For the general case a heuristic with a worst‐case performance ratio that increases in the number of scenarios is presented. Restricting the class of problem instances in a way that many reasonable practical problem instances satisfy allows for the derivation of a heuristic with a constant worst‐case performance ratio. Worst‐case performance analysis of approximation algorithms is classical in the field of combinatorial optimization, but in stochastic programming the authors are not aware of any previous results in this direction. © 2003 Wiley Periodicals, Inc. Naval Research Logistics, 2003     

并行时域自适应算法在电大目标瞬态散射中的应用

基于电磁场时域积分方程(TDIE)数值技术计算复杂目标的瞬态散射特性,其计算量和内存需求大,采用时域自适应算法(TDAIM)降低了TDIE的计算规模。在研究TDAIM并行算法的基础上,开发了基于.NET Remoting的电磁场分布式数值计算方案。数值结果表明,该方案显著提高了TDAIM的计算效率,为解决电大目标瞬态电磁散射问题提供了一条有效途径。     

准静态受遮蔽的Rician衰落信道中空时格码的性能上界

针对准静态受遮蔽的 Rician 衰落信道,在收端已知信道状态信息的条件下,文中分析了空时格码的性能.推导了码字差相关阵为对角阵的空时格码系统的成对差错概率的确切表达式,得到了相应空时格码在准静态受遮蔽的 Rician 衰落信道中的性能上界。仿真结果表明,在遮蔽不太严重的信道中,文中得到的 union 界与相应编码系统的实际性能有很好的近似度。     

车载UWB SAR干扰自适应抑制新方法

在分析车载超宽带合成孔径雷达(UWB SAR)探测系统中干扰类型及其产生机理的基础上,针对Tsaipei Wang等提出的批处理方法的局限性,提出一种基于迭代运算的图像域自适应干扰抑制方法.该方法利用基于韦布尔分布的恒虚警率(CFAR)检测器提高干扰信号估计的准确性;采用迭代方法在线估计和更新参数,实现干扰的自适应抑制.实测数据处理结果表明:该方法可有效抑制自信号干扰,且结构上适于实时处理,可满足车载UWB SAR探测系统对图像预处理的要求.     

态势估计中的二维模糊空间知识处理

空间推理是态势估计中的一个重要组成部分。针对海战态势估计的要求,建立了模糊确定性下二维空间点、区域,以及几种对态势估计中常用的空间关系的数学模型和用来确定空间关系的计算方法。     

高超声速滑翔飞行器变轨段自适应跟踪制导方法

针对高超声速滑翔飞行器变轨段大偏差条件下的标准轨迹跟踪问题,提出一种基于权值矩阵自适应修正的变轨段跟踪制导方法。分析了变轨段主要控制方式和标准轨迹特性;将简化的纵向运动方程在标准轨迹附近线性化;采用将误差项引进线性二次型性能指标加权矩阵的方式,设计了改进的权值自适应修正跟踪制导方法。CAV-H飞行器仿真分析表明,该方法能够实现高超声速滑翔飞行器变轨段高精度自适应跟踪制导,对初始及过程偏差具有良好的鲁棒性。