Reinforcement Learning Based Quantization Strategy Optimal Assignment Algorithm for Mixed Precision

The quantization algorithm compresses the original network by reducing the numerical bit width of the model,which improves the computation speed. Because different layers have different redundancy and sensitivity to databit width. Reducing the data bit width will result in a loss of accuracy. Therefore, it is difficult to determinethe optimal bit width for different parts of the network with guaranteed accuracy. Mixed precision quantizationcan effectively reduce the amount of computation while keeping the model accuracy basically unchanged. In thispaper, a hardware-aware mixed precision quantization strategy optimal assignment algorithm adapted to low bitwidth is proposed, and reinforcement learning is used to automatically predict the mixed precision that meets theconstraints of hardware resources. In the state-space design, the standard deviation of weights is used to measurethe distribution difference of data, the execution speed feedback of simulated neural network accelerator inferenceis used as the environment to limit the action space of the agent, and the accuracy of the quantization model afterretraining is used as the reward function to guide the agent to carry out deep reinforcement learning training. Theexperimental results show that the proposed method obtains a suitable model layer-by-layer quantization strategyunder the condition that the computational resources are satisfied, and themodel accuracy is effectively improved.The proposed method has strong intelligence and certain universality and has strong application potential in thefield of mixed precision quantization and embedded neural network model deployment.

Strategies for Effective Comments on Students’ Written Assignments

Abstract:Comments on students’assignments are important for English course.The loopholes in teachers’working schemes-comments of their students’written assignments are the clues around which the paper spreads.In the process of teaching,some English teachers pay so much attention to the contents taught in class but ignore the comments on students’written assignments.Consequently,it is not applicable for students to correct their mistakes in time and they probably make the same mistakes repeatedly.

Two-stage stochastic programming with robust constraints for the logistics network post-disruption response strategy optimization

Logistics networks (LNs) are essential for the transportation and distribution of goods or services from suppliers to consumers. However, LNs with complex structures are more vulnerable to disruptions due to natural disasters and accidents. To address the LN post-disruption response strategy optimization problem, this study proposes a novel two-stage stochastic programming model with robust delivery time constraints. The proposed model jointly optimizes the new-line-opening and rerouting decisions in the face of uncertain transport demands and transportation times. To enhance the robustness of the response strategy obtained, the conditional value at risk (CVaR) criterion is utilized to reduce the operational risk, and robust constraints based on the scenario-based uncertainty sets are proposed to guarantee the delivery time requirement. An equivalent tractable mixed-integer linear programming reformulation is further derived by linearizing the CVaR objective function and dualizing the infinite number of robust constraints into finite ones. A case study based on the practical operations of the JD LN is conducted to validate the practical significance of the proposed model. A comparison with the rerouting strategy and two benchmark models demonstrates the superiority of the proposed model in terms of operational cost, delivery time, and loading rate.

Control strategies and experimental verifications of the electromagnetic mass damper system for structural vibration control

The electromagnetic mass damper (EMD) control system, as an innovative active control system to reduce structural vibration, offers many advantages over traditional active mass driver/damper (AMD) control systems. In this paper, studies of several EMD control strategies and bench-scale shaking table tests of a two-story model structure are described. First, two structural models corresponding to uncontrolled and Zeroed cases are developed, and parameters of these models are validated through sinusoidal sweep tests to provide a basis for establishing an accurate mathematical model for further studies. Then, a simplified control strategy for the EMD system based on the pole assignment control algorithm is proposed. Moreover, ideal pole locations are derived and validated through a series of shaking table tests. Finally, three benchmark earthquake ground motions and sinusoidal sweep waves are imposed onto the structure to investigate the effectiveness and feasibility of using this type of innovative active control system for structural vibration control. In addition, the robustness of the EMD system is examined. The test results show that the EMD system is an effective and robust system for the control of structural vibrations.

A Two-Stage Genetic Algorithm for the Truck-Door Assignment Problem with Limited Capacity Vehicles and Storage Area

In this paper,we investigate a kind of truck-door assignment problem in a crossdock.Multiple related realistic constraints such as the capacity of trucks,the capacity of temporary storage area,the time windows of trucks as well as the operational time inside the crossdock are considered and therefore,make the assignment problem strongly NP-hard and very challenging.A two-stage genetic algorithm is proposed to solve this problem and the experimental results show the superiority of this approach in comparison to CPLEX solver in terms of effectiveness and efficiency especially for the large-scale instances.That is,the decision making process can speed up and a near optimal truck-door assignment scheme can be obtained within a relatively short time by applying this two-stage approach in practice.

基于主动重心的青年高血压患者心肺运动时序数据增强

精准医疗的逐步兴起,如挖掘青年高血压患者的心肺运动时序数据,可以了解不同个体对有氧运动训练的响应性,有助于提高患者高血压管理计划的制定效率,更有效地实现有氧运动干预的治疗。开展该研究的瓶颈之一在于难以获取充足的样本数据。为了解决获取数据难度大、成本高等问题,利用加权动态时间规整重心平均算法来进行时间序列数据增强,重点针对重心选择和权重分配进行了研究。针对重心选择问题,首次引入了主动重心的概念,提出了代表性重心与多样性重心选择策略,改善了数据增强的效果。此外,针对现有权重分配策略的不足,提出了随机权重距离递减分配策略,避免了合成重复样本,进一步提升了模型的泛化能力。实验结果表明,在该研究背景下同时考虑重心选择与权重分配进行数据增强,可以进一步提升青年高血压患者有氧运动干预疗效预测的准确性。

核心素养导向的英语作业设计优化策略研究

针对国家“双减”政策和“新课标”,从实际情况出发,分析中小学英语作业的现状及存在的问题,结合问卷调查结果及相关理论,构建核心素养导向的中小学英语作业设计优化框架,探索“双减”背景下中小学英语作业设计优化策略,促进学生思维力、创造力、想象力的发展和核心素养的提升。

改进FCOS的复杂场景口罩佩戴检测算法

针对在复杂场景口罩佩戴检测中存在的多尺度、多角度和遮挡等问题,提出一种基于改进FCOS的复杂场景口罩佩戴检测算法。在算法的骨干网络中引入Res2Net的分组残差连接结构,提高网络对不同尺度口罩的特征提取能力,并在其中集成可变形卷积,拓展其对未知形状物体的建模能力;设计一种集成注意力机制的特征金字塔,为不同的特征通道赋予不同的权重,抑制无用的特征信息;根据目标口罩的相关统计特征自动地划分正负样本,提高不同尺度口罩的样本质量,并引入Generalized Focal Loss联合训练样本的分类分数和定位质量分数,提升算法性能。实验结果表明,在复杂场景下的口罩佩戴检测中,该改进算法的mAP相比于原始FCOS提高6.7个百分点,同时与一些主流的目标检测算法相比,该改进算法也具有更好的效果和鲁棒性。

基于分配策略优化算法的智能防空任务分配

针对分配策略最优算法在大规模场景中求解速度不足的问题,基于马尔可夫决策过程,将深度强化学习与其相结合,将大规模防空任务分配问题进行智能化求解。根据大规模防空作战特点,利用马尔可夫决策过程对智能体进行建模,构建数字战场仿真环境;设计防空任务分配智能体,通过近端策略优化算法,在数字战场仿真环境中进行训练。以大规模防空对抗任务为例,验证了该方法的可行性和优越性。

基于选择回溯的相控阵雷达-目标分配算法

针对多雷达多目标场景下现有雷达-目标分配算法中存在传感器切换率过高、算法耗时过长的问题,文中提出一种基于选择回溯的雷达-目标分配方法。该方法首先根据目标位置、雷达散射截面积大小等信息计算下一时刻雷达能够稳定跟踪目标的概率;选择其中概率较高,即失跟风险较低目标并直接沿用上一时刻分配结果以降低算法计算量;对概率较低,即失跟风险较高目标进行重新分配以提高跟踪稳定性;在分配过程中,当雷达资源出现饱和时采用“回溯”操作对已分配目标进行再分配以提高算法性能。仿真实验表明,与传统分配方法相比,所提方法能充分利用上一时刻分配结果,兼顾考虑了算法性能与复杂度,综合性能较优。

自适应多密度峰值子簇融合聚类算法

经典的密度峰值聚类算法在计算局部密度时过分依赖截断距离,在分配非中心点时易出现连锁效应,且人工选取聚类中心点的方式难以识别出密度不均匀簇的聚类中心。针对该问题,提出一种自适应多密度峰值子簇融合聚类算法。考虑样本的邻域信息,将自然邻居的思想引入密度峰值聚类中,实现了样本点局部密度的自适应计算;为发现稀疏密度簇的中心,提出一种簇中心自动选取策略用于确定初始子簇中心,并使用一种两阶段分配策略降低连锁效应发生的概率;提出一种基于K近邻相似度的度量准则,将相似度高的子簇进行融合,得到最终的聚类结果。在二维合成数据集以及UCI数据集上,相较经典的密度峰值聚类算法以及近年来对其改进的算法,该算法表现出更优异的性能。

改进Yolov5的无人机目标检测算法

无人机场景下航拍图像存在密度高、目标小、覆盖范围广等特性,使得现有的目标检测器容易出现错检漏检的现象,为了提高识别的精度,提出了一种改进Yolov5的目标检测模型。通过采用梯度流丰富的C2F模块增加模型的特征提取能力。引入上采样算子CARAFE(content-aware reassembly of features)增加感受野进行数据特征融合,提升特征金字塔网络性能。通过采用全局性动态标签分配策略,提高模型识别准确率。通过VisDrone2019数据集验证表明,改进后的模型平均精度mAP值达到65.3%,较传统模型提升了24.7个百分点,可以更加准确地完成航拍过程中针对目标的检测任务。

基于改进多目标简化群优化算法求解动态武器目标分配问题

针对动态武器目标分配算法中计算耗时较长,在动态多目标条件下应用智能算法时普遍存在求解效率低的问题,提出一种多目标简化群优化算法。该算法将简化群优化算法扩展到多目标领域,将全变量更新机制与谐波步长策略相结合,来平衡算法的探索和开发能力,将算法与其他算法用多目标优化测试函数进行多样性和收敛性对比测试分析。实验结果表明,该算法在求解多样性和收敛性上有一定优势,应用于动态武器目标分配问题求解,具有良好的计算效率和求解精度。

Optimal Dispatch for Flexible Uncertainty Sets in Multi-energy Systems: An IGDT Based Two-stage Decision Framework

The operational stability and economy of multi-energy systems(MES)are threatened by various uncertainties,such as variable renewable energy power,energy demands,and weather conditions.Most of the existing methods for the dispatch decisions of MES are based on the prescribed probability distribution or uncertainty sets of random variables,which have many disadvantages,such as potential infeasibility and over-conservatism.In this paper,we propose a novel dispatch model for MES that integrates dispatch decision making,uncertainty set selection,and operational cost control into a unified framework.First,the deterministic dispatch model of MES is introduced,in which the physical characteristics of district heating systems and buildings are fully considered.Then,a novel decision framework that combines the two-stage dispatch strategy and info-gap decision theory(IGDT)is proposed for MES,where the uncertainty set is flexible and can be optimized based on the operational cost budget.Finally,a revised algorithm,based on the column-and-constraint generation method,is proposed for the model.Case studies are performed on MES that includes a 33-bus distribution system and a heating network modified from a real 51-node network located in Jinlin Province,China.The results verify the effectiveness of the proposed method.

作业数据驱动的大规模因材施教——上海市闵行区“智慧教育示范区”创建实践

在“智慧教育示范区”创建和基础教育数字化转型的大背景下,为了打造优质均衡的教育体系,满足每个学生的发展需求,上海市闵行区在区域层面构建了区校两级作业系统,梳理了作业业务的构成要素,对作业业务各个环节进行量化。通过伴随式采集数据,形成区域作业数据驱动大规模因材施教的模型,并将知识状态追踪模型、学科核心能力评价模型、学习品质模型、学生综合画像模型等四个模型作为子模型。通过模型应用,开展基于知识状态的个性化辅导、学习策略和学习意志力的分层教学、学习投入的分层教学等实践,形成通过作业数据发掘差异的新路径。基于区域探索性数据分析,证明通过作业数据可以发掘学生的“材”,且发掘的结果既符合教师已有经验,又能弥补经验的不足。经实践证明,精准“施教”具有可行性,可为数据驱动的大规模因材施教提供参考。

煤矿井下群机器人高效任务分配算法

松散型合作群机器人系统在现阶段煤矿辅助机器人作业中具有广泛应用前景。但松散型合作群机器人系统的任务分配过程未向划分过程进行反馈,导致任务划分与分配过程高效性与合理性不足。针对该问题,提出一种基于改进型鲁宾斯坦协商策略的煤矿井下群机器人高效任务分配算法。根据群机器人系统任务划分与分配过程的多方博弈特点,将鲁宾斯坦协商策略由双方博弈向多方共同博弈方向延伸,提出多方协商博弈的“出价-讨价-还价”规则。从机器人个体执行能力与任务执行情况差异的角度出发,提出基于机器人个体单位时间任务完成量的折扣因子计算方法,以及基于各分配周期任务执行情况的任务完成状态反馈参数模型,以实现任务的动态划分与分配。通过3组机器人合作执行煤矿矿区的整体监测任务,对算法性能开展实验验证,结果表明:(1)算法3(采用改进型鲁宾斯坦协商策略)的任务划分与分配效率较算法1(将每组无人机数量与运行速度乘积的比例直接作为3组无人机任务划分与分配的标准)、算法2(使用多方共同协商的鲁宾斯坦协商策略,但不考虑任务完成状态反馈参数)分别提升了30.10%,18.29%。(2)基于算法3的3组无人机执行任务的平均最大时间差为42 s,较算法1、算法2分别优化了77.66%,65.29%,这是由于算法3通过引入任务完成状态反馈参数,及时对任务参与方的任务执行过程进行评估,将任务的分配和执行过程向任务的划分阶段进行反馈,使任务的划分与分配更加准确。

免疫进化策略及其在二次布局求解中的应用

利用一种免疫进化策略算法来求解二次布局问题。免疫进化策略保留了一般进化策略的随机全局搜索能力，又进一步借鉴生物免疫机制中抗体的多样度保持机制和基于浓度的调节机制，较好地维持了进化个体的多样性。求解结果相当完美。

基于精英改选机制的粒子群算法的空战纳什均衡策略逼近

针对无人机协同攻击的动态多策略性,应用纳什均衡概念,考虑联合生存概率和武器消耗等因素,融合双方价值函数计算和双矩阵对策纳什均衡点的求解方法,建立一种多战斗步空战动态目标分配优化模型.提出基于精英改选机制的粒子群(elite re-election particle swarm optimization,ERPSO)算法,在群体极值引导能力不足时,通过对其克隆、变异和重新初始化等操作增加个体的多样性,保留传统粒子群算法结构简单、快速收敛等优点,改善算法易于陷入局部极小的问题.将ERPSO算法应用于目标分配模型求解纳什均衡点,获取更为精确的双方混合策略,确保实时性和准确性,验证了模型和方法的有效性.

足球机器人多智能体协作策略

文章提出了足球机器人多智能体系统协作策略,它由三部分组成:角色执行器、角色分配器和避撞任务处理器。机器人角色分为进攻、防守和守门。角色任务处理器决定每一角色运动到何位置。角色分配器在每个采样周期中,根据视觉信息,动态地改变每一机器人的角色。避撞任务处理器预测两个机器人相撞的可能性,并改变机器人目标位置以避免碰撞。角色任务执行器执行由4阶bezier曲线产生运动轨迹。带有PID控制的足球机器人跟踪确定的bezier曲线。该文提出的策略已成功应用到GDUT_TAIYI队。

基于Shapley值法的共同配送利益分配研究

分析了由Shapley提出的用于解决n人合作对策问题的Shapley值法,给出了此方法分配共同配送伙伴利益的过程;并用此方法解决了共同配送企业之间的利益分配问题。根据Shapley值法的不足提出了修正方案。从实例中可以看出,修正后的结果更具有真实性和实际意义。