This paper introduces a new exact and smooth penalty function to tackle constrained min-max problems. By using this new penalty function and adding just one extra variable, a constrained rain-max problem is transforme...This paper introduces a new exact and smooth penalty function to tackle constrained min-max problems. By using this new penalty function and adding just one extra variable, a constrained rain-max problem is transformed into an unconstrained optimization one. It is proved that, under certain reasonable assumptions and when the penalty parameter is sufficiently large, the minimizer of this unconstrained optimization problem is equivalent to the minimizer of the original constrained one. Numerical results demonstrate that this penalty function method is an effective and promising approach for solving constrained finite min-max problems.展开更多
We investigate the superconvergence properties of the constrained quadratic elliptic optimal control problem which is solved by using rectangular mixed finite element methods.We use the lowest order Raviart-Thomas mix...We investigate the superconvergence properties of the constrained quadratic elliptic optimal control problem which is solved by using rectangular mixed finite element methods.We use the lowest order Raviart-Thomas mixed finite element spaces to approximate the state and co-state variables and use piecewise constant functions to approximate the control variable.We obtain the superconvergence of O(h^(1+s))(0<s≤1)for the control variable.Finally,we present two numerical examples to confirm our superconvergence results.展开更多
近几年卷积神经网络作为深度学习最重要的技术,在图像分类、物体检测、语音识别等领域均有所建树。在此期间,由多层卷积神经网络组成的深度神经网络横空出世,在各种任务准确性方面具有显著提升。然而,神经网络的权重往往被限定在单精度...近几年卷积神经网络作为深度学习最重要的技术,在图像分类、物体检测、语音识别等领域均有所建树。在此期间,由多层卷积神经网络组成的深度神经网络横空出世,在各种任务准确性方面具有显著提升。然而,神经网络的权重往往被限定在单精度类型,使网络体积相较于特定硬件平台上的内存空间更大,且floating point 16、INT 8等单精度类型已无法满足现在一些模型推理的现实需求。为此,提出一种以子图为最小单位,通过判断相邻结点之间的融合关系,添加了丰富比特位的混合精度推理算法。首先,在原有单精度量化设计的搜索空间中增加floating point 16半精度的比特配置,使最终搜索空间变大,为寻找最优解提供更多机会。其次,使用子图融合的思想,通过整数线性规划将融合后的不同子图精度配置,根据模型大小、推理延迟和位宽操作数3个约束对计算图进行划分,使最后累积的扰动误差减少。最终,在ResNet系列网络上验证发现,所提模型精度相较于HAWQ V3的损失没超过1%的同时,相较于其他混合精度量化方法在推理速度方面得到了提升,在ResNet18网络中推理速度分别提升18.15%、19.21%,在ResNet50网络中推理速度分别提升13.15%、13.70%。展开更多
安全约束机组组合(Security-constrained Unit Commitment,SCUC)问题作为制定发电计划的核心环节,在电力系统优化调度等方面具有十分重要的意义。针对考虑故障态约束后SCUC问题规模庞大、难以求解的情况,提出了一种基于辅助优化问题的...安全约束机组组合(Security-constrained Unit Commitment,SCUC)问题作为制定发电计划的核心环节,在电力系统优化调度等方面具有十分重要的意义。针对考虑故障态约束后SCUC问题规模庞大、难以求解的情况,提出了一种基于辅助优化问题的故障态安全约束削减方法。首先引入与具体故障态安全约束相关的辅助优化问题,从而建立判别相应故障态安全约束是否冗余的充分必要条件。然后探究冗余故障态安全约束辨识过程的具体加速方法,包括松弛辅助优化问题方法,使用可行性判据进行故障态安全约束预分类方法,以及多线程并行计算方法。最后,在IEEE118测试系统上对所提方法的正确性和有效性进行了仿真验证。展开更多
基金supported by the Grant of the Academy of Mathematics and System Science of Chinese Academy of Sciences-The Hong Kong Polytechnic University Joint Research Institute (AMSS-PolyU)the Research Grands Council Grant of The Hong Kong Polytechnic University (No. 5365/09E)
文摘This paper introduces a new exact and smooth penalty function to tackle constrained min-max problems. By using this new penalty function and adding just one extra variable, a constrained rain-max problem is transformed into an unconstrained optimization one. It is proved that, under certain reasonable assumptions and when the penalty parameter is sufficiently large, the minimizer of this unconstrained optimization problem is equivalent to the minimizer of the original constrained one. Numerical results demonstrate that this penalty function method is an effective and promising approach for solving constrained finite min-max problems.
基金supported by Guangdong Province Universities and Colleges Pearl River Scholar Funded Scheme(2008)National Science Foundation of China 10971074+1 种基金the National Basic Research Program under the Grant 2005CB321703Hunan Provincial Innovation Foundation For Postgraduate CX2009B119.
文摘We investigate the superconvergence properties of the constrained quadratic elliptic optimal control problem which is solved by using rectangular mixed finite element methods.We use the lowest order Raviart-Thomas mixed finite element spaces to approximate the state and co-state variables and use piecewise constant functions to approximate the control variable.We obtain the superconvergence of O(h^(1+s))(0<s≤1)for the control variable.Finally,we present two numerical examples to confirm our superconvergence results.
文摘近几年卷积神经网络作为深度学习最重要的技术,在图像分类、物体检测、语音识别等领域均有所建树。在此期间,由多层卷积神经网络组成的深度神经网络横空出世,在各种任务准确性方面具有显著提升。然而,神经网络的权重往往被限定在单精度类型,使网络体积相较于特定硬件平台上的内存空间更大,且floating point 16、INT 8等单精度类型已无法满足现在一些模型推理的现实需求。为此,提出一种以子图为最小单位,通过判断相邻结点之间的融合关系,添加了丰富比特位的混合精度推理算法。首先,在原有单精度量化设计的搜索空间中增加floating point 16半精度的比特配置,使最终搜索空间变大,为寻找最优解提供更多机会。其次,使用子图融合的思想,通过整数线性规划将融合后的不同子图精度配置,根据模型大小、推理延迟和位宽操作数3个约束对计算图进行划分,使最后累积的扰动误差减少。最终,在ResNet系列网络上验证发现,所提模型精度相较于HAWQ V3的损失没超过1%的同时,相较于其他混合精度量化方法在推理速度方面得到了提升,在ResNet18网络中推理速度分别提升18.15%、19.21%,在ResNet50网络中推理速度分别提升13.15%、13.70%。
文摘安全约束机组组合(Security-constrained Unit Commitment,SCUC)问题作为制定发电计划的核心环节,在电力系统优化调度等方面具有十分重要的意义。针对考虑故障态约束后SCUC问题规模庞大、难以求解的情况,提出了一种基于辅助优化问题的故障态安全约束削减方法。首先引入与具体故障态安全约束相关的辅助优化问题,从而建立判别相应故障态安全约束是否冗余的充分必要条件。然后探究冗余故障态安全约束辨识过程的具体加速方法,包括松弛辅助优化问题方法,使用可行性判据进行故障态安全约束预分类方法,以及多线程并行计算方法。最后,在IEEE118测试系统上对所提方法的正确性和有效性进行了仿真验证。
