基于快速神经网络架构搜索的鲁棒图像水印网络算法

为解决深度学习在图像水印算法中计算量大且模型冗余的问题，提高图像水印算法在抵抗噪声、旋转和剪裁等攻击时的鲁棒性，提出基于快速神经网络架构搜索（neural architecture search，NAS）的鲁棒图像水印网络算法。通过多项式分布学习快速神经网络架构搜索算法，在预设的搜索空间中搜索最优网络结构，进行图像水印的高效嵌入与鲁棒提取。首先，将子网络中线性连接的全卷积层设置为独立的神经单元结构，并参数化表示结构单元内节点的连接，预先设定结构单元内每个神经元操作的搜索空间；其次，在完成一个批次的数据集训练后，依据神经元操作中的被采样次数和平均损失函数值动态更新概率；最后，重新训练搜索完成的网络。水印网络模型的参数量较原始网络模型缩减了92%以上，大大缩短了模型训练时间。由于搜索得到的网络结构更为紧凑，本文算法具有较高的时间性能和较好的实验效果，在隐藏图像时，对空域信息的依赖比原始网络更少。对改进前后的2个网络进行了大量鲁棒性实验，对比发现，本文算法在CIFAR-10数据集上对抵抗椒盐噪声和旋转、移除像素行（列）等攻击优势显著；在ImageNet数据集上对抵抗椒盐高斯噪声、旋转、中值滤波、高斯滤波、JPEG压缩、裁剪等攻击优势显著，特别是对随机移除行（列）和椒盐噪声有较强的鲁棒性。     

Diffusive search with spatially dependent resetting

We consider a stochastic search model with resetting for an unknown stationary target $a\in \mathbb{R}$ with known distribution $\mu$. The searcher begins at the origin and performs Brownian motion with diffusion constant $D$. The searcher is also armed with an exponential clock with spatially dependent rate $r=r\left(\cdot \right)$, so that if it has failed to locate the target by the time the clock rings, then its position is reset to the origin and it continues its search anew from there. Denote the position of the searcher at time $t$ by $X\left(t\right)$. Let $E_0^{\left(r\right)}$ denote expectations for the process $X\left(\cdot \right)$. The search ends at time $T_a=inf\{t\ge 0:X\left(t\right)=a\}$. The expected time of the search is then ${\int }_{\mathbb{R}}\left(E_0^{\left(r\right)}{T}_a\right)\mu \left(da\right)$. Ideally, one would like to minimize this over all resetting rates $r$. We obtain quantitative growth rates for $E_0^{\left(r\right)}{T}_a$ as a function of $a$ in terms of the asymptotic behavior of the rate function $r$, and also a rather precise dichotomy on the asymptotic behavior of the resetting function $r$ to determine whether $E_0^{\left(r\right)}{T}_a$ is finite or infinite. We show generically that if $r\left(x\right)$ is of the order ${\left|x\right|}^{2l}$, with $l>-1$, then $log{E}_0^{\left(r\right)}{T}_a$ is of the order ${\left|a\right|}^{l+1}$; in particular, the smaller the asymptotic size of $r$, the smaller the asymptotic growth rate of $E_0^{\left(r\right)}{T}_a$. The asymptotic growth rate of $E_0^{\left(r\right)}{T}_a$ continues to decrease when $r\left(x\right)\sim \frac{D\lambda }{x^2}$ with $\lambda >1$; now the growth rate of $E_0^{\left(r\right)}{T}_a$ is more or less of the order ${\left|a\right|}^{\frac{1+\sqrt{1+8\lambda }}{2}}$. Note that this exponent increases to $\infty$ when $\lambda$ increases to $\infty$ and decreases to 2 when $\lambda$ decreases to 1. However, if $\lambda =1$, then $E_0^{\left(r\right)}{T}_a=\infty$, for $a\ne 0$. Our results suggest that for many distributions $\mu$ supported on all of $\mathbb{R}$, a near optimal (or optimal) choice of resetting function $r$ in order to minimize ${\int }_{{\mathbb{R}}^d}\left(E_0^{\left(r\right)}{T}_a\right)\mu \left(da\right)$ will be one which decays quadratically as $\frac{D\lambda }{x^2}$ for some $\lambda >1$. We also give explicit, albeit rather complicated, variational formulas for ${inf}_{r\gneqq 0}{\int }_{{\mathbb{R}}^d}\left(E_0^{\left(r\right)}{T}_a\right)\mu \left(da\right)$. For distributions $\mu$ with compact support, one should set $r=\infty$ off of the support. We also discuss this case.     

The maximum number of columns in supersaturated designs with

Cheng and Tang [Biometrika, 88 (2001), pp. 1169–1174] derived an upper bound on the maximum number of columns that can be accommodated in a two‐symbol supersaturated design (SSD) for a given number of rows () and a maximum in absolute value correlation between any two columns (). In particular, they proved that for (mod ) and . However, the only known SSD satisfying this upper bound is when . By utilizing a computer search, we prove that for , and . These results are obtained by proving the nonexistence of certain resolvable incomplete blocks designs. The combinatorial properties of the RIBDs are used to reduce the search space. Our results improve the lower bound for SSDs with rows and columns, for , and . Finally, we show that a skew‐type Hadamard matrix of order can be used to construct an SSD with rows and columns that proves . Hence, we establish for and for all (mod ) such that . Our result also implies that when is a prime power and (mod ). We conjecture that for all and (mod ), where is the maximum number of equiangular lines in with pairwise angle .     

一种自适应人工蜂群算法求解U型顺序相依拆卸线平衡问题

产品拆卸过程中零部件之间会相互干扰影响任务作业时间,基于该情形构建了多目标U型SDDLBP优化模型,并提出一种自适应ABC算法。所提算法设计了自适应动态邻域搜索方法,以提高局部开发能力;采用了轮盘赌与锦标赛法结合的分段选择法,以有效评价并选择蜜源进行深度开发;建立了基于当前最优解的变异操作,以提高全局探索能力快速跳出局部最优。最后,通过算例测试和实例分析验证算法的高效性。     

面向远海岛礁群的双向物流网络规划

针对若干远海岛礁群系统封闭性的分布特点,将中心岛礁选址、各岛礁建设泊位数量和规模、仓库储存量、运输船型和数量以及航线设置、班期组织等作为规划内容,以整个运输网络系统总成本最低为原则,综合考虑了仓库存储理论,建立了兼顾“选址-库存-路径”的规划模型。结合所建模型特点,提出了双层并行搜索遗传算法,构建了以海上运输为基本交通方式的双向物流系统。最后,以我国南海为例进行建模并求解,采用不同算法并通过多组不同规模的算例进行演算比较,验证了所建模型和算法的有效性与合理性。     

A lightweight cell switching and traffic offloading scheme for energy optimization in ultra-dense heterogeneous networks

One of the major capacity boosters for 5G networks is the deployment of ultra-dense heterogeneous networks (UDHNs). However, this deployment results in a tremendous increase in the energy consumption of the network due to the large number of base stations (BSs) involved. In addition to enhanced capacity, 5G networks must also be energy efficient for it to be economically viable and environmentally friendly. Dynamic cell switching is a very common way of reducing the total energy consumption of the network, but most of the proposed methods are computationally demanding, which makes them unsuitable for application in ultra-dense network deployment with massive number of BSs. To tackle this problem, we propose a lightweight cell switching scheme also known as Threshold-based Hybrid cEll swItching Scheme (THESIS) for energy optimization in UDHNs. The developed approach combines the benefits of clustering and exhaustive search (ES) algorithm to produce a solution whose optimality is close to that of the ES (which is guaranteed to be optimal), but is computationally more efficient than ES and as such can be applied for cell switching in real networks even when their dimension is large. The performance evaluation shows that THESIS significantly reduces the energy consumption of the UDHN and can reduce the complexity of finding a near-optimal solution from exponential to polynomial complexity.     

Nonmonotone Inexact Newton Method for the Extended Linear Complementarity Problem

This paper presents a nonmonotone inexact Newton-type method for the extended linear complementarity problem (ELCP). We first reformulate the optimization system of the ELCP problem into a system of smoothed equations. Then we solve this system by a nonmonotone inexact Newton-type algorithm. The global convergence is obtained and numerical tests for some classes of ELCP include linear complementarity, horizontal linear complementarity, and generalized linear complementarity problems are also given to show the e?ciency of the proposed algorithm.