一种高效的函数可微的全局优化模拟退火算法

针对函数可微的全局优化问题,将最速下降法,Newton法和罚函数法引入模拟退火算法中,提出了一种高效的模拟退火算法.该算法可以求得可微函数优化问题的全局最优解,且具有计算量小,效率高的特点.利用罚函数将约束优化问题转化为无约束优化问题后,可以利用提出的算法进行求解.数值算例表明,提出的算法能够高效地求解无约束及带约束的函数可微的全局优化问题.     

线性支持向量机优化问题的一种光滑算法

线性支持向量机的无约束优化模型的目标函数不是一个二阶可微函数,因此不能应用一些快速牛顿算法来求解。提出了目标函数的一种光滑化技巧,从而得到了相应的光滑线性支持向量机模型,并给出了求解该光滑线性支持向量机模型的Newton-Armijo算法,该算法是全局收敛的和二次收敛的。     

广义支持向量机的多项式光滑函数法

为了求解广义支持向量机(GSVM)的优化问题,将带有不等式约束的原始优化问题转化为无约束优化问题,由于此无约束优化问题的目标函数不光滑,所以引入一族多项式光滑函数进行逼近,实验中可以根据不同的精度要求选择不同的逼近函数。用BFGS算法求解。实验结果表明,该算法和已有的GSVM的求解算法相比,更快地获得了更高的测试精度,更适合大规模数据集的训练。因此给出的GSVM的求解算法是有效的。     

一类非线性极小极大问题的粒子群-邻近点算法

针对每个分量函数都是凸函数的离散型非线性极小极大问题,提出一种全局收敛的粒子群-邻近点混合算法。该算法利用极大熵函数将极小极大问题转化为一个光滑函数的无约束凸优化问题;利用邻近点算法为外层算法,内层算法采用粒子群算法来优化此问题;数值结果表明,该算法数值稳定性好、收敛快,是求解此类非线性极小极大问题的一种有效算法。     

求解一类凸优化问题的区间迭代算法

研究了一类非线性带约束的凸优化问题的求解.利用Kuhn-Tucker条件将凸优化问题等价地转化为多变元非线性方程组的求解问题.基于区间算术的包含原理及改进的Krawczyk区间迭代算法,提出一个求解凸优化问题的区间算法.对于目标函数和约束函数可微的凸优化,所提算法具有全局寻优的特性.在数值实验方面,与遗传算法、模式搜索法、模拟退火法及数学软件内置的求解器进行了比较,结果表明所提算法就此类凸优化问题能找到较多且误差较小的全局最优点.     

基于最佳一致光滑逼近的孪生支持向量机研究

孪生支持向量机本质为两个二次规划问题,对于其目标函数中约束变量取正号不可微特性,提出一种基于最佳一致逼近的多项式光滑函数构建方法。分别以Bernstain多项式和Chebyshev多项式进行正号函数最佳一致有效光滑逼近。重点突出Chebyshev多项式的最佳一致逼近过程,使用Remez算法构造最佳一致Chebyshev多项式,讨论各阶Chebyshev多项式逼近状况。最后综合最佳一致逼近多项式和样本适应度构建目标优化函数,采用快速Newton-Armijo算法求解目标优化函数,基于UCI数据验证了方法的优越性。     

应用非单调线搜索求解一类互补问题

考虑一类含非Lipschtizian连续函数的非线性互补问题。引入plus函数的一类广义光滑函数,讨论其性质。应用所引入函数将互补问题重构为一系列光滑方程组,提出一个具有非单调线搜索的Newton算法求解重构的方程组以得到原问题的解。在很弱的条件下,该算法具有全局收敛性和局部二次收敛性。利用该算法求解一自由边界问题,其数值结果显示该算法是有效的。     

一个新的连续可微的单参数填充函数

填充函数法是求解非线性全局优化问题的有效方法。针对无约束优化问题,在目标函数及其梯度利普希兹连续的基础上,提出了一个新的连续可微的单参数填充函数,并研究了该填充函数的相关性质。最后,给出了一个填充函数算法,数值实验表明,该填充函数是有效的且算法是可行的。     

基于极大熵差分进化混合算法求解非线性方程组*

针对非线性方程组,给出了一种新的算法——极大熵差分进化混合算法。首先把非线性方程组转换为一个不可微优化问题;然后用一个称之为凝聚函数的光滑函数直接代替不可微的极大值函数,从而可把非线性方程组的求解转换为无约束优化问题,利用差分进化算法对其进行求解。计算结果表明,该算法在求解的准确性和有效性均优于其他算法。     

分段光滑的半监督支持向量分类机

为了解决半监督支持向量分类优化模型中的非凸非光滑问题,基于分段逼近的思想提出了一个分段函数,并以此逼近非凸非光滑的目标函数。给出的分段函数可以根据不同的精度要求选择不同的逼近参数,同时构造出基于上述分段函数的光滑半监督支持向量机模型。采用了LDS(Low Density Separation)算法求解模型,分析了其对对称铰链损失函数的逼进精度。理论分析和数值实验结果都证明分段光滑的半监督支持向量机的分类性能和效率优于以往提出的光滑模型。     

压缩感知优化问题的等价表示及其目标罚函数方法

首先定义了压缩感知优化问题的一个等价表示问题,证明了这个等价表示问题的最优解也是压缩感知优化问题的最优解。然后定义了它的一个具有2阶以上的光滑性的目标罚函数,给出了一个迭代求解算法,证明了所提算法的收敛性定理。定理表明,可以通过求解目标罚函数来获得压缩感知优化问题的近似最优解,该方法为研究和解决实际的压缩感知问题提供了一个新的工具。     

A study on the use of heuristics to solve a bilevel programming problem

The bilevel programming problem is characterized as an optimization problem that has another optimization problem in its constraints. The leader in the upper level and the follower in the lower level are hierarchically related where the leader's decisions affect both the follower's payoff function and allowable actions, and vice versa. One difficulty that arises in solving bilevel problems is that unless a solution is optimal for the lower level problem, it cannot be feasible for the overall problem. This suggests that approximate methods could not be used for solving the lower level problem, as they do not guarantee that the optimal solution is actually found. However, from the practical point of view near‐optimal solutions are often acceptable, especially when the lower level problem is too costly to be exactly solved thus rendering the use of exact methods impractical. In this paper, we study the impact of using an approximate method in the lower level problem, discussing how near‐optimal solutions on the lower level can affect the upper level objective function values. This study considers a bilevel production‐distribution planning problem that is solved by two intelligent heuristics hierarchically related: ant colony optimization for solving the upper level problem, and differential evolution method to solve the lower level problem.     

Stochastic optimal open-loop feedback control: Approximate solution of the Hamiltonian system

Considering a dynamic control system with random model parameters and using the stochastic Hamilton approach stochastic open-loop feedback controls can be determined by solving a two-point boundary value problem (BVP) that describes the optimal state and costate trajectory. In general an analytical solution of the BVP cannot be found. This paper presents two approaches for approximate solutions, each consisting of two independent approximation stages. One stage consists of an iteration process with linearized BVPs that will terminate when the optimal trajectories are represented. These linearized BVPs are then solved by either approximation fixed-point equations (first approach) or Taylor-Expansions in the underlying stochastic model parameters (second approach). This approximation results in a deterministic linear BVP, which can be handled by solving a matrix Riccati differential equation.     

Bad-scenario-set Robust Optimization Framework With Two Objectives for Uncertain Scheduling Systems

This paper proposes a robust optimization framework generally for scheduling systems subject to uncertain input data, which is described by discrete scenarios. The goal of robust optimization is to hedge against the risk of system performance degradation on a set of bad scenarios while maintaining an excellent expected system performance. The robustness is evaluated by a penalty function on the bad-scenario set. The bad-scenario set is identified for current solution by a threshold, which is restricted on a reasonable-value interval. The robust optimization framework is formulated by an optimization problem with two conflicting objectives. One objective is to minimize the reasonable value of threshold, and another is to minimize the measured penalty on the bad-scenario set. An approximate solution framework with two dependent stages is developed to surrogate the biobjective robust optimization problem. The approximation degree of the surrogate framework is analyzed. Finally, the proposed bad-scenario-set robust optimization framework is applied to a scenario job-shop scheduling system. An extensive computational experiment was conducted to demonstrate the effectiveness and the approximation degree of the framework. The computational results testified that the robust optimization framework can provide multiple selections of robust solutions for the decision maker. The robust scheduling framework studied in this paper can provide a unique paradigm for formulating and solving robust discrete optimization problems.      

基于均值改进控制策略的昂贵约束并行代理优化算法

针对具有黑箱特性的昂贵约束优化问题及工程中计算资源利用率不高问题,提出了新的基于均值改进控制策略的并行代理优化算法.该算法为了减少仿真建模计算负担,选取Kriging近似模型对目标函数和约束函数进行近似估计.在Kriging模型基础上,利用均值改进与新增试验样本间的不等关系构建具有距离特性的控制函数.算法的均值改进控制策略通过控制函数调整最大改进值,实现样本设计空间的多点填充.算法适用范围:1)计算成本主要来自于仿真估计而非优化;2)复杂的工程或商业软件内部无法修改的昂贵仿真问题.数值算例和仿真案例表明:该算法可有效获取近似最优解,减少仿真试验次数的同时弱化均值改进准则的贪婪特性.相比于其他多点填充策略,均值改进控制策略可有效提升算法计算效率.此外,算法获取优化问题近似最优解的稳定性和精度均具有一定优势.     

Difference approximation of optimal periodic control problems

The constrained optimal periodic control problem is approximated by a sequence of discretized problems in which the system of differential equations of the basic continuous problem is replaced by a system of one–step difference equations. Two kinds of approximate optimal controls are derived from the optimal solutions of discretized problems: the first in the form of a step function and the second in the form of a special trigonometric polynomial generated by a positive kernel. Sufficient conditions for approximate solutions to be weakly convergent to the optimal solution of the basic problem are given. Certain improvements in the difference approximation considered are discussed and potential applications given.     

A new solution to optimization-satisfaction problems by a penalty method

This paper is concerned with an optimization-satisfaction problem to determine an optimal solution such that a certain objective function is minimized, subject to satisfaction conditions against uncertainties of any disturbances or opponents' decisions. Such satisfaction conditions require that plural performance criteria are always less than specified values against any disturbances or opponents' decisions. Therefore, this problem is formulated as a minimization problem with the constraints which include max operations with respect to the disturbances or the opponents' decision variables. A new computational method is proposed in which a series of approximate problems transformed by applying a penalty function method to the max operations within the satisfaction conditions are solved by usual nonlinear programming. It is proved that a sequence of approximated solutions converges to a true optimal solution. The proposed algorithm may be useful for systems design under unknown parameters, process control under uncertainties, general approximation theory, and strategic weapons allocation problems.     

基于LS-SVM求解非线性常微分方程组的近似解

提出了一种基于最小二乘支持向量机(LS-SVM)的改进方法求解非线性常微分方程组初值问题的近似解.利用径向基核函数(RBF)可导的特点对LS-SVM模型进行改进,将含核函数导数形式的LS-SVM模型转化为优化问题进行求解.方法可在原始对偶集中获得近似解的最佳表示,所得近似解连续可微,且精度较高.给出数值算例,通过与真实解的对比验证了所提方法的准确性和有效性.     

Approximate optimization of systems with high-dimensional uncertainties and multiple reliability constraints

A novel approach is proposed to solve reliability-based optimization (RBO) problems where the uncertainty dimension can be large and where there may be many reliability constraints. The basic idea is to transform all reliability constraints in the target RBO problem into non-probabilistic ordinary ones by a pilot analysis. It will be shown that such a pilot analysis only requires a single run of the modified subset simulation (called the parallel subset simulation) regardless the number of the reliability constraints. Once the reliability constraints are approximated by the ordinary ones, the RBO problem can be solved as if it is an ordinary optimization problem. The resulting optimal solution should be approximately feasible, and the corresponding objective function value is minimized under the approximate constraints. Three numerical examples are investigated to verify the proposed novel approach. The results show that the approach may be capable of finding approximate solutions that are usually close to the actual solution of the target RBO problem.     

Decision-making based on approximate and smoothed Pareto curves

We consider bicriteria optimization problems and investigate the relationship between two standard approaches to solving them: (i) computing the Pareto curve and (ii) the so-called decision maker’s approach in which both criteria are combined into a single (usually nonlinear) objective function. Previous work by Papadimitriou and Yannakakis showed how to efficiently approximate the Pareto curve for problems like Shortest Path, Spanning Tree, and Perfect Matching. We wish to determine for which classes of combined objective functions the approximate Pareto curve also yields an approximate solution to the decision maker’s problem. We show that an FPTAS for the Pareto curve also gives an FPTAS for the decision-maker’s problem if the combined objective function is growth bounded like a quasi-polynomial function. If the objective function, however, shows exponential growth then the decision-maker’s problem is NP-hard to approximate within any polynomial factor. In order to bypass these limitations of approximate decision making, we turn our attention to Pareto curves in the probabilistic framework of smoothed analysis. We show that in a smoothed model, we can efficiently generate the (complete and exact) Pareto curve with a small failure probability if there exists an algorithm for generating the Pareto curve whose worst-case running time is pseudopolynomial. This way, we can solve the decision-maker’s problem w.r.t. any non-decreasing objective function for randomly perturbed instances of, e.g. Shortest Path, Spanning Tree, and Perfect Matching.