一种计算矩阵特征值特征向量的神经网络方法

当把Oja学习规则描述的连续型全反馈神经网络(Oja-N)用于求解矩阵特征值特征向量时,网络初始向量需位于单位超球面上,这给应用带来不便.由此,提出一种求解矩阵特征值特征向量的神经网络(1yNN)方法.在lyNN解析解基础上得到了以下结果:初始向量属于任意特征值对应特征向量张成的子空间,则网络平衡向量也将属于该空间;分析了lyNN收敛于矩阵最大特征值对应特征向量的初始向量取值条件;明确了lyNN收敛于矩阵不同特征值的特征子空间时,网络初始向量的最大取值空间;网络初始向量与已知特征向量垂直,则lyNN平衡解向量将垂直于该特征向量;证明了平衡解向量位于由非零初始向量确定的超球面上的结论.基于以上分析,设计了用lyNN求矩阵特征值特征向量的具体算法,实例演算验证了该算法的有效性.1yNN不出现有限溢,而基于Oja-N的方法在矩阵负定、初始向量位于单位超球面外时必出现有限溢,算法失效.与基于优化的方法相比,lyNN实现容易,计算量较小.     

神经网络BP学习算法动力学分析

研究神经网络BP学习算法与微分动力系统的关系.指出BP学习算法的迭代式与相 应的微分动力系统数值解Euler方法在一定条件下等价,且二者在解的渐近性方面是一致的. 给出了神经网络BP学习算法与相应的微分动力系统解的存在性、唯一性定理和微分动力系统 的零解稳定性定理.从理论上证明了神经网络的学习在一定条件下与微分动力系统的数值方法 所得的数值解在渐近意义下是等价的,从而借助于微分动力系统的数值方法可以解决神经网络 的学习问题.最后给出了用改进Euler方法训练BP网的例子.     

具有不等式路径约束的微分代数方程系统的动态优化

针对具有不等式路径约束的微分代数方程（Differential-algebraic equations，DAE）系统的动态优化问题，通常将DAE中的等式路径约束进行微分处理，或者将其转化为点约束或不等式约束进行求解.前者需要考虑初值条件的相容性或增加约束，在变量间耦合度较高的情况下这种转化求解方法是不可行的；后者将等式约束转化为其他类型的约束会增加约束条件，增加了求解难度.为了克服该缺点，本文提出了结合后向差分法对DAE直接处理来求解上述动态优化问题的方法.首先利用控制向量参数化方法将无限维的最优控制问题转化为有限维的最优控制问题，再利用分点离散法用有限个内点约束去代替原不等式路径约束，最后用序列二次规划（Sequential quadratic programming，SQP）法使得在有限步数的迭代下，得到满足用户指定的路径约束违反容忍度下的KKT（Karush Kuhn Tucker）最优点.理论上证明了该算法在有限步内收敛.最后将所提出的方法应用在具有不等式路径约束的微分代数方程系统中进行仿真，结果验证了该方法的有效性.     

Regression-based weight generation algorithm in neural network for solution of initial and boundary value problems

This paper introduces a new algorithm for solving ordinary differential equations (ODEs) with initial or boundary conditions. In our proposed method, the trial solution of differential equation has been used in the regression-based neural network (RBNN) model for single input and single output system. The artificial neural network (ANN) trial solution of ODE is written as sum of two terms, first one satisfies initial/boundary conditions and contains no adjustable parameters. The second part involves a RBNN model containing adjustable parameters. Network has been trained using the initial weights generated by the coefficients of regression fitting. We have used feed-forward neural network and error back propagation algorithm for minimizing error function. Proposed model has been tested for first, second and fourth-order ODEs. We also compare the results of proposed algorithm with the traditional ANN algorithm. The idea may very well be extended to other complicated differential equations.     

Sequential synthesis of the time-optimal control in real time

The paper considers the sequential synthesis method for time-optimal control of linear systems. The method is based on piecewise constant finite controls that ensure approximate solutions for time-optimal control problems. A sequence of finite controls is thereafter transformed into the optimal control. The appropriate computations are reduced to a sequence of linear algebraic equation problems and the integration of a matrix differential equation over the intervals of control switching points and final point change. It is proven that the sequence of finite controls converge to the optimal control. The sliding mode conditions are obtained, as well as the control structure modifications for the motions on switching manifolds. The initial approximations reducing considerably computational complexity are considered. The computational algorithm, together with the modelling and numerical results, are presented.     

A genetic algorithm-based search space splitting pattern and its application in hydraulic and coastal engineering problems

This article reports a search space splitting pattern that can be applied to genetic algorithms in order to ensure that the entire search space is investigated. Hence, by keeping the genetic algorithm simple, in a reasonable time and with a high degree of accuracy, the initial solutions can be improved toward the global optimum point. The simplicity of the presented method is an advantage that makes it useful for applied hydraulic and coastal engineering problems. The performance of the proposed method was evaluated by a benchmark optimization problem, Levy No. 5, and three hydraulic and coastal engineering problems: inverse problem of Manning’s equation, the equation of equilibrium beach profiles, and the settling velocity equation of natural sediment particles. The results indicated that the nonlinear complex problems can be solved by the proposed method with a high degree of accuracy. The proposed genetic algorithm-based search space splitting pattern can either be used exclusively or alternatively it can be combined with improved operators in the literature.

     

On the simulation of nonlinear bidimensional spiking neuron models

Bidimensional spiking models are garnering a lot of attention for their simplicity and their ability to reproduce various spiking patterns of cortical neurons and are used particularly for large network simulations. These models describe the dynamics of the membrane potential by a nonlinear differential equation that blows up in finite time, coupled to a second equation for adaptation. Spikes are emitted when the membrane potential blows up or reaches a cutoff θ. The precise simulation of the spike times and of the adaptation variable is critical, for it governs the spike pattern produced and is hard to compute accurately because of the exploding nature of the system at the spike times. We thoroughly study the precision of fixed time-step integration schemes for this type of model and demonstrate that these methods produce systematic errors that are unbounded, as the cutoff value is increased, in the evaluation of the two crucial quantities: the spike time and the value of the adaptation variable at this time. Precise evaluation of these quantities therefore involves very small time steps and long simulation times. In order to achieve a fixed absolute precision in a reasonable computational time, we propose here a new algorithm to simulate these systems based on a variable integration step method that either integrates the original ordinary differential equation or the equation of the orbits in the phase plane, and compare this algorithm with fixed time-step Euler scheme and other more accurate simulation algorithms.     

Artificial neural networks for solving ordinary and partialdifferential equations

We present a method to solve initial and boundary value problems using artificial neural networks. A trial solution of the differential equation is written as a sum of two parts. The first part satisfies the initial/boundary conditions and contains no adjustable parameters. The second part is constructed so as not to affect the initial/boundary conditions. This part involves a feedforward neural network containing adjustable parameters (the weights). Hence by construction the initial/boundary conditions are satisfied and the network is trained to satisfy the differential equation. The applicability of this approach ranges from single ordinary differential equations (ODE), to systems of coupled ODE and also to partial differential equations (PDE). In this article, we illustrate the method by solving a variety of model problems and present comparisons with solutions obtained using the Galerkin finite element method for several cases of partial differential equations. With the advent of neuroprocessors and digital signal processors the method becomes particularly interesting due to the expected essential gains in the execution speed.     

多自由度Vanderpol振子极限环计算

提出了一种确定二阶多自由度Vanderpol振子周期轨道及周期的并行算法。通过改变系统的时间尺度,将系统周期转化为系统的变量,进而通过多分裂波形松弛法将边值问题转化为初值问题进行求解;同时为了计算系统周期,设计了周期迭代算法。最后通过数值示例进行仿真,进行了强有力的论证。     

用线性多步法训练神经网络

研究用微分方程数值解法－－线性多步法替代神经网络的学习算法,指出在一定条件下神经网络的BP学习问题与求解一个相应的微分系统在渐近意义下是等价的,从而求解微分动力系统的数值解法也可用于神经网络的学习,给出了训练神经网络的Milne方法和BP－Milne结合算法以及Hamming方法和BP－Hamming结合算法,并以9点两类模式、随机模式识别和石油地质中沉积微相模式识别等3个问题为例进行了实验,实验结果表明利用微分动力系统的数值解法进行神经网络的学习是可行的。     

用神经网络计算矩阵特征值与特征向量

该文研究用神经网格求解一般实对称矩阵的全部特征向量的问题。详细讨论了网络的平均态度合的结构并建立了平衡态集合的构造定理。通过求解简单的一维微分方程求出了网络的解析表达式。这一表达式是由对称矩阵的特征值与特征向量表达的、因而非常清晰利用解的解析表达式分析了网络的解的全局渐近行为。提出了用一些单位向量作为网络初始值计算对称矩阵的全部特征值与特征向量的具体算法。     

Mean square numerical solution of random differential equations: Facts and possibilities

This paper deals with the construction of numerical solutions of random initial value differential problems. The random Euler method is presented and the conditions for the mean square convergence are established. Numerical examples show that random Euler method gives good results even if the sufficient convergence conditions are not satisfied.     

Computer aided symbolic solution of multi-point boundary value problems occurring in physics and engineering

The author has elsewhere published papers demonstrating applicability of computer algebraic and symbol manipulation in obtaining solutions to ordinary and partial differential equations by Piccard's method, steepest descent and various forms of the Newton-Kantorovich theorems and applying them to non-trivial problems in engineering, physics and, especially, celestial mechanics. In this paper, the Taylor series will be developed permitting expansion about any point and for any boundary conditions for any order derivative at arbitrary points, i.e., the general multi-point boundary valve problem (MPBVP) will be solved. The symbolic algorithm developed is written in PL/1-Formac and produces the Taylor series solution for any nonlinear differential equation in which the highest order derivative may be algebraically isolated. This program permits the continuation of this solution on intervals of the independent variable, in the manner of polynomial splines. This program permits symbolic solutions, e.g. in terms of a symbolic initial condition. However, such a solution requires enormous main storage. PL/1-Formac was used due to its general availability and its compatibility with relatively small main storage, even as small as 200 K bytes. Two parameters are available for attaining a given numerical accuracy; the order of the Taylor expansion and the number of continuation intervals into which the solution range is divided. Experiments show that high accuracy can be obtained by judiciously selecting these two parameters in order to counterbalance truncation error against numerical round-off error. Extensive additional documentation of this procedure has been performed on scores of problems occurring in the applied mathematics literature.     

连续时间 Hopfield网络模型数值实现分析

讨论使用Euler方法和梯形方法在数值求解连续时间的Hopfield网络模型时,离散时间步长的选择和迭代停止条件问题.利用凸函数的定义研究了能量函数下降的条件,根据凸函数的性质分析它的共轭函数减去二次函数之差仍为凸函数的条件.分析连续时间Hopfield网络模型的收敛性证明,提出了一个广义的连续时间Hopfield网络模型.对于常用的Euler方法和梯形方法数值求数值实现连续时间Hopfield网络,讨论了离散时间步长的选择.由于梯形方法为隐式方法,分析了它的迭代求算法的停止条件.根据连续时间Hopfield网络的特点,提出改进的迭代算法,并对其进行了分析.数值实验的结果表明,较大的离散时间步长不仅加速了数值实现,而且有利于提高优化性能.     

Iterative methods for model reduction by domain decomposition

It is proposed a method to reduce the computational effort to solve a partial differential equation on a given domain. The main idea is to split the domain of interest in two subdomains, and to use different approximation methods in each of the two subdomains. In particular, in one subdomain we discretize the governing equations by a canonical scheme, whereas in the other one we solve a reduced order model of the original problem. Different approaches to couple the low-order model to the usual discretization are presented. The effectiveness of these approaches is tested on numerical examples pertinent to non-linear model problems including Laplace equation with non-linear boundary conditions and compressible Euler equations.     

A two-point difference scheme for computing steady-state solutions to the conservative one-dimensional Euler equations

An implicit finite-difference method is presented for obtaining steady-state solutions to the time-dependent, conservative Euler equations for flows containing shocks. The method uses a two-point central-difference scheme for the flux derivatives with dissipation added at supersonic points via the retarded density concept. Application of the method to 1-dimensional nozzle flow equations for various combinations of subsonic and supersonic boundary conditions show the method to be very efficient. Residuals are typically reduced to machine zero in approximately 35 time steps for 50 mesh points. For 1-dimensional Euler calculations, it is shown that the scheme offers two advantages over the more widely-used three-point schemes. The first is in regard to application of boundary conditions, and the second relates to the fact that the two-point algorithm is well-conditioned for large time steps.     

Nonlinear vibration and dynamic instability analysis nanobeams under thermo-magneto-mechanical loads: a parametric excitation study

The nonlinear vibration behavior and dynamic instability of Euler–Bernoulli nanobeams under thermo-magneto-mechanical loads is the main objective of the present paper. Firstly, a short Euler–Bernoulli nanobeam is modeled and exposed to an external parametric excitation. Based on the nonlocal continuum theory and nonlinear von Karman beam theory, the nonlinear governing differential equation of motion is derived. Secondly, to transport the partial differential equation to the ordinary differential equation, Galerkin method is applied. Then, multiple scales method, as an analytical approach, is used to solve the equation. At the end, modulation equation of Euler–Bernoulli nanobeams is obtained. Then, to evaluate the dynamic instability of the system, trivial and nontrivial steady-state solutions are discussed. Emphasizing the effect of parametric excitation, for considering the instability regions, bifurcation points are studied and investigated. As a result, it can be observed that the damping coefficient plays an effective role as well as parametric excitation in stability and frequency response of the system.

     

基于祖冲之类方法的多体动力学方程保能量保约束积分

针对一类多体动力学问题导出的微分 代数方程,提出一种保能量、保约束的算法.该算法基于祖冲之类方法和欧拉中点保辛差分,利用祖冲之类方法保证在时间格点上精确满足约束方程,避免约束违约问题;并进一步证明该算法在时间格点上可以精确保能量.数值算例进一步验证该算法的可靠性.     

改进变步长算法在实时飞行仿真中的应用

飞行模拟机是民航训练飞行员的重要装备,飞行仿真系统是飞行模拟机的重要系统之一.飞行仿真要建立飞机的全量运动方程,利用数值算法解算运动方程达到仿真飞机飞行状况的目的.采用四元数法表示的飞机欧拉方程能够克服普通欧拉方程奇异性,但用定步长方法解算会产生较大累积误差,所以必须采用变步长方法.实时飞行仿真要求较高的实时性与逼真度,通过仿真试验分析,确定了采用一种改进变步长2阶Runge-Kutta法,该方法具有迭代次数少,解算精度高,实时性强的优点.利用该改进变步长算法,编写了实时飞行系统仿真软件,软件中使用相应算法解决了变步长算法选择步长与系统迭代定时时间不匹配的矛盾,实现了精确的实时仿真.     

Averaging on Manifolds by Embedding Algorithm

We will propose a new algorithm for finding critical points of cost functions defined on a differential manifold. We will lift the initial cost function to a manifold that can be embedded in a Riemannian manifold (Euclidean space) and will construct a vector field defined on the ambient space whose restriction to the embedded manifold is the gradient vector field of the lifted cost function. The advantage of this method is that it allows us to do computations in Cartesian coordinates instead of using local coordinates and covariant derivatives on the initial manifold. We will exemplify the algorithm in the case of SO(3) averaging problems and will rediscover a few well known results that appear in literature.