第11章 约束问题的线性化方法.pptVIP

11.3.1可行方向算法 可行方向算法 例：可行方向算法 例：可行方向算法 例：可行方向算法 … 可行方向算法修正 ε微扰法 Topkis–Veinott方法 11.3.2单纯形方法推广 单纯形方法回顾 约束标准型： 基本解： 相对收益： 基本变量的 选取与替换： 新的可行基本解： 最优化准则： 所有非基本变量的相对收益大于或等于0 单纯形方法推广到线性约束问题：凸单纯形方法 相对收益： 最优化准则： 最优解可能不在顶点，非基本变量可能不为0 约束标准型： 基本解： 相对收益： 最优化准则： 线性搜索 凸单纯形算法 凸单纯形算法 11.3.3既约(Reduced)梯度方法 类似于无约束优化的梯度算法（Cauchy算法）。搜索方向d 为梯度的负方向 约化梯度为 ，即凸单纯形算法中非基本量的相对收益。可以证明，它实际上是在约束条件(m个)下的以非基本变量为独立变量(n-m)的梯度： 称为约化梯度，是在非基本变量子空间中的梯度。 11.3.3既约(Reduced)梯度方法 基本量的变化： 非基本量子空间 中的搜索方向： 保证x在定义域内： 确定搜索方向 11.3.3既约(Reduced)梯度方法 11.3.3既约(Reduced)梯度方法 约化梯度方法的加速 共轭梯度 准牛顿方法 Since ?2 and g12 are linear functions, they need not be replaced by piecewise linear functions. However, the other two, ?1 and g11, will need to be approximated. Suppose we choose to construct our approximation over the interval 0  x1  3, and suppose we use four equidistant grid points for this purpose. The data required for the piecewise approximation are tabulated below: * Variable x3 is a slack variable inserted to convert the constraint to an equality. * In fact, the simple definition of the active constraint set used in the basic form of the feasible direction algorithm can not only slow down the progress of the iterations but also lead to convergence to points that are not Kuhn–Tucker points (ref. 2, Chap. 13). This type of false convergence is known as jamming and, loosely speaking, occurs because the steps that the method takes become shorter and shorter as the direction vectors alternate between closely adjacent boundaries. The steps become shorter not because line search optima are reached but because a nearby constraint not considered in the direction generation subproblem is encountered. Various modifications have been presented that can guarantee the elimination of jamming. We will briefly consider two of these: the -perturbation method [1] and the method of Topkis and Veinott * Suppose we apply the same algorithm but replace the linear objective function coefficients c with the linearized form of a no