第八章 高层理解－－目标识别.ppt

因此使间隔最大等价于使 最小；而要求分类线对所有样本正确分类，就要求： 过两类样本中离分类面最近的点且平行于最优分类面的超平面H1和H2上的训练样本就是使上式等式成立的那些样本，称为支持矢量。 （1） 4、最优分类面的求解 在（1）式的约束下，求下面函数的最小值： 为此，可以定义Lagrange函数： 若 为最优解，则： 可以证明：支持向量的 不等于零，而其它样本的 均等于零。由此，可以得到最优分类函数： 上式求和实际上是对支持向量来进行的。 5、支持向量机 这些观念可以扩展到非线性可分的情况下。可以增加一个松弛项到式子中。这个式子是非线性可分训练样本到分类空隙的方差总和。这个总和是通过控制最小错分的参数来权衡的。第二个问题是怎样处理非线性分类边界。这可以通过用核函数影射训练样本到高维特性空间来完成。在高维空间里，分类边界是线性的，常用的核函数是径向基函数，定义如下 和 表示两个样本。 值是用户控制的参数用来决定高斯核的宽度。 支持向量机的基本思想：首先通过非线性变换将输入空间变换到一个高维空间，然后在这个空间中求取最优线性分类面，而这种非线性变换是通过定义适当的内积函数实现的。 8.8 Genetic Algorithms Genetic algorithms: use a natural evolution mechanisms to search for the maximum of an object function; Generation: 1 2 3 4 1 011 001 110 010 2 011 110 110 010 Three basic operation Reproduction: is responsible for the survival of the fittest and for the death of others based on a probabilistic treatment; it copies strings with highest fitness into the next generation of strings; Crossover: randomly choosing a position for the border of each pair of strings, and to produce new strings by swapping all characters between the beginning of the string pairs and the border position; Mutation: is randomly to change one charater of some string of the population from time to time; algorithm: Create a starting population of code strings, and fine the value of their objective function; Probabilistically reproduce high fitness string in the new population , remove poor fitness strings; Construct new strings combining reproduced code strings from the previous population; From time to time, chang one character of some string at random; Order code strings of the current population according to the value of their objective function (fitness); If the maximum achieved string fitness does not increase over several step, stop. 8.10 Simulated annealing Basi notion: Definition: Simulated annealing searches for a minimum of an objective function that represents the goodness of some complex system. Imaging a sugar bowl freshly filled