Nelder-Mead is effective at discovering local minima but is guaranteed only to converge to a local minimum under certain conditions. For this reason, as with the pattern search algorithm, we make heuristic improvements to the algorithm such as testing for convergence by simplex size and random initialization after first restart. Random initialization helps substantially in quickly discovering unique and important local minima but makes the optimization weakly deterministic. Listing 8-5 is a lightweight algorithm for Nelder-Mead optimization as outlined at the beginning of this section. Similarly to the pattern search algorithm, the Nelder-Mead optimizer is best utilized by setting K very high and stopping manually later.

对一个连续变量进行分箱，因此这个连续变量的作用就相当于分类变量。

除了为图形中的主要部分添加标签，我们还经常需要为单个观测或分组观测添加标签。

surface and level plots

通过报告我们可以知道，质量差的钻石具有更高的价格，最差的钻石颜色是J（微黄），最差的纯净度是I1（肉眼可见内含物）。质量差的钻石似乎价格更高，造成这一现象的原因是一个重要的混淆变量：钻石的重量(carat)。重量是确定钻石价格的单一因素中最重要的一个，而且质量差的钻石往往更重一些。通过报告我们还可以知道一些钻石有非常大的残差。