讲稿里复合误差那一段是本页最该看的，因为幻灯上只有两个名词和一张示意图：

「…if it's more likely to make a mistake, it might actually drift away from the distribution even more. And at that point, maybe here it's actually very far away, and it makes maybe an even bigger mistake. So this is often referred to as compounding errors. Because they can start drifting off the distribution when you make mistakes, those mistakes can then compound as you roll out your policy.」

然后她把它和监督学习的根本差别接上：

「The distribution of states that are visited by your policy is actually different from the distribution of states that's visited by the expert, and this means that it differs fundamentally from super[vised learning]」

「differs fundamentally」—— 这不是程度问题，是类别问题。

而紧接着那一句把上下半讲缝在了一起：「this is one of the reasons why fitting the distribution really well is very important. Because if you don't fit it well, and you learn only the mean, for example, then you're even more likely to have these compounding errors. Even if you do fit it well, you might still make mistakes and might drift off the distribution.」

注意最后半句：拟合得好也只是推迟问题，不解决问题。这就是为什么需要下两页的办法。

CS224R 第二讲 · 模仿学习：行为克隆与 DAgger