## 2012年4月15日，星期日

### 互动式PAC

P很好地 A近似地 C直立的

#### 4条评论：

1. Ran into your blog. You might be interested in another application of a 弗里德曼风格 inequality (http://arxiv.org/abs/1002.4058). Here, the data can be adversarial (e.g. not necessarily iid), but the algorithm can act randomly and still do well. This doesn't get into game theory, however, because of how regret is defined -- this may be a limitation of the model.

1. 感谢您指出，我'把它放在我的阅读清单上。

Re: game theory, I may have been too hasty in declaring the future to be all game theory :) proposition 2 of http://arxiv.org/abs/1104.5070 indicates for OCO-style problems there is always a minimax strategy for the adversary which is oblivious (but nonstationary). So maybe statistical reasoning will be sufficient to master these problems.

2. 好吧，我显然需要花一会儿时间，但总的观点是显而易见的：您可以在对抗性环境中使用mar，以证明有关您自己的随机化结果。

万岁mar！