random effects 虽然说的是参数上的 randomness,但也可以理解为是给 error term 的协方差矩阵加了些 structure 条件:本来假设数据是 i.i.d. 的,error 的 covariance matrix 是 diagonal 的。现在因为有了分组等更多的信息,我们就不再假设 diagonal 了。因此 random effect model 也叫 variance components model。
另外也可以往 meta learning 的方向去理解,不同组其实对应了不同参数的模型(不同的 tasks),它们有一定的共性,我们就把共同的参数定义为 fixed effects,每个组自己独特的参数就是 random effects,这些组的 random effects 我们假设它们是服从一个分布的(they are samples from task distribution)。mixed model 就是把这些 tasks 放在一起解了,然后再进一步得到每个模型的参数估计。
Auto-Bidding in Large-Scale Auctions: Learning Decision-Making in Uncertain and Competitive Games
摘要和阿里巴巴游戏介绍,我发现就是做翻译,所以说本质没有差异。
2024.10.16
部分工作,不知是不是年纪大了,看这段的时候:又寒冷、又饥饿、又头晕!!
Challenge: This task is not a traditional constrained optimization problem because when bidding for each opportunity, it is impossible to know all future opportunities in advance due to the randomness of arriving patterns, making it hard to obtain a closed-form analytical solution.
Advertising is a competitive game where multiple agents bid simultaneously using their confidential bidding strategies. The game is also dynamic, with competitors continually adapting their strategies. Therefore, auto-bidding agents need to perceive the dynamic gaming environment, model the connection between bid and performance, and then carefully bid for each impression opportunity by taking into account the preceding bids and performance.
Additionally, participants need to consider how to adhere to the CPA constraint, given complex features such as the uncertainty of prediction and sparse data. In this competition, we’ll have 2 tracks. The first is the classic track, where any optimization method is allowed. The second is the AIGB track, encouraging the use of generative models for modelin