Learn to Play Maximum Revenue Auction

摘要

Auctions for allocating resources and determining prices have become widely applied for services over the Internet, Cloud Computing, and Internet of Things in recent years. Very often, such auctions are conducted multiple times. They may be expected to gradually reveal participants' true value distributions, with which, it eventually would result in a possibility to fully apply the celebrated Myersons optimal auction to extract the maximum revenue, in comparison to all truthful protocols. There is however a subtlety in the above reasoning as we are facing a problem of exploration and exploitation, i.e., a task of learning the distribution and a task of applying the learned knowledge to revenue maximization. In this work, we make the first step effort to understand what economic settings would make this double task possible exactly or approximately. The question opens up greater challenges in the wider areas where auctions are conducted repeatedly with a possibility of improved revenue in the dynamic process, most interestingly in auctioning cloud resources.