算法博弈论 W36 Mechanism design basics
会在笔记中记录 Ιωάννης 的搞笑希腊口音的英语
帕拉漏 parallel
显仙秀里 essentially
神拿里欧 scenario
珠 true
多大儿 total
度 two
Auction
- single-item auction
- sponsored search auction
Single-item auctions
Scenario
a seller wants to sell an item
n potential buyers (interested in the item, ready to act strategically to get it)
Question: how do we design how the item should be sold?
Intermediate problem: what do the buyers want?
- buyer i has no-negative val of the item
- valuate is private, unknown to other potential buyers and the seller
- utility
- 0 if buyer doesn’t get the item
- if buyer gets the item and pays
Sealed bid Auction
- every buyer i submits bid to the seller
- seller decides who should get the item
- the seller decides the price of items
First-price auction
highest bidder gets/wins the item and pays equal to bid
for bidder: difficult to decide how to behave
for seller: difficult to predict what will happen
Second-price auction
highest bidder wins the item and pays the second highest bid
dominant strategy: if it maximizes utility for agent, no matter what other buyers decide to do
dominant strategy is the subset of best response
Theorem: in a second-price auction, every buyer i has as dominant strategy to submit private valuation as bid for the item
Proof?
Let B be the highest bid among other buyers
2 cases
- utility: 0. no incentive to bid since it will lead utility to negative.
- utility:
for bidder: easy to decide how to behave
for seller: easy to predict what will happen
Theorem: in second-price auction, truthful bidder is guaranteed non-negative utility
DSIC
DSIC aka dominant strategy incentive compatible: if truthful bidding is a dominant strategy for every potential buyer i and gives non-negative utility.
efficiency issues
- who gets the item? the highest bidder
- social welfare (the sum of utility of all buyers here equals winner’s valuation-winner’s pay) maximization
second-price auctions are ideal
- it’s DSIC
- it’s social welfare maximizing
- it can be implemented in time complexity polynomial (linear)
Sponsored search auctions
竞价拍卖
a simple model
k slots: 1, 2, …, k
advertisers who, for a particular query, wish to have their ad in one of the slots
every slot j has a click-through-rate (CTR) : how often users watch an ad at j-th slot
CTR of a slot does not depend on the ad hosted
CTRs:
advertiser i: has a private valuation for every click of its ad and values having the ad at slot j as
Goals:
- DSIC auction
- social welfare maximization: the tot val of the advertisers for clicks is as high as possible
- simplicity: low computational complexity
generalized second-price auction
advertisers: submit bids to Google
Google: interprets the bids as valuations per click and assigns the ad with j-th highest valuation per click to the j-th slot
Google: defines as price per click at slot j the (j+1)-th highest bid
when is welfare maximized?
when higher a times with higher v
NOT DSIC!!! truth-telling here is not dominant strategy (举个很简单的例子就可以证明)
Social welfare maximization? Not necessarily
exercise 2.1 prove that third-price auction is not DSIC
assume there are 3 bidder with true value 13, 11, 10.
the first guy 13 pay 10.
the second guy’s utility is 0, it incentives it to bid higher than 13, which utility will increase to (11-10)=1 i. e. the mechanism encourage telling lies.