CSC304 Lecture 9 Mechanism Design w/ Money: More examples of VCG, winner determination and truthful approximation CSC304 - Nisarg Shah 1
VCG Recap β’ π π€ = π β = argmax πβπ΅ Ο π π€ π (π) β Ο πβ π π€ π π β β’ π π π€ = max Ο πβ π π€ π π π β’ Procedure β’ Step 1: Choose the allocation to maximize social welfare β’ Step 2: Payment charged to each agent π is the externality that π imposes on others o [Max welfare of others | π absent] β [welfare of others | π present] Under π β CSC304 - Nisarg Shah 2
VCG Recap β’ Four properties β’ Maximize social welfare β’ Dominant strategy incentive compatibility (DSIC) β’ No payments to agents β’ Individual rationality (IR) β’ Vickrey auction satisfies the first two β’ VCG adds Clarkeβs pivot rule to satisfy all four CSC304 - Nisarg Shah 3
VCG Example β’ In the last lecture, we sawβ¦ β’ Additive valuations: agent has value π€ π π for each π , π€ π π = Ο πβπ π€ π π β’ Unit-demand valuations: Still have π€ π π for each π , π€ π π = max πβπ π€ π π o Goods are βsubstitutesβ β’ Another exampleβ¦ β’ Complementary goods: value of the whole exceeds the sum of values of its parts CSC304 - Nisarg Shah 4
VCG Example β’ A chair ( π ) and a table ( π’ ) π€ 1 π = 3 π€ 2 π’ = 4 β’ Allocation? β’ Payment? π€ 3 {π, π’} = 6 CSC304 - Nisarg Shah 5
VCG Example β’ A chair ( π ) and a table ( π’ ) π€ 1 π = 3 π€ 2 π’ = 4 β’ Allocation? β’ Payment? π€ 3 {π, π’} = 8 CSC304 - Nisarg Shah 6
VCG Example: Seller as Agent β’ Seller ( π ) wants to sell his car ( π ) to buyer ( πΆ ) β’ Seller has a value for his own car: π€ π π β’ Individual rationality for the seller mandates that seller must get revenue at least π€ π π β’ Idea: Add seller as another agent, and make his values part of the welfare calculations! CSC304 - Nisarg Shah 7
VCG Example: Seller as Agent π€ π π = 3 π€ πΆ π = 5 β’ What ifβ¦ β’ We give the car to buyer when π€ πΆ π > π€ π (π) and β’ Buyer pays seller π€ πΆ π : Not DSIC for buyer! β’ Buyer pays seller π€ π (π) : Not DSIC for seller! CSC304 - Nisarg Shah 8
VCG Example: Seller as Agent π€ π π = 3 π€ πΆ π = 5 β’ Allocation? Mechanism takes $3 from buyer, and gives β’ Buyer gets the car (welfare = 5 ) $5 to the seller! β’ Payment? β’ Need external subsidy β’ Buyer pays: 3 β 0 = 3 β’ Seller pays: 0 β 5 = β5 CSC304 - Nisarg Shah 9
Problems with VCG β’ Difficult to understand in complex settings β’ Need to reason about what allocation would maximize welfare if agent π was absent β’ Only cares about welfare, not revenue β’ Though, as we will see in a few lectures, gets pretty good revenue β’ With sellers and buyers, need external subsidy β’ Actually, cannot get individual rationality, DSIC, no subsidy, and constant approximation of welfare β’ Might be computationally difficult to implement β’ Computing welfare maximizing allocation may be hard CSC304 - Nisarg Shah 10
Single-Minded Bidders β’ Combinatorial auction for a set of π items π β’ Each agent π has β’ Value π€ π if receives a subset π π β π β’ Value 0 if doesnβt get a superset of π π β’ βSingle - mindedβ β’ Welfare-maximizing allocation: β’ Find a subset of players π with the highest total value such that their sets π π are disjoint CSC304 - Nisarg Shah 11
Single-Minded Bidders β’ Reduction to the Weighted Independent Set (WIS) problem in a graph β’ NP-hard to find the welfare-maximizing allocation β’ Note: not even thinking about computing payments yet β’ In fact, hard to approximately optimize welfare 1 2 βπ ) approximation (unless ππ β πππ ) o No O(π β’ Luckily, a simple greedy algorithm gives π -approximation (i.e., OPT/GREEDY β€ π ) CSC304 - Nisarg Shah 12
Greedy Algorithm β’ Input: (π€ π , π π ) for each agent π β’ Output: Agents with mutually independent π π β’ Greedy Algorithm: β’ Sort the agents. Go over them one-by-one. Accept each bid if no requested item is previously allocated. β’ Sort by what? β’ π€ 1 β₯ π€ 2 β₯ β― β₯ π€ π ? π -approximation π€ 1 π€ 2 π€ π π 1 β₯ π 2 β₯ β― π π ? π -approximation β’ π€ 1 π€ 2 π€ π π 1 β₯ π 2 β₯ β― π π ? π -approximation [Lehmann et al. 2011] β’ CSC304 - Nisarg Shah 13
Greedy Algorithm β’ (allocation rule, payments) truthful if and only if β’ Allocation is monotonic: If agent π wins with (π€ π , π π ) , it β² β₯ π€ π and π π β² β π π β² , π π β² ) where π€ π must win with (π€ π β’ Payments are critical prices: Agent π pays the least value (s)he could have reported and still won. |π π | β’ π π = π€ π β β π πβ β’ π β is the smallest index π such that π π β© π π β β and π π β© π π = β for all π < π, π β π β’ If agent π reports less than this value, agent π gets π π first, and π loses. CSC304 - Nisarg Shah 14
Moral β’ VCG can sometimes be too difficult to implement β’ May look into approximately maximizing welfare β’ Can set the payments right if the allocation rule is monotone β’ Need for approximation is due to computational considerations β’ Later in mechanism design without moneyβ¦ β’ Canβt use payments to ensure truthfulness β’ Will need to approximate welfare just to get truthfulness, even without computational considerations CSC304 - Nisarg Shah 15
Sponsored Search Auctions CSC304 - Nisarg Shah 16
Sponsored Search Auctions β’ Suppose the search engine receives a search query β’ π advertisement slots β’ β Clickthrough ratesβ : π 1 β₯ π 2 β₯ β― β₯ π π β₯ π π+1 = 0 β’ π advertisers (bidders) For convenience β’ Bidder π derives value π€ π *per click* β’ Final value to bidder π for receiving slot π = π€ π β π π β’ Without loss of generality, π€ 1 β₯ π€ 2 β₯ β― β₯ π€ π β’ Age-old question: β’ Who gets which slot, and how much should they pay? CSC304 - Nisarg Shah 17
Sponsored Search : VCG β’ VCG β’ Maximize welfare: π th bidder gets π th slot ( 1 β€ π β€ π ) β’ Payment of π th bidder? β’ Increase in social welfare to others if π abstains β’ Bidders π + 1 through π + 1 get βupgradedβ by one slot π+1 β’ Payment of bidder π = Ο π=π+1 π€ π β (π πβ1 β π π ) π πβ1 βπ π π+1 β’ Payment to bidder π βper clickβ = Ο π=π+1 π€ π β π π β’ Not very intuitiveβ¦ CSC304 - Nisarg Shah 18
Sponsored Search : VCG β’ What happens if all clickthrough rates are same? β’ π 1 = π 2 = β― = π π > π π+1 = 0 β’ Payment of bidder π per click π πβ1 βπ π π+1 β’ Ο π=π+1 π€ π β = π€ π+1 π π β’ Bidders 1 through π pay the value of bidder π + 1 β’ Familiar? VCG for π identical items CSC304 - Nisarg Shah 19
Sponsored Search : GSP β’ Generalized Second Price Auction (GSP) β’ For 1 β€ π β€ π β’ Bidder π gets slot π β’ Bidder π pays the bid of bidder π + 1 β’ A natural extension of the second price auction β’ We already saw that this is not truthful even with two identical slots β’ Highest bidder paying 2 nd highest bid β wants to lower bid to become 2 nd highest bidder and pay 3 rd highest bid CSC304 - Nisarg Shah 20
Sponsored Search : GSP β’ Truth-telling is not a Nash equilibrium ο β’ But there is a good Nash equilibrium that realizes the VCG outcome, i.e., maximizes welfare and generates as much revenue as VCG βΊ [Edelman et al. 2007] β’ Even the worst Nash equilibrium gives 1.282 - approximation to welfare ( πππ΅ β€ 1.282 ) and generates at least half the revenue of VCG [Caragiannis et al. 2011, Dutting et al. 2011, Lucier et al. 2012] CSC304 - Nisarg Shah 21
VCG vs GSP β’ VCG β’ Truthful in dominant strategy β more confidence that players will bid truthfully β’ Theoretical welfare/revenue guarantees will hold β’ Though players might still misreportβ¦ β’ Difficult to understand β’ GSP β’ Need to rely on players reaching a Nash equilibrium β’ Good welfare and revenue β’ Easy to understand CSC304 - Nisarg Shah 22
VCG vs GSP β’ Google uses GSP β’ Facebook used GSP, but switched to VCG β’ They argue that maximizing welfare has two benefits β’ Advertisers are happy β attract more advertisers β more long-term revenue β’ Users are happy (?!) β users use FB more β more slots to sell β more long-term revenue β’ No consensus CSC304 - Nisarg Shah 23
Sponsored Search Reality β’ Value is proportional to clickthrough rate β’ Could it be that users clicking on the 2 nd slot are more likely buyers than those clicking on the 1 st slot? β’ Ad engines also want to produce quality results β’ An advertiser having a high value for a slot does not necessarily mean his ad is appropriate for the slot β’ Theoretical analysis does not take into account market competition β’ Advertiser divide their budget among ad engines CSC304 - Nisarg Shah 24
Recommend
More recommend