Download presentation
Presentation is loading. Please wait.
1
Information, Incentives, and Mechanism Design
Nick Gravin Course itcs.sufe.edu.cn/~nick Textbooks available at : Mechanism Design & Approximation: jasonhartline.com/MDnA/MDnA-ch1to8.pdf Game Theory, Alive: homes.cs.washington.edu/~karlin/GameTheoryBook.pdf
2
Course structure (Tentative)
Week 1, 2: Mechanism Design and Approximation Overview (Chapter 1, MDA) Topics: mechanism design, approximation, philosophy thereof, first-price auction, second-price auction, lottery, posted-pricings Week 2, 3, 4: Equilibrium (Chapter 2, MDA) Topics: Bayes-Nash equilibirum, dominant strategy equilibrium, single-dimensional agents, BNE characterization, revenue equivalence, uniqueness, revelation principle, incentive compatibility. Week 4, 5, 6, 7, 8: Optimal Mechanism Design (Chapter 3, MDA) Topics: single-dimensional mechanism design, surplus-optimal mechanism (VCG), revenue-optimal mechanism (Myerson), amortized analysis, virtual values, revenue curves, revenue linearity. Week 8, 9, 10, 11: Bayesian Approximation (Chapter 4, MDA) Topics: reserve pricing, posted pricing, prophet inequalities, correlation gap, monotone hazard rate distributions. Week 12, 13: Price of Anarchy in games (Chapter 8, GTA) Topics: selfish routing, existence of equilibrium, affine latencies, network formation games, market sharing games. Week 14, 15: Stable Matchings and allocations (Chapter 10, GTA) Topics: applications, Gale-Shapley algorithm, properties of Gale-Shapley algorithm, truthfulness considerations
3
Grading Homework (50%) Final exam (50%) Course report (50%) [bonus]
5 assignments (your score β best 4 out of 5). Work in pairs (same score for both) No copying, No late turn in. Final exam (50%) Based on problem solving recommendation: do homework. Course report (50%) [bonus] read a research paper/chapter in a book, write a report.
4
Economic & Computational Systems
Individuals take actions optimize their own (usually selfish) goals System combines actions to produce an outcome applies its laws and restrictions. Systems producing good outcomes: Spectrum auctions, Residency matching program, Auctions for online advertising, etc. Bad outcomes: Financial markets meltdowns, traffic jams, scandals in Olympics games.
5
Example I: Residency Matching
Medical student graduates in USA needs to be assigned to Hospitals for residency. Preferences: MGH > > Northwestern Memorial H. > > Pennsylvania hospital >
6
Example II: Adward auctions
Auction: buyers compete for better click probability Click-through rate/ success probability umbrella Ads Slot 1 Slot 2 Slot 3 0.5 0.2 0.1 30 billion $ of revenue for Google!
7
Example III: spectrum auctions
How to reallocate spectrum Run a double Auction Buys spectrum: From inefficient provides Meet budget constraint Sells spectrum: Complex preferences demand & value uncertainty Run an auction
8
Example IV: Congestion regulation
People choose: the fastest route between π π & π‘ π are selfish => suboptimal outcomes Government: introduces tolls to induce socially optimal outcomes Similar model for selfish routing of packages in a network.
9
Remarks These systems are complex Successful systems have :
individualsβ strategy spaces are large difficult to optimize over Successful systems have : simple rules individual strategies are simple This course: trade-off between optimality vs. simplicity simplicity => robustness, easy computations, practicality. Means: theory of approximation. compare the loss of performance of a simple practical mechanism to the complex optimal mechanism.
10
Let us isolate this issue:
Example IV: a model Network of routers π=({πππ’π‘πππ }, {πΏππππ }) π strategic users { π 1 , π‘ 1 β¦ π π , π‘ π } Routers/links: capacity constraints We need a mechanism for congestion control! Issues: dynamic demands complex network structure strategic user behaviour etc. Let us isolate this issue: Competition for 1 link
11
Single-item Allocation Problem
Definition [single-item problem] is given by 1 indivisible item available π strategic agents competing for the item each agent π has a value π£ π β₯0 for receiving the item. Objective: maximize social surplus (social welfare), i.e., max π£ π , π π’πβ π‘βππ‘ π β gets the item. Optimal solution: max π π£ π - allocate to the highest value Notation: π£ (1)
12
Mechanism I π 1 : You (say, value 100 RMB):
ask agents to report their values β agent π reports π π , select the agent π β with the highest report β π β = πππ max π π π allocate the item to π β You (say, value 100 RMB): How do you bid? You bid highest number you can think of! How to address: Bids are meaningless? Pay for overbidding?
13
No payments: Lottery Definition [Lottery]
select a uniformly random agent, and allocate the item to this agent. Theorem Lottery is an π approximation. Proof. The expected social welfare (SW) = 1 π πβ[π] π£ π The optimum = π£ (1) = max πβ[π] π£ π β€ πβ π π£ π β€πβ
(ππ). Bound is tight, when π£ 1 =1, π£ 2 = π£ 3 =β¦= π£ π =π.
14
Single-item Auction Def [Single-item auction] Solution to the allocation problem that solicits bids, picks a winner, determines payments. Def [First-price auction] ask agents to report their values (β agent π reports π π ), select π β with highest report ( β π β = πππ max π π π ), allocate the item to agent π β charge this winning agent her bid π π β . Used in government procurements Not easy to calculate your best strategy A natural example: Social welfare π£ π β = π£ π β β π π β + π π β agentβs utility sellerβs ,
15
How to play? Def [utility] agent π derives utility π’ π
π’ π = π£ π β π π , if she wins in the auction π’ π =0, if she does not get the item Assumption: agents want to maximize their utility if responding to others β play (pure) Nash Equilibrium. uncertainty about bids β maximize utility in expectation. Experiment: 20 people got π£ π βΌπ[0,4] usd randomly split into 10 pairs and played first-price auction winner of each auction would get π£ π β π π usd. Over-biding compared to equilibrium β (Assumption?)
16
Ascending-price Auction
Def [ascending-price auction] a.k.a. English auction. (gradually) raise an offered price up from zero, let agents to drop out if donβt want to win at the offered price, stop at the price π β : the second-to-last agent drops out, item β the last remaining agent, charge her π β . Agentβs strategy (value π£): drop at price π<π£ β certainly donβt get the item β bad idea. continue above π>π£ β if win, pay more than π£ β bad idea. β dominant strategy: drop at π=π£.
17
Ascending-price Auction
View at the system: every agent follows dominant strategy second-highest bidder π£ (2) will drop highest bidder gets the item, pays π£ (2) β utility π£ (1) β π£ (2) SW: π£ (1) = π£ (1) β π£ (2) + π£ (2) β optimal social surplus. Theorem The ascending-price auction maximizes the social surplus in dominant strategy equilibrium. Question: How to extend this to routing example? Issues: too slow to implement in practice ascending prices donβt find socially optimal routing scheme.
18
Second-Price Auction Def [Second-price auction] a.k.a. Vickrey auction
accept sealed bids, item β the agent π β with the highest bid, charge π β the second-highest bid. Strategically equivalent to the ascending price auction Theorem Truthful bidding is dominant strategy.
19
Truthfulness Theorem Truthful bidding is dominant strategy.
Proof Fix all values π£ 1 ,β¦, π£ π . Let π£ π = max πβ π π£ π . For agent i: Case 1. bid π π > π£ π β wins, pays π£ π β π’ π = π£ π β π£ π . Case 2. bid π π < π£ π β loses β π’ π =0. π max utility: if π£ π > π£ π wants case 1, if π£ π < π£ π wants case 2. Bidding π π = π£ π achieves that β dominant strategy. Corollary Second-price auction maximizes social surlplus. Observations: π£ π - infimum of bids π can make and still win; π£ π - agent πβs critical value = the price charged to π; π£ π - a function only of the other agentsβ reports.
20
General routing problem
Def [second-price routing mech.] a.k.a. Vickrey-Clarke-Groves solicit sealed bids, find the set of messages that can be routed simultaneously with the largest total value, charge the agents of each routed message their critical values. Theorem The second-price routing mechanism has truthful bidding as a dominant strategy. Corollary The mechanism maximizes the social surplus. Proof similar to 2nd-price auction
21
Can let only 1 message through
Example C Optimal value: 5$ (1) π΄βπ΅ & (4) πΆβπ· Can let only 1 message through A D What are the payments for senders of messages (1) and (4)? π£ 1 (π΄βπ΅)=1$ π£ 2 (π΄βπ·)=2$ π£ 3 π΅βπΆ =3$ π£ 4 (πΆβπ·)=4$ Critical value for (1)? B
22
Example C π£(π΄βπ΅,πΆβπ·)= 0$+4$ >π£ π΅βπΆ =3$ A D
What are the payments for senders of messages (1) and (4)? π£ 1 (π΄βπ΅)=0$ π£ 2 (π΄βπ·)=2$ π£ 3 π΅βπΆ =3$ π£ 4 (πΆβπ·)=4$ Critical value for (1)? B What if (1) bids 0, does she win? Yes. β π π΄βπ΅ =0$
23
Example C A D What are the payments for
senders of messages (1) and (4)? π£ 1 (π΄βπ΅)=1$ π£ 2 (π΄βπ·)=2$ π£ 3 π΅βπΆ =3$ π£ 4 (πΆβπ·)=4$ Critical value for (4)? B
24
Example C π£(π΄βπ΅,πΆβπ·)=1$+1$ <π£ π΅βπΆ =3$ A D What are the payments for
senders of messages (1) and (4)? π£ 1 (π΄βπ΅)=1$ π£ 2 (π΄βπ·)=2$ π£ 3 π΅βπΆ =3$ π£ 4 (πΆβπ·)=1$ Critical value for (4)? B What if (4) bids 1, does she win? No. β π πΆβπ· >1$
25
Example C π£(π΄βπ΅,πΆβπ·)=1$+2$ =π£ π΅βπΆ =3$ A D What are the payments for
senders of messages (1) and (4)? π£ 1 (π΄βπ΅)=1$ π£ 2 (π΄βπ·)=2$ π£ 3 π΅βπΆ =3$ π£ 4 (πΆβπ·)=2$ Critical value for (4)? B What if (4) bids 2, does she win? Tie. β π πΆβπ· =2$
26
VCG: collusion C A D What are the payments for
senders of messages (1) and (4)? π£ 1 (π΄βπ΅)=1$ π£ 2 (π΄βπ·)=2$ π£ 3 π΅βπΆ =3$ π£ 4 (πΆβπ·)=4$ B
27
VCG: collusion C π£(π΄βπ΅,πΆβπ·)=5$+5$ >π£ π΅βπΆ =3$ A D
What are the payments for senders of messages (1) and (4)? π£ 1 (π΄βπ΅)=5$ π£ 2 (π΄βπ·)=2$ π£ 3 π΅βπΆ =3$ π£ 4 (πΆβπ·)=5$ B What if (1) & (4) both bid 5? Win, π π΄βπ΅ = π πΆβπ· =0$
28
Posted Pricing (single item)
Question: What if even 2nd-price auction is too complicated? Def [uniform-pricing] for a given price π£ , mechanism serves the first agent willing to pay π£ (break ties in arrival order). E.g., πΉ=π 0,1 , πΉ π§ =π§, πΉ=π 0,2 , πΉ π§ = π§ 2 Assumption: Distributional knowledge of demand. π£ π βΌπΉ i.i.d. for all πβ[π], where πΉ π§ = Pr π£βΌπΉ [π£β€π§] is a known cdf (cumulative density function).
29
Posted pricing (single item)
2nd price auction: ax ante probability of π winning is 1 π Theorem For values drawn i.i.d. from any distribution F, the uniform pricing of π£ = πΉ β1 (1β 1 π ) is an π πβ1 β1.58 approximation to the optimal social surplus. Proof On the board Before the values are drawn Mimic by posted price: π£ = πΉ β1 (1β 1 π )
30
Discussion Vickrey-Clarke-Groves (VCG) routing. Algorithmic theory:
Can be tricky to compute the optimal solution. Actually, for Border Gateway Protocol (BGP) is NP-hard. Algorithmic theory: if computing the optimal solution is intractable try instead to compute an approximately optimal solution. Question: Can we use an approximation algorithm and still retain the dominant-strategy incentive property? Question: Is there a general theory for designing approximation mechanisms from approximation algorithms?
31
Theory: what we want Informative: pinpoints salient features of the environment and characteristics of the good mechanisms. Prescriptive: gives concrete suggestions for how a good mechanism should be designed. Predictive: The mechanisms that the theory predicts should be the same as the ones observed in practice. Tractable: The theory should not assume super-natural ability for the agents or designer to optimize.
32
Philosophy of approximation
Picassoβs βBull studyβ
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.