Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
CDS270: Optimization, Game and Layering in Communication Networks
Lecture 1: Static Games and Classical Mechanism Design
Lijun Chen
10/04/2006
Agenda
2
Strategic games and their solution conceptsStrategic form games and dominated strategiesNash equilibrium and correlated equilibrium
Classical mechanism design Incomplete information gamesIncentive-compatible mechanismVCG mechanism
Strategic game
3
Def: a game in strategic form is a triple
is the set of players (agents)is the player strategy space
is the player payoff function Notations
: the set of all profiles of player strategies
: profile of strategies: the profile of
strategies other than player
},,{ NiNi uSNG ∈∈=
NiS i
RSui → : i
NSSSS ×××= L21
),,,( 21 Nssss L=
),,,,,,,( 1121 Niii ssssss LLL +−− =i
4
Implicitly assume that players have preferences over different outcomes, which can be captured by assigning payoffs to the outcomesThe basic model of rationality is that of a payoff maximizerFirst consider pure strategy, will consider mixed strategy later
Example: finite game
5
4,3
2,1
3,0
5,1
8,4
9,6
6,2
3,6
2,8
L M
U
R
M
D
column
row
Example: Continuous strategy game
6
Cournot competitionTwo players: firm 1 and firm2Strategy : the amount of widget that firm produces The payoff for each firm is the net revenue
where is the price, is the unit cost for firm
],0[ ∞∈is
iiii scsspsssu −+= )(),( 2121
p ici
i
Dominated strategies
7
How to predict the outcome of a game?Prisoner’s Dilemma
Two prisoners will play (C,C)Def: a strategy is (weakly) dominated for player if there exists such that
-2,-2
-1,-5
-5,-1
-4,-4
D C
D
C
is
i ii Ss ∈′
iiiiiiii Ssssussu −−−− ∈≥′ allfor ),(),(
Iterated elimination of dominated strategies
8
Iterated elimination of dominated strategies
However, most of games are not solvable by iterated elimination of dominated strategies
2,8
4,3
2,1
3,0
5,1
8,4
9,6
6,2
3,6
L M
U
R
M
D
Nash equilibrium
9
Def: a strategy profile is a Nash equilibrium, if for all ,
For any , define best response function
Then a strategy profile is a Nash equilibriumiff
*s
i
iiiiiiii Ssssussu ∈≥ −− allfor ),(),( ***
ii Ss −− ∈
}. ),( ),(|{)( iiiiiiiiiiii SsssussuSssB ∈′∀′≥∈= −−−*s
).( **iii sBs −∈
Examples
10
Battle of the Sexes
Soccer
2,1
0,0
0,0
1,2
Ballet
Ballet
Soccer
Two Nash equilibria (Ballet, Ballet) and (Soccer, Soccer)
Cournot Competition
11
Suppose a price function Suppose cost Then, the best response function
Nash equilibrium satisfies , i.e.,
)}(1,0max{)( 2121 ssssp +−=+
10 21 ≤==≤ ccc
2/)1()(2/)1()(
112
221
cssBcssB
−−=−−=
)()(
122
211
sBssBs
==
)1()1(
2
1
cscs
−=−=
3/3/
Second price auction
12
An object to be sold to a player inEach player has a valuation of the object. We further assume The players simultaneously submit bids,The object is given to the player with highest bid. The winner pays the second highest bid.The payoff of the winner is his valuation of the object minus the price he pays. All other players’ payoff is zero.
N
i iv021 >>>> Nvvv L
Nbb ,,1 L
13
is Nash equilibriumPlayer 1 receives the object and pay , and has payoff . Player 1 has no incentive to deviate, since his payoff can only decrease.For other players, the payoff is zero. In order to change his payoff, he needs to bid more than
, but that will result in negative payoff. So, no player has incentive to change.
Question: are they more Nash equilibria?
),,(),,( 11 NN vvbb LL =
2v021 >−vv
1v
14
Not all games have (pure) Nash equilibriumMatching Pennies
-1,1
1,-1 -1,1
1,-1
Heads Tails
Heads
Tails
Mixed strategies
15
Let denote the set of probability distribution over player strategy spaceA mixed strategy is a probability mass function over pure strategiesThe payoff of a mixed strategy is the expected value of the pure strategy profiles
iΣ
iSi
ii Σ∈σ
ii Ss ∈
∑ ∏∈ ∈
=Ss
iNj
jji susu )())(( σ
Mixed strategy Nash equilibrium
16
Def: a mixed strategy profile is a (mixed strategy) Nash equilibrium if for all
A mixed strategy profile is a (mixed strategy) Nash equilibrium if for all
The payoff is the same for allThe payoff for each is not larger
*σ
i
iiiiiiii uu Σ∈≥ −− σσσσσ allfor ),(),( ***
*σ
i
iiiiiiii Sssuu ∈≥ −− allfor ),(),( *** σσσ
),( *iii su −σ )(supp *
iis σ∈
),( *iii su −σ )(supp *
iis σ∉
Example
17
Assume row (column) player choose “ballet” with probability ( ) and “soccer” with probability ( )
Mixed strategy Nash equilibrium is
Soccer
2,1
0,0
0,0
1,2
Ballet
Ballet
Soccer
pp−1 q−1
q
)1(20)1(01)1(10)1(02ppppqqqq−×+×=−×+×−×+×=−×+×
⎩⎨⎧
==
3/13/2
qp
Existence of Nash equilibrium
18
Theorem (Nash ‘50): Every finite strategic game has a mixed strategy Nash equilibrium.Example: Matching Pennies game has a mixed strategy Nash equilibrium (½, ½; ½, ½)
Proof: using Kakutani’s fixed point theorem. See section 1.3.1 of the handout for details
1,-1 -1,1
1,-1
Heads TailsHeads
Tails -1,1
Continuous strategy game
19
Theorem (Debreu ’52; Glicksberg ’52; Fan ’52): Consider a strategic game with continuous strategy space. A pure strategy Nash equilibrium exists if
is nonempty compact convex setis continuous in and quasi-concave in
Theorem (Glicksberg ’52): Consider a strategic game with nonempty compact strategy space. A mixed strategy Nash equilibrium exists if is continuous.
},,{ NiNi uSN ∈∈
iS
iSiu S
},,{ NiNi uSN ∈∈
iu
Correlated equilibrium
20
In Nash equilibrium, players choose strategies independently. How about players observing some common signals? Traffic intersection game
Two pure Nash equilibria: (stop, go) and (go, stop)One mixed strategy equilibrium: (½, ½; ½, ½)If there is a traffic signal such that with probability ½ (red light) players play (stop, go) and with probability ½ (green light) players play (go, stop). This is a correlated equilibrium.
0,0
2,2 1,3
Stop GoStop
Go 3,1
21
Def: correlated equilibrium is a probability distribution over the pure strategy space such that for all
A mixed strategy Nash equilibrium is a correlated equilibriumThe set of correlated equilibria is convex and contains the convex hull of mixed strategy Nash equilibria
)(⋅pi
iiis
iiiiiiii Stsstussusspi
∈≥−∑−
−−− , allfor 0] ),(),()[,(
Dynamics in games
22
Nash equilibrium is a very strong concept. It assumes player strategies, payoffs and ration-ality are “common knowledge”.“Game theory lacks a general and convincing argument that a Nash outcome will occur”.One justification is that equilibria arise as a result of adaptation (learning).
Consider repeated play of the strategic gamePlayers are myopic, and adjust their strategies based on the strategies of other players in previous rounds.
23
Best response
Fictitious play, regret-based heuristics, etcMany if not most network algorithms are repeated and adaptive, and achieving some equilibria. Will discuss these and networking games later in this course.
))(()1( tsBts iii −=+
Classical mechanism design (MD)
24
Mechanisms: Protocols to implement an outcome (equilibrium) with desired system-wide properties despite the self-interest and private information of agents.Mechanism design: the design of such mechanisms.Provide an introduction to game theoretic approach to classical mechanism design
Game theoretic approach to MD
25
Start with a strategic model of agent behaviorDesign rules of a game, so that when agents play as assumed the outcomes with desired properties happen
Incomplete information games
26
Players have private type Strategy is a function of a player’s type Payoff is a function of player’s typeAssume types are drawn from some objective distributionDef: a strategy profile is a Bayesian-Nash equilibrium if every players plays a best response to maximize expected payoff given its belief about distribution , i.e.,
Θ∈),,,( 21 Nθθθ L
iii Ss ∈)(θ
Rsu ii ∈)),(( θθ
),,,( 21 Np θθθ L
*s
)),(,()|(maxarg)( ** ∑−
−−−∈i
iiiiiiiisii ssups
θ
θθθθθ
i
)|( iip θθ−
Example: variant of Battle of the Sexes
27
Two types: either wants to meet the other or does notAssume row player wants to meet column player, but not sure if column player want to meet her or not (assign ½ probability to each case); and column player knows row player’s typeIf column player want to meet row player, the payoffs are Soccer
2,1
0,0
0,0
1,2
Ballet
Ballet
Soccer
28
If column player does not want to meet row player, the payoffs are
The Bayesian-Nash equilibrium is (Ballet, (Ballet, Soccer))
E[Ballet, (Ballet, Soccer)]= ½x2+ ½x0=1E[Soccer, (Ballet, Soccer)]= ½x0+ ½x1= ½
Soccer
2,0
0,1
0,2
1,0
Ballet
Ballet
Soccer
Stronger solution concepts
29
Def: a strategy profile is Ex post Nash equilibrium if every player ‘s strategy is best response whatever the type of others
Def: a strategy profile is dominant strategy equilibrium if every player ‘s strategy is best response whatever the type and whatever strategy of others
*s
i
iiiiiisii ssusi
−−−∈ θθθθ allfor )),(,(maxarg)( **
i
iiiiiisii ssusi
−−−∈ θθθθ ,s allfor )),(,(maxarg)( i-*
*s
Example: second price auction
30
The type is player valuationEach player submit bid A dominant strategy is to bidPlayers don’t need to know valuations (types), or strategies of others.
iv)( ii vb
iii vvb =)(*
Model of MD
31
Set of alternative outcomesPlayer has private information (type) Type defines a value function for outcome for each player Player payoff for outcome and paymentsThe desired properties are encapsulated in the social choice function
e.g., choose to maximize social welfare, i.e.,
O
i iθ
Rov ii ∈);( θOo∈ i
iiiii povou −= );();( θθ o
ip
Of →Θ:
o∑
∈=
iiOoouf );(maxarg)( θθ
32
The goal is to implement social choice function
A mechanism is defined by an outcome rule and a payment ruleA mechanism M implements social choice function if , where the strategy profile is an equilibrium solution of the game induced by M.
)(θf
…
)( 11 θs
)( NNs θ
Mechanism )(sgo =
)(),,( 1 sppp N =L},{ pgM =
OSg →:nRSp →:
)(θf)())(,),(( *
11* θθθ fssg NN =L
),,( *1
*Nss L
33
Properties of social choice functions and mechanisms
Pareto optimal:
Efficient:
Budget-balance:
A mechanism that implements the corresponding social choice functions is called Pareto optimal, efficient or budget-balanced mechanisms, respectively.
)),((),( j )),() ),every for if θθθθθθ scfuau(scfu(a,θuf(a jjii <∃⇒>≠
∑∈i
iia
avf ),(maxarg)( if θθ
∑ =i
ip 0)( if θ
Incentive-compatible mechanism
34
Revelation principle: any mechanism can be transformed into an incentive compatible, direct-revelation mechanism that implements the same social choice functionDirect-revelation mechanism is a mechanism in which player strategy space is restricted to their types
…
1θ
Nθ
Mechanism )(θgo =
)(),,( 1 θppp N =L},{ pgM =
35
Incentive-compatible means the equilibrium strategy is to report truthful information about their types (truth-revelation).
First price auction is not incentive-compatible. In first price auction, the buyer with highest bid gets the object and pays his bid. The second price auction is incentive compatible, direct-revelation mechanism.
Truthful mechanism
36
Truthful (aka “strategy-proof”) mechanism: truth-revelation is a dominant strategy equilibrium.
Very robust to assumption about agent rationality and information about each otherAn agent can compute its optimal strategy without modeling the types and strategies of others
Vickrey-Clarke-Groves mechanisms
37
VCG mechanism:Collect from agents
: select an outcome: agent pays ,
where
),,,( 21 Nθθθθ L=
)(θg ∑∈
∈i
iiOoovo );(maxarg* θ
)(θp i ∑∑≠≠
− −ij
jjij
ji
j ovov );();( * θθ
∑≠∈
− ∈ij
jjOo
i ovo );(maxarg θ
38
Theorem: VCG mechanism is efficient and truthful.Proof: VCG mechanism is the only mechanism that is efficient and strategy-proof amongst direct-revelation mechanisms.
∑ ∑≠ ≠
−− −+=
ij ijj
ijjjiiiii ovovovu );();();(),( ** θθθθθ
Combinatorial auction
39
Goods Outcomes: allocations , where and s are not overlapped.Agent valuation forGoal: allocate goods to maximizeApplications: wireless spectrum auction, course scheduling, …
P
),,( 1 NAAA L= PAi ⊆
iA
);( iii Av θ PAi ⊆
∑i
iii Av );( θ
40
Two items A and B; 3 agents (taken from Parkes)Valuation
Agent 3 wins AB and pays 10-0=10.
1
A B AB
32
5
00
05
0 12
5
5
41
Another valuation
Agents 1 and 2 win and each pays 7-5=2
1
A B AB
32
5
00
05
0 7
5
5
Remarks
42
Only consider the incentive issue: to overcome the self-interest of agentsNot discuss computational and informational issues. Will discuss these in distributed mechanism design and its applications in networking.