Title: Computational Issues in Game Theory Lecture 3: Other topics
1Computational Issues in
Game Theory Lecture 3
Other topics
- Edith Elkind
- Intelligence, Agents, Multimedia group (IAM)
- School of Electronics and CS
- U. of Southampton
2Plan
- Selfish routing
- Voting and voting manipulation
- Coalitional games
3Selfish routing model
- Intuition traffic from A to B
- individual drivers make decisions
- no singe car has a perceivable effect
- A directed graph G (V, E)
- Source-sink pairs (si, ti) for i1, , k
- rate ri 0 of traffic between si and ti
- For each edge e, a latency function le(.)
- assume le(.) 0, monotone, differentiable
r1 1
4Non-atomic Players
- Routing game infinitely many players of
infitezimal size (particles of flow) - individual player cannot affect the payoff for
others, but can affect its own payoff - can be viewed as a limit of an
atomic game as n ? infinity
5Selfish Flows
r 1
Goal minimize average travel time
- upper edge x
- lower edge 1-x
- average travel time 1/21/2 11/2 3/4
- But traffic on the lower edge is envious!
xx1(1-x) 2x-1 x ½
Envy-free (selfish) flow average travel time 1
6Selfish Agents
- Total latency is a measure of social welfare.
- Agents are selfish
- do not care about social welfare
- want to minimize personal latency
- (Wardrop)-Nash equilibrium no one wants
to change strategy, i.e., switch paths - equivalently all flow paths from si to ti
have the same travel
time. - unique WNE in pure strategies
- Cost of anarchy Worst Nash/Opt ?
- cost of stability Best Nash/Opt
7Cost of Anarchy General Latencies
- Previous examples cost of anarchy 4/3
- Corollary for any C gt 0, the cost of anarchy is
C. - What can we do now?
Opt xxd1(1-x) (d1)xd1-1 xOPT
1/(d1)1/(d1) 1 o(1) C(f) 1/(d1) o(1)
? 0
Nash all traffic chooses upper path, C(f) 1
8The price of anarchy bounds
- Bicriteria bound
Theorem if f is a NE flow for (G, r,
l) and f is feasible for (G,
2r, l), then C(f) C(f). - Linear latencies
Theorem if f is a NE flow for (G, r, l),
f is an optimal flow for
(G, r, l), and all
latencies are linear, then
C(f) 4/3 C(f). -
9Braesss Paradox
.5
Opt Nash, Latency 1.5
.5
What if we build a new road?
.5
0
Latency 1.75
Latency 1.5
Latency 2
.5
10Severity of Braesss paradox
- Let f be a NE flow for (V, E),
gE is a NE flow for (V, E), E ? E - what is maxEC(f)/C(gE)?
- n/2
- k (if you can only delete
k edges) - can we detect Braesss paradox?
- i.e., is there an E ? E s.t. C(gE) lt C(f)?
- no NP-hard
11Taxes can help!
C
- Braesss paradox
assign a tax of 1
to CD - the flow is optimal now!
- can we ensure optimal flow
by taxing the traffic? - yes marginal cost pricing (MCP)
- intuition each user pays for the extra delay
they cause - if users minimize latencytax, MCP induces
optimal flow - maybe at the expense of large taxes
- alternative goal minimize total disutility
x
0
A
B
0
0
x
D
12Voting
- n voters
- m candidates c1, , cm
- each voter has a preference over candidates
- total ordering ci gt cj gt
- voting rule
voters preferences ? candidates - or preferences ? total orders of candidates
13Voting rules examples
- Plurality each candidate gets 1 point from each
voter that ranks him first - Borda each candidate gets
- m-1 point from each voter that ranks him 1st
- m-2 points from each voter that ranks him 2nd
-
- Veto each candidate gets 1 point from each voter
that does not rank him last - Scoring rules each candidate gets
- ai points from each voter that ranks him ith
14Voting rules binary cup
Do most voters prefer A to B?
R1
F
R2
C
F
R3
C
15Voting rules more examples
- Copeland
- is score j majority prefers i to j
- candidate with the highest score wins
- Maximin
- is score min j v v prefers i to j
- candidate with the highest score wins
- STV
- if someone has more than n/2 votes, he wins
- else, find a candidate with the least number of
1st place votes and cross him out
from all ballots - repeat until there is a majority winner
16Voting rules criteria
- Pareto-optimality
- if everyone prefers i to j, j does not win
- not true for Binary Cup
- Condorset-consistency
- if i is such that for each j
majority of voters
prefers i to j, then i wins - not true for Borda
- Monotonicity
- if we move the winning candidate up in one of the
votes, he still wins - not true for STV
17Arrow theorem
- Let F output total orderings of candidates
- Independence of irrelevant alternatives (IIA)
- for any
- two profiles R(R1, , Rm) and S(S1, , Sm)
- pair of candidates a, b
- such that
- i prefers a to b in R iff i prefers a to b in S
(for each i) - F(R) and F(S) order a and b in the same way
- Theorem there is no F for m 3 candidates that
is - non-dictatorial
- Pareto-optimal
- satisfies IIA
18Manipulation
- Can in be in the voters best interest to
misrepresent his preferences? - 50 voters AgtBgtC
- 50 voters CgtAgtB
- 20 voters BgtAgtC
- voting rule plurality (draws resolved by a coin
toss) - any voter with preference BgtAgtC
is better off
voting AgtBgtC - in this scenario, STV is truthful
- but there are other examples
19Gibbard-Satterthwaite theorem
- For any voting rule for m3 candidates that is
- non-dictatorial
- symmetric wrt candidates
- there is a set of preference profiles such that
- one of the voters is better off voting
non-truthfully - Extends to randomized voting rules
- nonmanipulable rules are all of the form
- w.p. p, select a random voter and elect his top
candidate - w.p. 1-p, select a random pair of candidates,
and elect
the better of the two
20Escaping GS theorem computational barriers
- Plurality is easy to manipulate
- rank your favorite electable candidate first
- STV less so
- STV is NP-hard to manipulate
Bartholdi, Orlin91 - General methods for constructing
manipulation-resistant voting schemes Conitzer,
Sandholm03, Elkind, Lipmaa05
21Escaping GS theorem structured
preferences
c1
c4
c3
c2
c5
c6
c7
cv
- Candidates are ordered from left to right
- Each voter v has a favorite candidate cv
- If ci is between cv and cj, then ci gtv cj
- v may prefer c7 to c2
- Truthful voting rule
- order voters by their favorite candidates
- winner favorite candidate of the median voter
22Coalitional games
- Intuition players can cooperate to achieve goals
and then share the payoff - Payoff function coalitions ? utilities
- transferable utility each coalition has a value
- it is assumed that the players can share the
value - example team of workers building a house
- non-transferable utility for each coalition,
there is a value for each player - example trade agreement between countries
23Coalitional games examples
- Weighted voting games
- n players with weights w1, , wn, quota T
- w(C) Si?Cwi
- v(C)1 (C wins) if w(C) T, v(C)0 otherwise
- Network flow games
- network G(V, E), with source s and sink t, edge
capacities c1, , cE - players are edges
- v(C) is the max s-t flow that C can carry
24Coalitional games special classes
- Simple games value of each C is 0 or 1
- weighted voting games
- variant of network flow games (thresholded)
- Monotone games if S ? T, then v(S) v(T)
- weighted voting games with non-negative weights
(but not general WV games) - Superadditive games
if SnTØ, then
v(S U T) v(S)v(T) - network flow games
- usually not the case for simple games
- can assume that the grand coalition will form
25Dummy players
- a is a dummy if for any C, v(CUa)v(C)
- Weighted voting games e.g.,
- a has weight 1
- all other players have even weight
- quota is even
- Network flow games
- no loop-free s-t path containing a
- NP-hard to detect in many games
- e.g., WV games, NF games
a
26How to distribute payoffs fairness
- Shapley value
agents expected contribution - Intuition agents join coalition
in random order - how does the value change when i joins?
- Formally
- p a permutation of n players
- Pred(i, p) j p( j ) lt p( i )
- fi(v) 1/n! Sp v(Pred(i, p)U i ) -
v(Pred(i, p)
27Shapley value
axiomatic characterization
- efficiency Si fi(v) v(grand coalition)
- dummy property dummy players receive 0
- symmetry if for any C s.t. i, j not in C
we have v(C U i ) v(C U j ),
then fi(v)
fj(v) - additivity for two coalitional games v, w,
define vw by (vw)(C) v(C)w(C).
We
have fi(vw) fi(v)fi(w) - Shapley value is the only value distribution
scheme that satisfies 1 - 4 !
28Computing Shapley value
- Weighted voting games
- computing Shapley value is NP-hard
- hard to decide if i is a dummy (i.e., fi 0)
- unless PNP, there is no algorithm for computing
fi in time poly(n, log max wi) - poly(n, max wi) algorithms exist
- for w1, , max wi, count coalitions of weight
wi - dynamic programming
- Network flow games
- computing Shapley value is NP-hard
- hard to decide if an edge is a dummy
29How to distribute payoffs
stability
- p (p1, , pn) payoff vector
- core the set of payoff vectors such that noone
wants to deviate - p is in the core if for any J we have p(J) v(J)
- Core can be empty
- Lemma in simple games,
the core is empty ltgt
no player belongs
to
all winning coalitions
pj gt 0
J
p(J) lt 1
30Relaxing the Notion of the Core
- e- core p is in the e- core iff for any J
p(J) v(J) - e - each coalition C gets at least v(C) - e
- nonempty for large enough e
- least core smallest non-empty e-core
- if least core e - core
- there is a p s.t. p(J) v(J) - e for any J
- for any e lt e there is no p
s.t. p(J)
v(J) - e for any J
31Coalition structures motivation
- Definitions of Shapley value and core
distribute the value of the grand
coalition - makes sense for superadditive games
- less so for simple games
- Can several coalitions form simultaneously?
- weighted voting games
- suppose C1 and C2 are disjoint, w(C1) gt T, w(C2)
gt T - merging C1 and C2 is bad for social welfare
- Sometimes, grand coalition cannot form
- legal restrictions, cant negotiate
32Coalition structures definition
- Coalition structure CS C1, , Ck
a partition of all agents
into coalitions - each agent belongs to some Ci
- Payoffs are distributed within coalitions
- p has to satisfy p(Ci)v(Ci) for i 1, , k
- CS-core (CS, p) is in CS-core (i.e.,stable) if
no group of agents wants to deviate - for every S, p(S) v(S)
33CS-core vs. core
- Weighted voting game
- 5 players a, b, c, d, e
- weights 2, 2, 2, 3, 3 T 6
- 2 disjoint winning coalitions a, b, c, d, e
- core is empty
- ½ -core is non-empty (1/6, 1/6, 1/6, 1/4, 1/4)
- CS-core is non-empty
CSa, b, c, d, e, p(1/3, 1/3,
1/3, 1/2, 1/2)