CSE 522 – Algorithmic and Economic Aspects of the Internet
Instructors: Nicole ImmorlicaMohammad Mahdian
Previously in this class
Properties of social networks
Generative models for power law distribution and power law graphs
Generative models for small-world networks
This Lecture
Final remarks on small-world networks
Network formation games, and a short introduction to game theory
Geographic Routing
Experiments suggest that the first criterion that people use for forwarding a message is geographic proximity.
Kleinberg: In a 2-d grid with long-range contact probability proportional to dist –2, “geographic routing” works.
However, experiments show that this probability is closer to dist –1.
Geographic Routing, cont’d
Liben-Nowell et al., PNAS 2005: Justification: in Kleinberg’s model, population is
distributed uniformly on a 2-d grid, but in the real world the distribution is not uniform.
Model: probability of a long-range contact from u to v proportional to the inverse of the # of people that are closer to u than v.
Result: In this model, geographic routing works. Experiments on ~500,000 blogs on LiveJournal
confirms the assumption of the model.
Getting Closer or Drifting Apart? Rosenblat and Mobius, QJE 2004:
Technology has made it less costly to interact with people across the globe (‘global village’).
As a result, people become more selective in whom to interact/collaborate with.
Could this fragment the social network into clusters of like-minded people?
Prominent example: scientific community
Getting Closer or Drifting Apart? Model:
Agents of types A and B are arranged uniformly around a circle.
Each person collaborates with a fixed # of other people, and receives a payoff from each collaboration.
The payoff for collaborating with someone of the same type is higher.
Collaborating with someone who is not close has a cost C. Results:
As C decreases, individual separation (diameter) decreases, but group separation increases.
Experiments on co-authorship among economists (69-99)
Network Formation Games
Models that use formal game theoretic reasoning to study network formation Individuals in a network face economic incentives
to form or break links with other individuals Individuals make self-motivated decisions about
which links to form
Applications: professional network, Internet
Incentives in Networks
Each individual is a source of benefits (information, resources)
Others can share the benefits of an individual via formation of links
Link formation is costly (time, effort, money)
Given these incentives, which links will form?
Game Theory Framework
Set of players (agents) Each player selects a strategy from the set of
allowed strategies. A payoff function specifies how much each
player receives given the strategy profile.
An equilibrium is a strategy profile in which no player can benefit by unilaterally changing his strategy.
Network Formation Games
Players {1,…,n} are nodes in the network
Each player i must simultaneously choose some subset of {1,…,n} as his strategy si
A strategy profile defines a (directed) graph G Nodes are players Edge (i,j) is in G if j 2 si
Example: Graph
Players = {1, 2, 3, 4}
1
4 3
2s1 = {4} s2 = {3,4}
s3 = {4}s2 = {3}
Game Theory Framework
Let Ni = |si| be number of links i forms Let Ci be “connectedness” of i (definition
varies depending on model) Given a strategy profile (i.e., graph G), the
payoff for a player i is a function i(Ni,Ci) decreasing in Ni and increasing in Ci
Players seek to maximize their payoff
Example: Payoffs
E.g., i is number of nodes that i can reach via a directed path in G minus the number of links i forms
1
4 3
21 = 2 – 1 = 1 2 = 2 – 2 = 0
3 = 1 – 1 = 03 = 1 – 1 = 0
Equilibrium Networks
When do we expect a graph to be stable?
A graph G is a Nash equilibrium if no player has an incentive to unilaterally sever or create links, i.e. for any other strategy s’i of i, his payoff ’i in the resulting graph G’ is at most his payoff i in G
Example: Equilibrium Networks
Node 1 has an incentive to sever connection to 4 and instead form a connection to 2 for a resulting payoff of ’1 = 3 – 1 = 2
1
4 3
21 = 2 – 1 = 1’1 = 3 – 1 = 2
2 = 2 – 2 = 0
3 = 1 – 1 = 03 = 1 – 1 = 0
Strict Equilibria
What if there is another strategy for a player which does not change his payoff?
A graph G is a strict Nash equilibrium if each player’s strategy is his unique best-response, i.e. for any other strategy s’i of i, his payoff ’i in the resulting graph G’ is strictly less than his payoff i in G
Example: Strict Equilibria
Any unilateral deviation by a node strictly decreases his payoff
1
4 3
21 = 3 – 1 = 2 2 = 3 – 1 = 2
3 = 3 – 1 = 23 = 3 – 1 = 2
Models: Bala and Goyal
Two models (Bala and Goyal, Econometrica 2000) One-way flow: A link can be used only by the
person who formed it to send information Two-way flow: A link between two people can be
used by either person Model is frictionless if value of information does
not decay with distance: Ci is number of nodes i can reach in G by a path of any length
Equilibria in Bala and Goyal
For any payoff function In both models, every Nash equilibrium is either
connected or empty In the one-way flow model, the only strict Nash
equilibria are the directed cycle and/or the empty network
In the two-way flow model, the only strict Nash equilibria are the center-sponsored star (one node connects to all others) and/or the empty network
Experimentation: Falk and Kosfeld Implemented game with 4 players
Players were offered 10 points (worth 65 cents each) for each player they had a direct or indirect connection to (including themselves)
Players were charged C points for each link they formed
There were five treatments: C = 5, 15, and 25 in one-way model and C = 5 and 15 in two-way model
Predictions vs Results
Treatment Strict NE Freq. of NE
Freq. of Strict NE
C=5, 1-way Circle 48% 41%C=15, 1-way Circle, ; 52% 52% (all circ.)C=25, 1-way Circle, ; 59% 59% (83% circ)C=5, 2-way Star 31% 0%
C=15, 2-way ; 9% 0%
Explanations: Symmetry of strategies/coordination issue Inequity aversion (people prefer equal payoffs) Concern for efficiency (empty graph gives no payoffs)
Dynamics in Bala and Goyal
Does not imply that equilibria are unique! For example, there are n possible stars. Can players find an equilibria?
Consider following best-response dynamic Start from an arbitrary initial graph In each period, each player independently
decides to “move” with probability p If a player decides to move, he picks a new
strategy randomly from his set of best responses to graph in previous period
Dynamics in Bala and Goyal
Theorem: In either model, the dynamic process converges to a strict Nash equilibrium network with probability one.
Simulations show that rate of convergence is quite rapid.
Accounting for Distances
Bala and Goyal Value of information decays by a factor of for
each link traversed (model with “friction”) Results similar to frictionless models still hold
Fabrikant et al. (PODC 2003) Value of connection to j for a node i is -d(i,j) (and
payoff function is linear) Nash equilibria become slightly more complex
(e.g., trees are Nash equilibria in some cases)
Model: Fabrikant et al.
The payoff incurred by player i is i = - Ni – jd(i,j)
where Ni is the number of links formed by i and d(i,j) is the distance between i and j in the underlying undirected network (two-way flow model).
Equilibria: Fabrikant et al.
For < 1, complete graph is only Nash equilibrium
For > n2, all Nash equilibria are trees
Conjecture: For some constant, all strict Nash equilibria are trees.
Upcoming paper in SODA 2006 disproves this.
Efficiency of the Equilibria
The social welfare or efficiency of a strategy profile in a game is defined as the sum of payoffs of all players
The price of anarchy of a game is the ratio of least-efficient Nash equilibrium to the most-efficient strategy profile (which need not be an equilibrium)
Theorem [Fabrikant et al.]: For any tree Nash equilibrium T, the welfare of T to the optimum is at most 5.
Other Network Formation Games
What if agents cooperate to form links?
Cooperative Game Theory
Players cooperate to achieve a common goal (e.g., building a network).
Achieving goal has a value for each agent. In a transferable utility game, agents must
additionally decide how to share this value among each other.
As in non-cooperative game theory, analyze stable situations, but now must consider coalitions as well as individuals.
Cooperative Network Formation Jackson and Wolinsky
Studied network formation as a cooperative game with transferable utilities, in particular individuals can share cost of links.
Showed there are natural situations in which no efficient network is pairwise stable for any utility-transfer rule.