NOBEL WP2 2004. Szept. 2-3. [email protected]
Game Theory in Inter-domain Routing
LÓJA Krisztina - SZIGETI János - CINKLER Tibor
BME TMIT
Budapest, Hungary
NOBEL WP2 2004. Szept. 2-3. [email protected]
Contribution to D15
• The research is a part of WP2 A.2.3
• Contributed to D15 Chapter 6.5
The Influence of Advertised Costs / Weights on the Utilization and Incomes of Operators
– Multiprovider network – competition
NOBEL WP2 2004. Szept. 2-3. [email protected]
Game theory: decision theory with at least two players where the decision of a players has an effect on the payoff of other players.
Non-cooperative game theory: the players can not make binding agreements.
What we need for a game theoretic model:
-the set of players
-the set of strategies
-the payoffs (that depends on all the decision)
-the information they have when they decide
Game Theory – Overview
NOBEL WP2 2004. Szept. 2-3. [email protected]
Mapping of Game Theory Objects
Player
Domain
Strategy
Connection
Cost Multiplier
2.0×
Payoff
Income /
Revenue
Information
Various
Information
connection
NOBEL WP2 2004. Szept. 2-3. [email protected]
Inter-domain Routing
There are more domains (Autonomous Systems, AS)
•The traffic goes through them
•They decide the price and get a profit
•The traffic travels on the cheapest route
•The payoff of a domain depends on
the decision of other domains as well.
AS 1
AS 2 AS 3
NOBEL WP2 2004. Szept. 2-3. [email protected]
Game Theoretic Model for Inter-domain Routing
Players: domains (autonomous systems)
Decision: price of the advertised connections
Set of strategies: the integers or the real numbers or an interval if the price has an upper and a lower bound or a set of prices
Payoff: the profit of the domains, the price multiplied with the congestion
Information: the players know the network and in our first models they also know the traffic demand
NOBEL WP2 2004. Szept. 2-3. [email protected]
Path Changing Situation
S T1 1
1
1
1
3
33
Path A Cost: 5Path B Cost: 9
1.S T
1 11
1
1
4
33
Path A Cost: 6Path B Cost: 9
2.S T
1 11
1
1
5
33
Path A Cost: 7Path B Cost: 9
3.
S T1 1
1
1
1
6
33
Path A Cost: 8Path B Cost: 9
4.S T
1 11
1
1
7
33
Path A Cost: 9Path B Cost: 9
5.S T
1 11
1
1
8
33
Path A Cost: 10Path B Cost: 9
6.
NOBEL WP2 2004. Szept. 2-3. [email protected]
We can determine for the game:
- global optimum (when the summarized payoff of all the players is maximal)
- Nash equilibrium (the stable point of the game, when no player has an incentive to deviate unilaterally, it is not always optimal)
- Pareto optimum (when the players cannot change strategies so that no player will have lower payoff and at least one of them higher payoff)
What can we do?
NOBEL WP2 2004. Szept. 2-3. [email protected]
We can determine for every domain:
- the maximal payoff
- the minimal payoff
- the reservation cost (the payoff that a domain can guarantee for itself, the maximum of the minimal payoffs)
NOBEL WP2 2004. Szept. 2-3. [email protected]
Example 1
AS 1
AS 2 AS 3
S 1
S 2S 3
T 1
T 2 T 3
Let we assume that
•all the connections in a domain have the same price
•one unit of traffic is transferred from Si to Ti (i=1,2,3)
•the traffic travels on the cheapest route, if it is not unique, the traffic is divided uniquely
•the price is a positive integer not higher than an upper bound, the payoff = price *congestion
•it is a one-shot game and the players decide simultaneously Global optimum: the price is the highest possible price in every domain
Reservation cost can be reached when the price is the highest
Pareto optimum: the price is the highest possible price in every domain
Although it is not a Nash equilibrium, if a domain unilaterally deviates
and wants lower price, it gets higher profit
NOBEL WP2 2004. Szept. 2-3. [email protected]
AS 1
AS 2 AS 3
Let we assume, that the prices of different connections of the domain can be different, and there is an upper bound on every path, if the price is higher, the users will not use that path.
In this case there are more Nash equilibria, all of them are Pareto optimal but in the case of simultaneous decisions the players cannot choose the same Nash equilibrium without previous discussion.
Example 2
NOBEL WP2 2004. Szept. 2-3. [email protected]
AS 1
AS 2 AS 3
S 1
S 2S 3
Let we assume that the users do not use a path if its price is higher than an upper bound, and the connections have an upper bound in the capacities.
The global optimum is not Nash equilibrium. It would be a Nash equilibrium, however, if we could assume that a traffic do not choose the cheaper path if it is longer.
Nash equilibrium is useful when we model dynamic pricing with repeated games.
Example 3
NOBEL WP2 2004. Szept. 2-3. [email protected]
Possibilities of refining the model
- many domains, many source-destination pairs
- instead of one-shot game we can use repeated games
- we can take into consideration:
•the dependence of the traffic demand on the price of the route
•the finite capacities of the connections
•that the domains do not know the traffic demands
Simulations will help us when the model is too difficult.
NOBEL WP2 2004. Szept. 2-3. [email protected]
Simulation
•Network topology:•3 domains•3 sources•6 wavelengths
•Traffic:•Symmetric between sources•Average load is 4 (66%)
•Strategy:•The domain calculates the allocation cost for each advertised connection and multiplies it by its constant factor {1.0, 1.2, 1.4, 1.6, 1.8, 2.0}, this means we have 6×6×6 = 216 scenarios
A
BC
S1S2
S3
NOBEL WP2 2004. Szept. 2-3. [email protected]
Results for Domain A
11.2
1.41.6
1.82
1
1.2
1.4
1.61.8
20
200
400
600
800
1000
1200
C multiplier
B multiplier
Income of domain A
11.2
1.41.6
1.82
1
1.2
1.4
1.61.8
20
200
400
600
800
1000
1200
1400
1600
C multiplier
B multiplier
Income of domain A
11.2
1.41.6
1.82
1
1.2
1.4
1.61.8
20
200
400
600
800
1000
1200
1400
1600
1800
C multiplier
B multiplier
Income of domain A
11.2
1.41.6
1.82
11.2
1.41.6
1.82 0
1
2
3
4
5
6
C multiplier
B multiplier
Unserved demandsA multiplier: 1.0 A multiplier: 1.4 A multiplier: 2.0
11.2
1.41.6
1.82
1
1.2
1.4
1.6
1.82
0
0.5
1
1.5
2
2.5
3
3.5
4
C multiplier
B multiplier
Unserved demands
11.2
1.41.6
1.82
1
1.2
1.4
1.6
1.82
0
1
2
3
4
5
6
C multiplier
B multiplier
Unserved demands
NOBEL WP2 2004. Szept. 2-3. [email protected]
Results for Domain B
11.2
1.41.6
1.82
11.2
1.41.6
1.82 0
1
2
3
4
5
6
C multiplier
B multiplier
Unserved demandsA multiplier: 1.0 A multiplier: 1.4 A multiplier: 2.0
11.2
1.41.6
1.82
1
1.2
1.4
1.6
1.82
0
0.5
1
1.5
2
2.5
3
3.5
4
C multiplier
B multiplier
Unserved demands
11.2
1.41.6
1.82
1
1.2
1.4
1.6
1.82
0
1
2
3
4
5
6
C multiplier
B multiplier
Unserved demands
11.2
1.41.6
1.82
1
1.2
1.4
1.61.8
20
200
400
600
800
1000
1200
1400
C multiplier
B multiplier
Income of domain B
11.2
1.41.6
1.82
1
1.2
1.4
1.6
1.82
0
200
400
600
800
1000
1200
1400
C multiplier
B multiplier
Income of domain B
11.2
1.41.6
1.82
1
1.2
1.4
1.6
1.82
0
200
400
600
800
1000
1200
1400
1600
C multiplier
B multiplier
Income of domain B
NOBEL WP2 2004. Szept. 2-3. [email protected]
Results for Domain C
11.2
1.41.6
1.82
11.2
1.41.6
1.82 0
1
2
3
4
5
6
C multiplier
B multiplier
Unserved demandsA multiplier: 1.0 A multiplier: 1.4 A multiplier: 2.0
11.2
1.41.6
1.82
1
1.2
1.4
1.6
1.82
0
0.5
1
1.5
2
2.5
3
3.5
4
C multiplier
B multiplier
Unserved demands
11.2
1.41.6
1.82
1
1.2
1.4
1.6
1.82
0
1
2
3
4
5
6
C multiplier
B multiplier
Unserved demands
11.2
1.41.6
1.82
B: 1.0B: 1.2
B:1.4B:1.6
B:1.8B:2.0
0
200
400
600
800
1000
1200
1400
C multiplierB multiplier
Income of domain C
11.2
1.41.6
1.82
1
1.2
1.4
1.6
1.82
0
200
400
600
800
1000
1200
1400
C multiplier
B multiplier
Income of domain C
11.2
1.41.6
1.82
1
1.2
1.4
1.6
1.82
0
200
400
600
800
1000
1200
1400
1600
1800
C multiplier
B multiplier
Income of domain C
NOBEL WP2 2004. Szept. 2-3. [email protected]
Conclusions
• Small variance of the advertised connection cost may have a great impact on the traffic routing (significant steps in the revenue figure)
• Even in our small test there are results that require explanation
• The simulation results depend on the traffic pattern (order of demands) – various patterns and finding average is needed