Inefficiency of Nash Equilibrium in Network Based Games∗* Advisor : Dr. Anil Seth Indian Institute of Technology Kanpur Mayank Sachan(Y6927253) Mrinmaya Sachan(Y6268) Indian Institute of Technology Kanpur April 19, 2010

Abstract

We study the Inefficiency of Nash Equilibrium in two important Network Based Games, namely non cooperative routing games and network formation games. Selfish agents are represented by nodes in a network and edges serve as links between them. The objective of the agents is to route traffic through the congested network. We consider important models and examples of the two types of games and survey optimal bounds on the price of stability and the price of anarchy in these models in a range of applications.



This report was submitted as part of CS497 and CS697 courses done at IIT Kanpur during Spring of 2010.

1

1

Introduction

In a general setting, an objective function is introduced that enables us to quantify the inefficiency of equilibria, and in particular to deem certain outcomes of a game optimal or approximately optimal. The definition of objective function is such that it intutivelty denotes the overall social benifit of all the players in the game.

We reiterate the basic concepts in our fundamental network models, and consider a few popular models in detail and their stability and efficiency. We also investigate how measures of the inefficiency of equilibria can guide mechanism and network design.

The main focus of our study was in understanding the bounds obtained in the price of anarchy and the price of stability of such models which in turn guides us to the understanding of the stabilty and efficiency of such games.

2

Structure of Contents

In Section 3, we start by giving the motivation behind the study of inefficiency of Nash Equilibrium of the network based games in detail. Then in Section 4, we illustrate some of the background definitions that will be useful in understanding the concepts illustrated further. The definitions are not given rigorously as we stress more on giving a intuitive explaination of them here.

In Section 5, we introduce Selfish Routing Games and study previous works on existence of pure strategy Nash Equilibrium in such games. We also study the existing bounds on Price of Anarchy and the Price of Stability in these cases.

Further in Section 6, we choose two general network formation games. After defining the model, cost structure, payment functions, utility & optimal outcome of such games we investigate their efficiency and stability. We also study some efforts to bound the Price of anarchy and stability of this model in detail.

2

3

Motivation

A study of inefficiency is necessary in games where it is prohibitively expensive to assign a central agent which makes all the players coordinate amongst themselves. A good bound on the price of anarchy guarantees that selfish behavior of the agents does not have severe consequences, and thus the benefit of imposing additional control over players actions is relatively small.

These network models have vast applications where the absence of any central authority allows the independent agents to act selfishly but in turn the greedy behavior of such agents affects the optimal performance of the overall network.

The models in turn facilitate a quantitative study of the trade-off between efficiency and stability in network formation. We pick up a range of simple network based games that model distinct ways in which selfish agents might create and evaluate networks. The varying trade-off between efficiency and stability is what motivated us to study them and evaluate their quality.

3

4 4.1

Background Definitions Nash Equilibrium

For an n player game with utility profile Ui for player i, i = 1, 2, ...n a strategy vector S=(S1 ,S2 ,...Sn ) is said to be a Nash Equilibrium if and only if for each player i and for every strategy S 0 where S and S 0 differ only in the it h component

Ui (S) ≥ Ui (S 0 ) Informally, Nash Equilibrium is a state where no player has any incentive to deviate unilaterally from it. A game can have either a pure-strategy or a mixed Nash Equilibrium (in the latter pure strategies are chosen stochastically with a fixed frequency). Every game has atleast one mixed strategy Nash Equilibrium. In a general game, there can be multiple Nash Equilibria. At the same time, no pure strategy Nash Equilibrium may exist.

4.2

Optimal Outcome

For a game, we define an objective function over all the outcomes called the “Social Good” (or “Social Cost”) of the outcome. Social Good (or alternatively Social Cost) is the sum of Utilities(or Costs) of all players in a given outcome. The strategy vector S which maximizes the Social Good (or minimizes the Social Cost) is called an “Optimal Outcome” of the game.

4.3

Price of Anarchy and the Price of Stability

In any general game, the “Price of Anarchy” is defined as the ratio of the worst value of the objective function in an equilibrium state of the game to that of an optimal outcome. Similarly, the “Price of Stability” of a game is defined as the ratio between the best objective function value of any one of its equilibria to that of an optimal outcome.

4

From the definition, it trivially follows that for any game the Price of Anarchy and the Price of Stability is never less than unity and that the Price of Anarchy is always greater than or equal to its Price of Stability. In games which exhibit only a single Nash Equilibrium, both quantities are essentially the same.

It is worth noting that a game with multiple equilibria may have a large price of anarchy even if only one of its equilibria is highly inefficient. Clearly, the price of stability is better suited to differentiate between games in which all equilibria are inefficient to those in which only some equilibrium are inefficient.

5

Selfish Routing Games

In this setting we study the effect of each player who routs certain traffic in a large communication network, such as the Internet, that has no central authority. We are given a network in the form of a directed graph G = (V, E), where V denotes the vertex set and E the directed edge set together with a set (s1, t1), ..., (sk, tk) of sourcesink vertex pairs. Each player is associated with a source-sink pair and his task is to route a certain amount of traffic from source to the sink.

Each edge has a cost which is a function of the amount of traffic that it uses. A player pays for the amount of traffic it routes through each of these edges. In the model, selfish players choose routes to minimize the cost incurred i.e. in an equilibrium outcome, all players choose a path of minimum cost.

An optimal outcome of this game is one which minimizes the sum of costs paid by all the players.

Routing Games can be further divided into two categories, Non-atomic and Atomic Selfish Routing Games. Intuitive difference between them is that in the former, each commodity represents a large population of individuals each of whom controls a negligible amount of traffic while in the latter, each commodity represents a single player who must route a significant amount of traffic on a single path.

5

We study the existence of Nash Equilibrium and the bounds on Price of Anarchy and the Price of Stability in each of these models.

5.1

Existence of pure strategy Nash Equilibrium in Routing Games

In has been proved in [2] that each instance of Non Atomic Selfish Routing game exhibits a pure strategy Nash Equilibrium which is essentially unique. That is, all the equilibrium flows of the non atomic game have the same cost. Intuitively since each player carries a negligible amount of traffic, it will always chose a path such that the sum of marginal cost of all the edges in the path are minimized. If we look deeper, it can be verified that optimal outcomes of the non atomic selfish routing games are equilibrium outcomes of the network in which the cost of each edge is replaced by the marginal cost function of each edge.

Next in the case of Atomic Selfish Routing Games it can be proved with the AAE counterexample in [2] that they do not necessarily exhibit any unique Nash Equilibrium. Similarly, there need not exist any Pure Strategy Nash Equilibrium in Atomic Selfish Routing Games. However, by placing additional restrictions on atomic instances we can guarantee the existence of a Pure Strategy Nash Equilibrium . One restriction is by having a constraint that each player routes the same amount of traffic. Another being the constraint that all edge cost functions are affine cost functions (i.e. of the form ax + b where x denotes the amount of traffic routed through the edge).

5.2

Price of Anarchy and the Price of Stability in Selfish Routing Games

Since each instance of a Non-atomic routing game exhibits a unique Nash Equilibrium, Price of Anarchy and the Price of Stability coincide in every Non Atomic instance of Selfish Routing games. It has been further proved in [2] that the Price of Anarchy in the non atomic instances depends only on the non linearity of the network cost function and not on any other factor i.e. not the network size, the network structure or the number of commodities. 6

In the case of Atomic selfish routing games, the Price of anarchy is not always defined as Pure equilibrium flows need not exist. But in the restricted Atomic Instance with affine cost functions (where equilibrium flows are guaranteed to exist), it has been √ 3+ 5 proved in [2] that the price of anarchy is at most 2 .

6

Network Formation Games

Network Formation Games are analogous to the practical scenario where the design and operation of the large computer networks (such as Internet) are carried out by the large number of independent service providers, all of whom seek to selfishly optimize the quality and cost of their own operations. Thise model in turn facilitates a quantitative study of the trade-off between efficiency and stability in network formation.

For such network formation games, the networks corresponding to Nash Equilibrium of such games is referred to as “Stable”. To consider the overall quality of the network the Social Cost is considered. Also, networks that optimize Social Cost are said to be “Optimal” or “Socially Efficient”. We henceforth consider some of the specific network formation games and study their stability and efficiency.

6.1

The Local Connection Game

This is a simple Network formation game where players can form links to other players. Players look to minimize the sum of their distances to all other nodes at the same time building as few edges as possible. Focus is still to study the inefficiency that results from the selfish behavior of these network builders.

7

6.1.1

The Model

As described in [3], the nodes in the graph are identified with players and the network is to be built. The strategy of a player is the set of edges which a player will build having itself as one of its end points. The distance between two nodes is counted by the number of hops. The cost of building an edge is taken to be constant α for all edges. Each player seeks to make the distances to all other nodes small, and to pay as little as possible. More P precisely, player us objective is to minimize the sum of costs and distances αnu + v dist(u, v), where nu is the number of edges bought by player u.

6.1.2

Bounds on the Price of Anarchy and the Price of Stability

As proved in [4] for α < 2 the social optimum is reached when the number of edges is maximum. Hence, the complete graph is the optimal outcome for α < 2. For α > 2, the social optimum is obtained by minimizing the number of edges, hence a star will be the optimal outcome for α > 2.

It has been shown in [3] that the price of anarchy is O(1) whenever α is O(n). More generally, the price of anarchy is O(1 + √αn ). It can also be trivially seen that for α > n2 , the Nash Equilibrium is a tree because any player will build an edge only if two nodes of the graph are non connected. This is because the cost of building an edge supersedes the gain obtained by reducing the distances to all the nodes.

It has also been proved in [4] that for α < n2 the price of anarchy is bounded by O(d) where further that d can be at most √ √ d is the diameter of the graph. It can be verified 2 α, hence bounding the price of anarchy to O( α).

Motivated by the results of upper bound of O(1) for these range of values of α, we ask the question whether an upper bound of O(1) can be obtained for all values of α in general.

For this we studied an important observation made in [4] that for any transient Nash Equilibrium tree T , the price of anarchy of T is less than a 5 (which is a constant). 8

The proof exploits the unique property of trees that they have at least one central node. A central node is one such that all the sub trees rooted at it have size < n2 where n is the total number of nodes in the tree.

Armed with this special property of trees, [4] makes a conjecture that for any network graph there always exists a constant A such that for all α > A, all the non transient equilibrium are trees. If we assume that this conjecture holds true, then a very important result that we always have a constant A such that for all α > A the price of anarchy is O(1) follows. However, this conjecture was disproved in the subsequent paper titled “On Nash Equilibrium of Network Creation Game” by Liam Roditty. More precisely, it was shown in [5] that for any positive integer n0 , there exists a graph built by n ≥ n0 players, which contains cycles and forms a non-transient Nash equilibrium, for any α . On the other hand, it was also shown in [5] that for α > 12nlogn every Nash Equilibrium of the graph is a tree. This further implies that for α > 12nlogn the price of anarchy is O(1).

The results were also improved for every value of α in [5]. Specifically, a constant √ upper bound for α ≤ n and for α ≥ 12nlogn was proved. For all intermediate 1 2 2 values, an upper bound of O(1 + (min( αn , nα )) 3 ) was derived. It still remains an open problem whether an O(1) upper bound exists for all values of alpha or not.

Further, we looked at a more complex problem on network formation where links are made not just individually but are introduced bilaterally with the consent of both players between which the corresponding link is created.

6.2

Network Formation Games with Bilateral Contracts

It is a Network Formation Game which we studied before where finite number of players wish to route traffic to each other. The major difference being that now the nodes contract bilaterally to form communication links with each other while in the previous game, each node had the liberty to set up a link individually, which could 9

be used by both the players between whom the link was created. The cost function of this model is also a little more complex.

Any outcome of this game is said to be “Stable” or in “Nash Equilibrium” if no node has an incentive to deviate unilaterally and no two pair of nodes have any incentive to deviate bilaterally.

Further the edges formed can be either directed (a practical counterpart for such models is Mobile Networks) which means that the traffic can flow only in a single direction or undirected (a practical counterpart for such models is Ad Hoc Networks) which means the traffic can flow in both the directions.

6.2.1

The Model

We will study both the models of such games. We build with a graph G(V, E) where V denotes the set of vertices and E denotes the set of edges of the graph. Each selfish player is associated with a vertex of the graph.

6.2.2

Cost Structure

In the model we will assume that the traffic is routed by the shortest path, the distance being measured in the number of hops. The model captures the cost in these three components: • The cost of routing traffic: We assume that node i experiences positive routing cost of ci per unit of traffic. Thus given a graph G the total cost experienced by node i is Ri (G) = ci fi (G) where fi (G) denotes the total traffic that transits through i added to the total traffic received by node i. • The cost of maintaining the network: We assume that each node experiences a maintenance cost of π per link incident on it. Hence effective maintenance cost of each link is 2π and the maintenance cost payed by each player is Mi (G) = πdi (G) where di (G) denotes the degree of that node. 10

• The cost of incomplete connectivity in the network: We assume that each node wishes to send a unit traffic to all other nodes and additional cost of λ per unit of traffic not sent is imposed on the node if it is unreachable from some node. Hence the disconnection cost is Di (G) = λ(n − ni (G)) where n is the total number of players and ni (G) is the number of nodes reachable from node i.

Hence, the total cost of node i in graph G is defined as Ci (G) = Ri (G) + Mi (G) + Di (G) The difference in cost to node i between graph G and graph G + ij is defined as the extra overhead of cost to be payed by node i due to addition of the edge (i, j) in the graph G. More formally, ∆(G, ij) = Ci (G + ij) − Ci (G)

6.2.3

Payment and Total Utility

Two nodes participating to form a communication link can exchange utility by making payments. We refer to the payment function as a vector P = (Pij , i, jεV ) where Pij denotes the payment made to player j by player i. Hence from Pn the above definiton, the total payment received by player i is Ti (P ) = j=1 (Pij − Pji ) Now the total utility of player i is defined as difference of the payment received by player i with the cost incurred by player i. More formally it is defined as Ui (P, G) = Ti (P ) − Ci (G)

The objective of each player in the game is to selfishly try to maximize its total utility function.

11

6.2.4

Contract Formation and The Game

The game is a single shot game in which each player simultaneously decides the set Ti of nodes it wishes to connect to and the set Fi of nodes it is willing to receive connections from. A contract (i, j) is formed iff j Ti and i F j.

A payment is made between two nodes i and j if and only if they form a contract otherwise the payment is zero between them. Payment function Q depends only on the nodes i , j and the resultant graph. It has the following two properties: • Monotonicity: If a node i can form contract with two nodes i and j then the payment made by i to j is more than the payment made by i to k if and only if the contract formation increases the cost of j more than the cost of k. Formally ∆Cj (G, ij) > ∆Ck (G, ik) if and only if Q(i, j; G + ij) > Q(i, k; G + ik). Informally, monotonicity requires that the payment to form a link must increase as the burden of forming that link increases on the accepting node. • Anti Symmetry: An antisymmetric contracting function has the property that at any feasible outcome of the game, the payment for a link ij does not depend on which node asked for the connection. Regardless of whether the contract (i, j) or (j, i) is formed, the direction and quantity of payment across the link ij remains the same. More formally Q(i, j; G) = −Q(j, i; G). Any Anti Symmetric payment function has the property that the sum of payments received by all the nodes in the graph sums up to zero.

6.2.5

Definition of Stability and Efficiency of the Network Formation Game

A instance of a network formation game is “Stable” if and only if the following two conditions hold: 12

• No unilateral deviations are profitable, i.e. no player can increase its utility by changing its strategy while the strategy of other players remains the same. • No bilateral deviations are profitable i.e. no two players can collude to increase their payoffs. More precisely, no two players can increase their utilities by changing their strategies while the strategies of all the other players remains the same.

Since the payment function is anti symmetric so the net payment made by all the players sums up to zero. Hence the social cost of the game is just the sum of individual cost of all the players. Outcome of the network formation game is said to “Efficient” if it has minimium social cost over all possible outcomes of this game with the given parameters.

6.2.6

Stable and Efficient Instances of Network Formation Games

In the undirected case, it has been proved in [6] that all the stable outcomes of the games are forests. This can be intuitively observed from the fact that the payoff model does not include any redundant links. Any redundant link only acts as an extra overhead in the cost of the player. This is true in this model but not true in other generic models of the network based games because the distance between two nodes plays no role in the utility function of this model.

In this game we can set a value of λ large enough so that it is never beneficial for a node to be in disconnected graph. In this case all the stable outcomes of the games are trees. It has also been shown [6] that a star with the center node having minimum ci (constant in the cost function of node i) among all the nodes always acts a stable outcome. Hence, we can conclude that for any given set of parameters, we can always find a stable outcome of the game. It has also been proved further [6] that this star is also MC efficient outcome of the game which means that it achieves a lower social cost among all minimally connected graphs.

If we look into the stable and efficient models in the directed case, we find that all the stable outcomes of the undirected model acts as a subset to the directed model. In the directed model also all the stable graphs are minimally connected. Furthermore, 13

there is a simple characterization of the set of payments made in equilibrium which states that if a directed edge is constructed from node i to node j then the positive payment will only me made from node i to node because this edge only increases the cost of node j without affecting its connectivity.

7

Conclusion

We considered many basic models of Network based games. The results obtained are very generic and can be applied to a vast domain of practical applications. However, by slightly changing the cost structures or utility functions or the definition of equilibrium we can obtain specific model applicable to more specific domains. One such example is considering the bilateral contract fromation game where the λ or π values for each node differ or when the cost function also depends on the distance between the two nodes.

8

Acknowledgements

We sincerely thank Dr. Anil Seth for his hard work, constant guidance and useful discussions on the problem. We are also thankful to the Department of Computer Science and Engineering, IIT Kanpur for providing us the opportunity to undertake this project.

References [1] Tim Roughgarden, Eva Tardos Introduction to the Inefficiency of Equilibria, in Algorithmic Game Theory (Cambridge University Press, 2007) [2] Tim Roughgarden. Routing Games, in Algorithmic Game Theory (Cambridge University Press, 2007) [3] Eva Tardos and Tom Wexler Network Formation Games and the Potential Function Method, in Algorithmic Game Theory (Cambridge University Press, 2007)

14

[4] Alex Fabrikant, Ankut Luthra, Elitza Maneva, Christos H. Papadimitiou, Scott Shenker, On a Network Creation Game, In PODC’03 July 13-16, 2003 [5] Susanne Albers, Stefan Eilts, Eyal Even-Dar, Yishay Mansour, Liam Roditty On Nash Equilibria for a Network Creation Game, in Proceedings of the 17th annual ACM-SIAM symposium on Discrete algorithm, 2006, 0-89871-605-5 [6] Jacomo Corbo, David Parkes, The Price of Selfish Behavior in Bilateral Network Formation, In PODC’03 July 17-20, 2005 [7] Esteban Arcaute, Eric Dallal, Ramesh Johari, Shie MannorDynamics and Stability in Network Formation Games with Bilateral Contracts, in Proceedings of the 46th IEEE Conference on Decision and Control, 2007

15

Inefficiency of Nash Equilibrium in Network Based Games

Apr 19, 2010 - ish agents are represented by nodes in a network and edges serve as ... it intutivelty denotes the overall social benifit of all the players in the ...

116KB Sizes 4 Downloads 275 Views

Recommend Documents

Nash Equilibrium in Discontinuous Games
Phone: 773$7028192; Fax: 773$702$8490; email: [email protected]. 1This literature has ... behaved best$reply correspondences (e.g., Nash 1950, 1951; Glicksberg 1952). To this end, ... to verify in practice. 2 .... be optimizing, and therefore ,x is

Definitions of Equilibrium in Network Formation Games
Apr 16, 2006 - Different models have been proposed to analyze the formation of bilateral links in small societies ... Financial support from the Lee Center for Advanced .... that they define, which we simply call the linking game with transfers.

The Nash Equilibrium
Bi(a-i) = fai 2 Ai : ui(ai;a-i) ¸ ui(bi;a-i. ) ... Bi(a-i. ) is a set which may or may not be a singleton.) So, for example, in a ..... final order of business in this chapter.

The Nash Equilibrium
A Nash equilibrium of a game G in strategic form is defined as any outcome. (a*1; ::: ... Bi(a-i. ) is a set which may or may not be a singleton.) So, for example, in a ...

Equilibrium Selection, Inefficiency, and Instability in ...
ε (v, α) be the set of v ∈ V(v) such that Γε(α, v) has no stable Nash equilibrium. Theorem 3. For any ε ∈ (0, 1), α ∈ A, and v > 0, the. Lebesgue measure of Vus ε (v, α) is positive.4. Proof. Suppose Γε(α, v) has a π-stable equilib

NETWORK FORMATION GAMES BASED ON ...
networks where the goal of the agents is to strategically pro- duce, disseminate and consume information) have been for- malized and studied for the first time in Zhang and van der. Schaar [6]-[7]. In these works, agents are considered to be heteroge

Cognitively Stable Generalized Nash Equilibrium in ...
Oct 15, 2015 - Local actions and generalized action profile. In a static game with unawareness ΓU = (G,F):. Definition. For every i ∈ N and Gk ∈ Gi , σ(i,k) ∈ ∆Ak i is called i's local action in Gk, where: Gi = {Gl ∈ G| for some Gk ∈ G,

A Note on Uniqueness of Bayesian Nash Equilibrium ...
Aug 30, 2011 - errors are my own. email: [email protected], website: ... second and main step of the proof, is to show that the best response function is a weak contraction. ..... In et.al, A. B., editor, Applied stochastic control in econometrics.

On the Generic Robustness of Nash Equilibrium to ...
Oct 3, 2016 - ... X matter for this definition because it is the domain of Ψ. In fact, a quite ...... (1973a) purification theorem and its extensions (e.g. van Damme ...

Reformulation of Nash Equilibrium with an Application ...
The set of Nash equilibria, if it is nonempty, is identical to the set of minimizers of real-valued function. Connect equilibrium problem to optimization problem.

A Note on Uniqueness of Bayesian Nash Equilibrium ...
Aug 30, 2011 - errors are my own. email: [email protected], website: ... second and main step of the proof, is to show that the best response ..... Each country has an initial arms stock level of yn ∈ [0,ymax], which is privately known.

Implementing the Nash Program in Stochastic Games
Jun 23, 2013 - game without state variables, a stochastic game with contracts is, .... problem, from his static setting to a stationary, infinite-horizon environment.

Finding Multiple Nash Equilibria in Pool-Based Markets
companies (GENCOs) optimize their strategic bids anticipating the solution of the .... straints of the ISO problem come from the energy balance and the limits on the ... better alternative than solving the combinatorial game creating combinations ...

Learning in Network Games - Quantitative Economics
Apr 4, 2017 - arguably, most real-life interactions take place via social networks. In our .... 10Since 90% of participants request information about the network ...

Communication equilibrium payoffs in repeated games ...
Definition: A uniform equilibrium payoff of the repeated game is a strategy ...... Definition: for every pair of actions ai and bi of player i, write bi ≥ ai if: (i) ∀a−i ...

Learning in Network Games - Quantitative Economics
Apr 4, 2017 - solely on observed action choices lead us to accept certain learning rules .... arguably, most real-life interactions take place via social networks.

Correlated Equilibrium and Concave Games
May 1, 2007 - I acknowledge financial support by The Japan Economic Research ... payoff function with respect to the player's own strategy and call the.

Equilibrium in Discontinuous Games without Complete ...
Nov 17, 2015 - i G N\I, player iIs best-reply correspondence is closed and has non-empty and convex values, then G possesses a pure strategy Nash equilibrium. 2. An Application to Abstract Games. We demonstrate here how Theorem 1.3 can be applied to

INEQUALITY AND INEFFICIENCY IN JOINT PROJECTS*
Dec 8, 2004 - complemented by similar efforts elsewhere in the network for the ... perfect equality of shares maximises social surplus over all share vectors.

Research Articles Generic inefficiency of equilibria in ...
... of reaching date-event σt. Thus, agent h's state prices u (cs) are summable. ..... Of course, all equilibria will be Pareto efficient if markets are complete. Huang.

New Insights on Generalized Nash Games with Shared Constraints ...
New Insights on Generalized Nash Games with Shared Constraints: Constrained and Variational Equilibria. Ankur A. Kulkarni. Uday V. Shanbhag. Abstract—We .... in the community of operations research and computational game theory it has been a common

Revisiting Generalized Nash Games and Variational ...
Sep 15, 2010 - A shared constraint is a requirement that the tuple x be constrained to lie in a set. C ⊆ Rm. In the generalized Nash game with shared ...

network formation games
Another central question about network formation is whether the networks formed by the agents are ...... used to study the influence of the degree of farsightedness on network stability. Level-K farsighted stability is a tractable concept with myopic

Culture-based fisheries - Network of Aquaculture Centres in Asia-Pacific
in their regular agricultural activities, which is chiefly rice farming. ... kets and relying on scale and distribution of production rather than raw intensity and ...