Forty-Eighth Annual Allerton Conference Allerton House, UIUC, Illinois, USA September 29 - October 1, 2010 Nash equilibria for spectrum sharing of two bands among two players Ilaria Malanchini Steven Weber Matteo Cesana Drexel University, Dept. of ECE [email protected] Drexel University, Dept. of ECE [email protected] Politecnico di Milano, DEI [email protected] Abstract—The spectrum sharing game and the quality of its equilibria have been widely studied in a variety of contexts. In this paper we consider two pairs of communicating users that share two bands of spectrum. Through the analysis of the Nash equilibria, we provide the conditions, with respect to the normalized signal and interference strengths, for the set of equilibria power allocations to coincide with the set of optimal allocations. In contrast, when these sets do not coincide, we characterize the quality of the equilibria using the price of stability and the price of anarchy measures. In the more general case of N pairs of transmit receive pairs in an ad hoc network, we provide simulation results of a simple distributed player power allocation update heuristic that improves the sum rate utility above that achieved by the equilibrium of splitting the power evenly between the two bands. I. I NTRODUCTION Growth in demand for wireless spectrum in the last few years has shown that the current policies for use of available wireless spectrum are increasingly inadequate. New, more flexible spectrum management between users and operators is necessary to improve the efficiency of spectrum usage. This problem of how to share spectrum has been widely addressed in the technical literature, especially in the context of heterogeneous and cognitive networks. Many of the adopted models are based on game theory since, in general, the quality of service perceived by a system (e.g., user, network) strictly depends on the behavior of the other entities. Therefore, the interaction and the competition among multiple systems can be analyzed using game theoretic models. In particular, noncooperative game theory is suitable in distributed networks, where control and management are inherently decentralized. In this work we consider a Gaussian interference game (GIG) [1], where two transmitter and receiver pairs (each pair being a player) spread power across two bands of spectrum, with each player subject to a sum power constraint. Each player’s payoff function is the sum of the Shannon rate achieved on each band, where the Shannon rate is calculated assuming random Gaussian codebooks and where interference is treated as noise. Each pair of users makes selfish decisions regarding which portion of the spectrum to use. This corresponds to deciding how to split the total power budget for transmission among the two sub-channels. We model the problem as a non-cooperative game in which selfish users aim at maximizing the achieved throughput. Related work. The spectrum sharing game has been widely addressed in the literature. In [1], the authors study 978-1-4244-8216-0/10/$26.00 ©2010 IEEE the spectrum sharing game for multiple players capable of spreading their power budget across a shared portion of spectrum. The authors investigate and analyze the equilibria of the game, focusing on the issues of fairness and efficiency. They also propose punishment strategies that allow competing entities to reach a fair and efficient operating point. The game between users is analyzed also in [2]. The authors consider the power control problem with SINR as objective function, in both the selfish and the cooperative scenario. They characterize the equilibria and identify the conditions where the Pareto and the Nash equilibrium coincide. The spectrum sharing problem has also been analyzed from the network’s point of view. In [3], the authors analyze the spectrum competition between two contending networks. They characterize the Nash equilibria of the game and discuss the different behavior varying the pathloss exponent. In [4], a similar model is proposed to analyze the spectrum sharing problem between two networks operators sharing two carriers. Based on the water-filling algorithm proposed in [5], the authors prove the existence of the Nash equilibria of the game, and characterize the different behavior varying the channel gains between the two contending communicating pairs. In [6], two system pairs sharing two frequency bands are considered. In particular, the model comprises an interference channel in parallel with an interference relay channel. For two different relaying strategies, the existence and the uniqueness of the equilibrium are analyzed. A similar scenario is discussed also in [7], where the authors consider a network with both shared and protected bands, and power allocation is similarly based on the water-filling solution. A formal proof that the game is supermodular is derived as well as conditions for the existence and uniqueness of the Nash equilbria. Spectrum sharing problem has been discussed also in the context of Cognitive Radio Networks [8]. Cognitive networks are emerging as a solution for the spectrum efficiency problem, allowing unlicensed users to access vacant portions of the spectrum. In [9] and [10], the authors consider the problem of spectrum sharing among secondary users. In contrast with other work, the power control is analyzed with a constraint on the total interference perceived by the primary system. Namely, the interference is measured using the interference temperature. This is a parameter that measures the power and bandwidth occupied by the secondary interference. In both the papers, the authors identify the Nash 783 equilibria and analyze their properties. Also in [11], the authors analyze the spectrum sharing problem with multichannels, considering the co-channel interference among secondary users and the interference temperature regulation imposed by primary systems. Existence and properties of the Nash equilibria are investigated. In [12], the authors model the spectrum competition among networks that have to decide both the channel and the power allocation. The Nash equilibria of the non-cooperative game are characterized and a cooperative technique is proposed in order to improve the opportunistic solution. In [13], the channel competition is played among selfish users, instead of networks. The authors characterized the equilibria and discuss the efficiency with respect to the optimal channel allocation. In [14], a spectrum selection game is proposed. Different from previous work, competitive users take decisions on which channels to use, not only depending on the number of other users that are sharing the same channel, but also taking into account the different parameters that characterize the available spectrum opportunity. Furthermore, a multi-stage game is proposed for the dynamic spectrum management. Contributions. We consider first a symmetric scenario where the received normalized interference powers (X) are equal and the received normalized signal powers (D) are equal, so that the problem has only these two parameters. By this parameter reduction we are able to explicitly characterize allocations that are socially optimal as well as all Nash equilibria, and thereby explicitly compute the Price of Stability (PoS) and Price of Anarchy (PoA) in terms of (X, D). We then consider an asymmetric scenario where the interference powers are unequal (X1 6= X2 ), and we characterize the socially optimal allocations and the Nash equilibria. Again, we are able to study the PoS and PoA, now as a function of the three key parameters (X1 , X2 , D). In the case of N players, we use our observation on the sum-rate optimal allocations in the 2-player asymmetric game to motivate a distributed player power allocation update heuristic that improves the sum rate utility above that achieved by the full-spread equilibrium. Our simulation results demonstrate that for networks with low spatial density a fixed full-spread equilibrium is socially optimal, but the gap between the fullspread equilibrium and the social optimum increases in the spatial density. This heuristic manages to achieve a sum utility above that of the full-spread equilibrium but is at a significant gap from the maximum sum utility. The paper is organized as follows. In §II, we describe the scenario and the game model for the 2-player case. §III and §IV analyze the 2-player and the N -player game, respectively. §V concludes the paper, discussing some open issues. II. S YSTEM M ODEL In this section, we describe the system and game theoretic model for the 2-player case that is analyzed in §III. The spectrum band is assumed to be divided into two orthogonal channels with equal bandwidths B. Players are indexed as 1, 2 and we let i ∈ {1, 2} indicate a player and j 6= i as the R1 D X1 Player 1 (P1 , P̄1 ) T1 U1 (P1 , P2 ) T2 D X2 Player 2 (P2 , P̄2 ) U2 (P1 , P2 ) R2 Fig. 1. Two transmitter and receiver pairs (players) split their power over two orthogonal bands. other player. The distance between the generic receiver i and the contending transmitter j is denoted by xi , whereas di is the distance between transmitter i and receiver i. To simplify the problem and reduce the number of model parameters we assume throughout this paper that d1 = d2 = d. Each pair has to decide how to split its total transmission power P between the two available bands, taking into account the behavior of the contending pair. Moreover, given a certain power budget P available for each one of the two pairs, the strategy space of the generic transmitter i is Pi ∈ [0, 1]. Namely, Pi is the fraction of P that pair i uses in the left band and P̄i = 1−Pi is the fraction of power in the right band. The utility (payoff) function of player i is defined as the sum achievable Shannon rate over the two bands when the interference from player j is treated as noise. We assume a channel model with pure pathloss attenuation so that the received power from a unit power transmission over a distance r is r−α , for α > 2 the pathloss exponent. Assuming a noise power of η̃ on each band, the player i utility function for transmission powers (P̃i , P̃j ) ∈ [0, P ]2 is ! ! P̃i d−α P̃¯i d−α Ũi = B log2 1 + +B log2 1 + η̃ + P̃j x−α η̃ + P̃¯j x−α i i (1) in units of bits per second. To summarize, under the above model assumptions, the spectrum sharing problem can be modeled as a non-cooperative game, where the players are the two transmitter and receiver pairs, the actions are the possible power splits over the two bands, and the payoffs are the sum of the achievable Shannon rates over the two bands. We remove nuissance parameters B, P, η, α from the model as follows. First note the payoff is linear in B and hence we define the normalized utility Ui = Ũi /B with units of spectral efficiency (bits per second per Hertz); this removes B. Multiply the numerator and denominator in the SINR in each band by 1/P and define the normalized transmission power as Pi = P̃i /P ∈ [0, 1], and the normalized noise power as η = η̃/P ; this removes P . Define the received signal to noise ratio (SNR) under maximum transmission power on a band as D = d−α /η, and the received interference to noise ratio (INR) under maximum transmission power on a band as Xi = x−α i /η; this removes η, α. The utility function becomes: Pi Di P̄i Di Ui = log2 1 + + log2 1 + . (2) 1 + Pj Xi 1 + P̄j Xi 784 Each one of the two players is assumed to be selfish. This means that each transmitter allocates power between the two bands trying to maximize the player’s achieved throughput (i.e., without taking into account the global optimum). Therefore, the stable operating points for the two players are the Nash equilibria. A Nash equilibrium is a pair (P1 , P2 ) ∈ [0, 1]2 from which neither player has incentive to unilaterally deviate. Since, in general, the Nash equilibrium reflects the selfish behavior of the players, it is often inefficient from the system point of view. Therefore, it is useful to compare the Nash equilibria of a game with the “globally” optimal solution, i.e., the one that could be achieved with a centralized control. Usually, this comparison is done in terms of social utility, e.g., the sum of the utility of all the players, UT = U1 + U2 . In particular, the social/global optimal sum utility is UT∗ = max (P1 ,P2 )∈[0,1]2 UT (P1 , P2 ). (3) This quantity exists by virtue of the fact that UT is continuous and bounded on [0, 1]2 . The “quality” of an equilibrium can be assessed using the concepts of Price of Stability (PoS) [15] and Price of Anarchy (PoA) [16]. They are, respectively, the ratio between the optimal solution that could be achieved by players in a centralized system and the best/worst Nash equilibrium. Namely, for NE ⊆ [0, 1]2 the set of Nash equilibria, we have: PoS = PoA = UT∗ max(P1 ,P2 )∈NE UT (P1 , P2 ) UT∗ . min(P1 ,P2 )∈NE UT (P1 , P2 ) (4) Clearly 1 ≤ PoS ≤ PoA where PoS = 1 when the social optimal allocation is a Nash equilibrium, and PoS = PoA when the Nash equilibrium is unique (or at least multiple equilibria all have common sum utility). Combining these observations, we have PoS = PoA = 1 when the unique Nash equilibrium coincides with the social optimal allocation. The aim of this analysis is to characterize the Nash equilibria of the game and assess their quality with respect to the optimal solution using the PoS and PoA. In particular, we characterize the PoS and PoA as a function of the three key parameters (D, X1 , X2 ). a game has selected the best response (or one of the best responses) to the other players’ strategies. To evaluate the best response function of the generic player i, we consider the partial derivative with respect to Pi (5) of the utility function (2). Therefore, the best response of player i is, by definition, the maximum of the utility function, over all possible strategies of player j. In other words, the best response function is given by the points in which the partial derivative is equal to zero. The above function is equal to zero when the numerator is null, that is: 1 X 1 Pi = + − Pj . (6) 2 D 2 For the sake of brevity, we do not report here the analysis that shows that these points are maxima of the function (and not minima). Since the domain of the function is [0,1], i.e., (Pi , Pj ) ∈ [0, 1]2 , the best response function cannot be outside this range. Therefore, we consider the following cases: 1 X 1 1 D Pi = + − Pj < 0 ⇐⇒ Pj > 1+ (7) 2 D 2 2 X 1 X 1 1 D Pi = + − Pj > 1 ⇐⇒ Pj < 1− (8) 2 D 2 2 X This implies that whenever Pi is less than 0 (or greater than 1) the best response is fixed to 0 (or 1). Finally, the best response for player i, Pi∗ (Pj ), is the following function: D 1 if X > D and 0 ≤ Pj <12 1 − X 1 D Pi∗ (Pj ) = 0 if X > D and 2 1 + X < Pj ≤ 1 1 X 1 else 2 + D 2 − Pj (9) These best response functions are illustrated in Fig. 2. Their intersection give the Nash equilibria, as stated in the following theorem. P1�P2� P1�P2� X<D 1.0 P2∗ (P1 ) 0.8 � 0.6 1 1 , 2 2 A. The symmetric case For this subsection we assume X1 = X2 = X. Generally speaking, the Nash equilibria of a game can be found using the best response functions. The best response is the strategy (or strategies) which produces the most favorable outcome for a player, taking other players’ strategies as given. Therefore, the Nash equilibrium is the point at which each player in � 2 1 2 0.2 � 1− 0.4 � X D 0.6� 0.8� 1.0 1 2 1+ 1+ X D � P1∗ (P2 ) 0.8 1 2 0.6 P1∗ (P2 ) � 1 0.2 This section analzes the quality of the equilibria in the 2-player game. In the first part we consider the symmetric case, i.e., when X1 = X2 = X, and in the second part we extend the analysis to the more general asymmetric case with X1 6= X2 . 1 2 � 0.4 III. T HE 2-P LAYER G AME X>D 1.0 X D � 0.4 X 1− D � 0.2 P2�P1� 1 2 1 1 , 2 2 �0.2 �0.4 D 1− X � P2∗ (P1 ) 1 2 0.6� 1 2 1+ 0.8� 1.0 � 1+ D X � � 1− D X � P2�P1� D X Fig. 2. The intersection of the best response functions P1∗ (P2 ), P2∗ (P1 ) give the Nash equilibria. The set of equilibria depends upon X ≶ D. Theorem 3.1: The 2-player symmetric game admits the following Nash equilibria: if X < D {(0.5, 0.5)} {(Pi , Pj ) : Pi + Pj = 1} if X = D NE = (10) {(0, 1), (0.5, 0.5), (1, 0)} if X > D Proof: A pure strategy Nash equilibrium is a point at which each player has selected the best response to the other 785 ∂Ui [(1 − 2Pj )Xi + (1 − 2Pi )D]D log2 e = ∂Pi (1 + Pj Xi + Pi D)(1 + (1 − Pj )Xi + (1 − Pi )D) players’ strategies. In this game, the point in which the two best responses cross is the solution of the following system: (1 − 2Pj )X + (1 − 2Pi )D = 0 (11) (1 − 2Pi )X + (1 − 2Pj )D = 0 that is Pi = Pj = 0.5. This is the unique equilibrium when X < D. However, when X > D the best responses cross in other two points, i.e., (0, 1) and (1, 0). When X = D the two best responses coincide with the line Pi + Pj = 1, i.e., there is an infinite number of equilibria. We assess the quality of the Nash equilibria via the PoS and PoA given in (4). Therefore, we consider the sum utility function UT (P1 , P2 ) in (13) with X1 = X2 = X. The globally optimal allocations are the maximizers of UT : Opt = arg max UT (P1 , P2 ) (12) (P1 ,P2 )∈[0,1]2 Claim 3.1: The set of optimal power allocations for the 2-player symmetric game is: if X < D̃ {(0.5, 0.5)} Opt = (14) {(Pi , Pj ) : Pi + Pj = 1} if X = D̃ {(0, 1), (1, 0)} if X > D̃ √ where D̃ = 1 + D − 1. Moreover, the function UT is concave for X > D̃, and has a saddle point at P = (0.5, 0.5) for X < D̃. Rationale: The function we want to maximize is a continuous twice differentiable function defined over the set C = [0, 1]2 . The necessary condition for a point (Pi , Pj ) to be a stationary point is the gradient to be zero. The two components of the gradient are given by (18) with Xi = Xj = X. We are looking for a pair (Pi , Pj ) that satisfy ∂UT (P1 , P2 )/∂P1 = 0 and ∂UT (P1 , P2 )/∂P2 = 0. For any point P = (P1 , P2 ) ∈ C define the projection Q(P) onto the line L = {(P1 , P2 ) : P1 + P2 = 1} as Q1 (P) = 0.5(1 + P1 − P2 ) and Q2 (P) = 0.5(1 + P2 − P1 ). We claim without proof that UT (Q(P)) ≥ UT (P) for all P ∈ C. With this claim it follows that max UT (P) = max UT (P) = max UT (P, P̄ ). P∈C P∈L P ∈[0,1] 7.38 7.38 7.375 7.379 7.37 7.365 7.378 7.36 7.377 7.355 7.376 7.35 7.375 7.345 7.374 1 7.34 1 0.5 0 P2 0 0.2 0.6 0.8 0.5 1 0 P2 P1 0 0.2 0.4 0.6 0.8 1 P1 The two previous theorems highlight how the equilibria for the two players depend upon (X, D). In particular, the optimal allocation is for both players to spread their power equally over the two bands when the interference is small relative to the signal (X < D̃), while the optimal allocation is to perform frequency division and put all power in (complementary) bands when the interference is large relative to the signal (X > D̃). Fig. 3 illustrates the function UT for X > D̃ (left) and X < D̃ (right). Finally, we can assess the quality of the Nash equilibria with respect to the optimal solution. To do this, we provide the analytical expression of the PoS and the PoA. Theorem 3.2: The Price of Stability (PoS) for the symmetric game is: ( 1 if X ≤ D̃ or X > D log2 (1+D) PoS = (19) if D̃ < X ≤ D 0.5D 2 log2 (1+ 1+0.5X ) and it has a maximum that is unbounded in D, i.e., limD→∞ maxX≥0 PoS(X, D) = ∞. Proof: The price of stability is given by (4). It is easy to see that D̃ ≤ D. Therefore there are two cases. When X ≤ D̃ or X > D the best equilibrium coincides with the global optimum, i.e., PoS = 1. Otherwise, PoS > 1. Namely, it is possible to see that the limX→D̃+ PoS = 1. Since: log2 (1 + D) =1 0.5D 2 log2 1 + 1+0.5 D̃ (20) it follows that the PoS is continuous at X = D̃. In contrast, the function is discontinuous at X = D. This point is also the maximum of the PoS, since the function is monotonically increasing in the interval (D̃, D). The upper bound of the PoS (as X → D− ) is: lim PoS = X→D − and it is easy to see that: lim log2 (1 + D) 2 log2 2(1+D) 2+D lim PoS = 1, lim D→0 X→D − (17) 0.4 Fig. 3. The sum utility function UT (P1 , P2 ) has a saddle at (0.5, 0.5) (left) for X > D̃ but is concave for X < D̃ (right). It follows that UT (0.5, 0.5) ≷ UT (0, 1) ⇔ X ≷ D̃. 7.381 7.385 (15) i.e., all global maximizers are in L. We now have a onedimensional optimization problem over P ∈ [0, 1]. We leave the fact that UT is concave for X < D̃ and has a saddle at (0.5, 0.5) for X > D̃ unproven. To find the threshold for which UT changes its shape and its set of maximizers, we evaluate UT at the key points P = (0.5, 0.5) and P = (0, 1) (equivalently, P = (1, 0)): 0.5D UT (0.5, 0.5) = 4 log2 1 + 1 + 0.5X UT (0, 1) = UT (1, 0) = 2 log2 (1 + D) (16) (5) lim PoS = ∞ D→∞ X→D + Therefore, the PoS is unbounded when D → ∞. 786 (21) (22) UT = log2 1 + Pi D 1 + Pj Xi + log2 1 + P̄i D 1 + P̄j Xi ∂UT (1 − 2Pj )Xi + (1 − 2Pi )D = D log2 e + ∂Pi (1 + Pj Xi + Pi D)(1 + P̄j Xi + P̄i D) + + log2 1 + P1�P2� 1.0 + log2 1 + P1�P2� X1 ≥ D, X2 ≥ D 1.0 (1, 0) 1.3 log2 (1 + D) � � 1 D 2 log 1 + 1+2 1 X 1.2 1.0 √ 0.5 1+D−1 {(0.5, 0.5)} P1 (P2 ) 0.8 P oA 0.6 2 1.1 P oS 1.0 0.4 1.5 X P2 (P1 ) {(0.5, 0.5)} 0.8 (0, 1) (2, 4, 1) 0.2 {(0, 1), (1, 0)} P1�P2� {(0.5, 0.5), (0, 1), (1, 0)} 0.4 0.6 0.8 X1 X2 = D P1 (P2 ) 1.0 lim log2 (1 + D) =∞ 0.5D 2 log2 1 + 1+0.5X (24) Therefore, the limX→∞ PoA = ∞. Fig. 4 shows PoS√and the PoA as a function of X for D = 1 (for which D̃ = 2 − 1 ≈ 0.414). Both curves are plotted to highlight the interval in which they coincide. Note that the PoS is in general bounded, and goes to infinity only when D goes to infinity. In contrast the PoA goes to infinity whenever X goes to infinity, i.e., for every D. The conclusion is that in general the PoS is very close to one, i.e., the best equilibrium is very close to the optimum, but at the same time, when X is large, e.g., in dense networks, the worst equilibrium could be “infinitely” worse than the optimal solution. A similar result is also given in [1], where the authors show that the inefficiency resulting from choosing to spread power equally among available bands can be arbitrarily large. B. The asymmetric case In this section, we consider the case in which Xi 6= Xj . In order to characterize the Nash equilibria of the game, we consider the partial derivative of the utility function of player i, reported in (5). The best response of each player is similar to the previous case, since the function of player i depends only on Xi . Namely, the best response for player i, (β2+ , 0) (2, 3/4, 1) 0.2 P1�P2� 0.4 0.6 0.8 X1 X2 ≤ D 1.0 P2�P1� 2 1.0 P2 (P1 ) 0.8 0.6 P2 (P1 ) (0.5, 0.5) (0.5, 0.5) 0.4 0.2 and it is unbounded, i.e., limX→∞ PoA = ∞. Proof: The PoA is given by (4). It is easy to see that (18) P2 (P1 ) 2 0.4 P1 (P2 ) 0.2 (2, 1/4, 1) (2, 1/2, 1) Theorem 3.3: The Price of Anarchy (PoA) for the symmetric game is: ( 1 if X ≤ D̃ log2 (1+D) PoA = (23) if X > D̃ 0.5D 2 log2 (1+ 1+0.5X ) (0.5, 0.5) P2�P1� 0.8 0.6 (13) X1 ≤ D, X1 X2 ≥ D2 (β2− , 1) P1 (P2 ) 0.4 0.2 D PoS and PoA for the 2-player symmetric case for D = 1. X→∞ P̄j D 1 + P̄i Xj 0.6 (0.5, 0.5) 0.2 1.0 P̂ Fig. 4. −Xj Pj Xj P̄j + (1 + Pi Xj )(1 + Pi Xj + Pj D) (1 + P̄i Xj )(1 + P̄i Xj + P̄j D) PoS,PoA 1.4 P∗ Pj D 1 + Pi Xj 0.2 Fig. 5. 0.4 0.6 0.8 1.0 P2�P1� 0.2 0.4 0.6 0.8 1.0 P2�P1� Best response functions for the 2-player asymmetric game. i.e., Pi∗ (Pj ), is the following function: D 1 if Xi > D and 0 ≤ Pj < 12 1 − X i ∗ 1 D Pi (Pj ) = 0 if X > D and 1 + < P ≤1 i j Xi 2 1 Xi 1 + − P else j 2 D 2 (25) The Nash equilibria are the intersections of the two best response functions P1∗ (P2 ), P2∗ (P1 ), illustrated in Fig. 5. The following theorem gives the Nash equilibria of the asymmetric game. Theorem 3.4: The 2-player asymmetric game has Nash equilibria given by (26). Proof: As in the previous case, the pure strategy Nash equilibrium is the point at which each player has selected the best response to the other players’ strategies. The two best responses always cross at (0.5, 0.5). This is the unique equilibrium when Xi Xj < D2 . However, when Xi Xj > D2 , the best responses cross in other two points. Namely, when both Xi and Xj are greater than D, these two points are (0, 1) and (1, 0). In contrast, when Xi ≤ D (and, consequently, i Xj > D), these points are Pi = 0.5 1 ± X D , with Pj = 0 and Pj = 1, respectively. When Xi Xj = D2 the two best responses coincide and there is an infinite number of equilibria. To assess the quality of the Nash equilibria, we characterize the optimal solution of the asymmetric game. Claim 3.2: The optimal solution for the asymmetric game is given by (27). Moreover, the function UT is concave for 787 {(0.5, 0.5)} {(P P ) : P = 0.5 + D (0.5 − P )} i j Xi i, j NE = i (0.5, 0.5), 0.5 1 + X , 0 , 0.5 1 − D {(0.5, 0.5), (0, 1), (1, 0)} Xi + Xj + Xi Xj < D, and has a saddle point at P = (0.5, 0.5) for Xi + Xj + Xi Xj > D. Rationale: The function we want to maximize and its partial derivative are reported in (13) and (18), respectively. As in the previous case, it is possible to show that the only point in which the gradient is zero is (0.5, 0.5). Therefore the maximum of the function can be attained at (0.5, 0.5) or at (0, 1) and (1, 0). We do not provide a formal proof of this claim. The threshold for which the total utility function changes its maximum (varying X1 and X2 ) can be found comparing its value at these two points. We obtain that: ,1 Xi Xj < D2 Xi Xj = D2 Xi ≤ D, Xi Xj > D2 Xi > D, Xj > D (26) X2 2 1.5 NE: (0.5,0.5), (0,1), (1,0) OPT: (0,1), (1,0) PoS=1 and PoA≥ 1 NE: (0.5,0.5), (β+2,0),(β−2,1) OPT: (0,1), (1,0) PoA ≥ PoS ≥ 1 (D,D) 1 X1X2=D2 NE: (0.5,0.5) OPT: (0,1), (1,0) PoS = PoA ≥ 1 0.5 UT (0.5, 0.5) ≷ UT (0, 1) ⇔ Xi + Xj + Xi Xj ≶ D. (28) The behavior of the Nash equilibria and the optimal solution (varying X1and X2 ) is highlighted in Fig. 6, where i βi± = 0.5 1 ± X D . Note that there exist four regions. 2 When Xi Xj > D , the game admits three equilibria and the optimum is in (0, 1) and (1, 0). In particular, when both Xi and Xj are greater than D, the PoS is one, since the best equilibrium and the optimum coincide. In contrast, when Xi (or Xj ) is less than D, the best equilibrium is worse than the optimum, then the PoS is greater than one. The PoA is in both the two cases greater than one. We do not report the expressions, that can be easily derived as done for the symmetric case. When Xi Xj < D2 and Xi Xj +Xi +Xj > D, the game admits a unique equilibrium, that does not coincide with the optimum, then PoS and PoA coincide and are greater than one. In contrast, below the curve given by (28), the optimal solution and the unique equilibrium coincide. Similar conditions on the uniqueness of the equilibrium are discussed also in [5] and [7]. Finally, along the curve Xi Xj = D2 the two best responses coincide and there exists an infinite number of equilibria. Note that the corresponding regions of the symmetric case, i.e., [0, D̃], [D̃, D], [D, ∞), can be identified along the line X1 = X2 . Xi D if if if if ~ ~ (D,D) NE:+(0.5,0.5), (0,β1),(1, β−1,) X1+X2+X1X2=D NE: (0.5,0.5) OPT: (0.5,0.5) PoS=PoA=1 0 0 0.5 OPT: (0,1), (1,0) PoA ≥ PoS ≥ 1 1 1.5 2 X1 D Fig. 6. Nash equilibria and optimal allocations as a function of X1 , X2 (for D = 1). Theorem 4.1: The strategy profile Pi = 0.5, ∀ i ∈ N is always a Nash equilibrium of the N -player game. Proof: The best response function Pi∗ (P−i ) of player i is obtained by solving ∂Ui (P)/∂Pi = 0 for Pi , where ∂Ui (P)/∂Pi is given by (32): X (1 − 2Pj )Xij + (1 − 2Pi )D = 0. (30) j6=i Projecting Pi∗ onto [0, 1] gives: 1 X Xij 1 1 Pi∗ (P−i ) = + − Pj . 2 D 2 j6=i (31) 0 Therefore, we have a system with N equalities and in N variables. It is straightforward to see that Pi = 0.5, ∀i ∈ N is always a solution of the previous system. This proves the IV. T HE N -P LAYER G AME theorem. We show why this solution is not guaranteed to be the only In this section, we consider the N -player game, i.e., there equilibrium of the game. First, we have to consider the case are N pairs of users competing for the same spectrum that is divided into two orthogonal sub-channels. Let N be the in which the N equalities are linearly dependent. In this case set of players. The utility function of the generic player i is we have a hyperplane of dimension N −1 on which there is an infinite number of solutions (Nash equilibria). Note that this the following: ! ! is consistent with the 2-player case. When X1 X2 = D2 the PD P̄ D system (30) is composed by two linearly dependent equalities P i P i +log2 1+ Ui = log2 1+ 1+ j6=i Pj Xij 1+ j6=i P̄j Xij and there is an infinite number of Nash equilibria. Therefore, (29) (31) is always a solution but it is not guaranteed to be the where Xij = x−α /η is the interference to noise ratio from only one. In [1], the authors provide a sufficient condition for ij transmitter j to receiver i when full power is used on a band, this solution to be the unique equilibrium. Furthermore, the and xij is the distance between transmitter j and receiver i. characterization of the best response for the N -player game We prove the following theorem, which is found in [1]. is discussed also in [4], [7]. 788 if Xi + Xj + Xi Xj > D {(0, 1), (1, 0)} {(Pi , Pj ) : Pi + Pj = 1} if Xi + Xj + Xi Xj = D Opt = {(0.5, 0.5)} if Xi + Xj + Xi Xj < D hP i j6=i (1 − 2Pj )Xij + (1 − 2Pi )D D log2 e ∂Ui P P = ∂Pi (1 + j6=i Pj Xij + Pi D)(1 + j6=i (1 − Pj )Xij + (1 − Pi )D) (27) (32) Since we are unable to characterize either all possible Nash equilibria or the optimal solution, we provide a numerical evaluation. The goal is to compare, through simulations, the optimal solution with the Nash equilibrium in which every player plays the strategy Pi = 0.5. Finally, in order to improve the Nash equilibria of the N player game, we propose a heuristic algorithm that starts from the Nash equilibrium in which all users play the strategy Pi = 0.5 and aims at improving the efficiency of the allocation. Namely, from the results of the 2-player game, we have found that whenever two pairs of users are such that Xij and Xji satisfy: Xij + Xji + Xij Xji > D (33) they should use different halves of the spectrum, instead of the whole band. Therefore, we can motivate the users to implement local moves towards the optimum, reducing the interference perceived from their neighbors. The goal of the heuristic is to approximate the global optimal solution only with local changes. Future work will investigate distributed protocols and mechanisms based on local moves that could be adopted by the users in order to improve their throughput. The heuristic is illustrated in Algorithm 1. Algorithm 1 Heuristic algorithm 1: Initialize Pi = 0.5 ∀ i = 1...N 2: while ∃(i, j)|Xij + Xji + Xij Xji > D and Pi = Pj = 0.5 do (0, 1), w.p. 0.5 3: Set (Pi , Pj ) = (1, 0), w.p. 0.5 4: end while We have implemented an instance generator able to create instances representing the network scenario. The generator takes as input the following parameters: the number of pairs N , the distance d between each transmitter and receiver and the edge L of the square arena where the N pairs are randomly deployed. Once all the pairs are deployed over the arena, the INRs Xij between each transmitter j and each receiver i are evaluated by the generator. The optimal solution has been obtained writing the optimization problem, i.e., (12)), in AMPL [17] and solving it with the non-linear solver MINOS [18]. Numerical results are discussed in the following. Namely, we consider an arena with L = 200 and d = 25. The parameters for the evaluation of the total utility are η = 10−3 , α = 4 and B = 1. An example of the considered topology is reported in Fig. 7. Hereafter, all the reported results are averaged values on 30 randomly generated instances. Fig. 7. Sample of the considered topology when N = 10 In Fig. 8, we report the total utility achieved by the users as a function of increasing N . When the number of users is small (say 2, . . . , 10) the optimum and the Nash equilibrium are very close. This is consistent with the theoretical results obtained for the 2-player game. In fact, when the number of pairs is small, they are far enough apart (i.e., the average Xij is small enough) and the optimal solution coincides with the equilibrium, i.e., each player splits power equally over the two bands. As the spatial density of the network increases, the average interference increases, and the equal split equilibrium achieves an increasingly poor sum utility relative to the optimal utility. The heuristic is able to improve the sum utility above that achieved by the equal split equilibrium, but there is still a significant gap between the heuristic sum utility and the optimal. V. O PEN I SSUES In this paper, we have considered a scenario composed by multiple pairs of users sharing two bands of the spectrum. We have modeled the problem as a non-cooperative game and we have analytically characterized the Nash equilibria. Furthermore we have provided measures of their quality with respect to the optimal solution, both for the 2-player and the N -player case. We conclude this paper briefly discussing two open issues that are currently part of the ongoing work. A. The optimal solution of the asymmetric 2-player game As reported in §III-B, the optimal solution of the 2-player game is based only on a conjecture, and we do not provide a formal proof. However, we aim at providing a formal proof, 789 Fig. 8. Top: total utility (optimal, full-spread equilibrium, heuristic) for the N -player game as a function of N . Bottom: price of anarchy of the full-spread equilibrium. based on the following idea. As done for the asymmetric case, we want to show that the point (0.5, 0.5) is the only stationary point of the interior of the domain set, therefore the maximum should be in (0.5, 0.5) or at the boundary of the domain. Our result is supported also by the Karush-KuhnTucker conditions, that show that the points (0, 1) and (1, 0) are candidate to be maxima of the utility function only when Xi + Xj + Xi Xj > D. [5] W. Yu, G. Ginis, and J. M. Cioffi, “Distributed multiuser power control for digital subscriber lines,” IEEE Journal on Selected Areas in Communications, vol. 20, no. 5, pp. 1105–1115, 2002. [6] E. V. Belmega, B. Djeumou, and S. Lasaulce, “Resource allocation games in interference relay channels,” in GameNets, Istanbul, Turkey, 2009, pp. 575–584. [7] R. Mochaourab and E. Jorswieck, “Resource allocation in protected and shared bands: uniqueness and efficiency of Nash equilibria,” in GAMECOMM, Pisa, Italy, 2009, pp. 1–10. [8] S. Haykin, “Cognitive radio: brain-empowered wireless communications,” IEEE Journal on Selected Areas in Communications, vol. 23, no. 2, pp. 201–220, 2005. [9] J. Jia and Q. Zhang, “A non-cooperative power control game for secondary spectrum sharing,” in IEEE International Conference on Communications. ICC, 2007, pp. 5933–5938. [10] R. Di Taranto, H. Yomo, and P. Popovski, “Two players noncooperative iterative power control for spectrum sharing,” in IEEE PIMRC, 2008, pp. 1–5. [11] Y. Wu and D. H. K. Tsang, “Distributed power allocation algorithm for spectrum sharing cognitive radio networks with QoS guarantee,” in IEEE INFOCOM, 2009, pp. 981–989. [12] G. Hosseinabadi, M. H. Manshaei, and J.-P. Hubaux, “Spectrum sharing games of infrastructure-based cognitive radio networks,” Technical Report LCA-REPORT-2008-027, 2008. [13] M. Fèlègyhzi, M. Čagalj, and J. Hubaux, “Efficient MAC in cognitive radio systems: a game-theoretic approach,” IEEE Transactions on Wireless Communications, vol. 8, no. 4, pp. 1984–1995, 2009. [14] I. Malanchini, M. Cesana, and N. Gatti, “On spectrum selection games in cognitive radio networks,” in IEEE Global Telecommunications Conference. GLOBECOM, 2009, pp. 1–7. [15] E. Anshelevich, A. Dasgupta, J. Kleinberg, E. Tardos, T. Wexler, and T. Roughgarden, “The price of stability for network design with fair cost allocation,” in IEEE FOCS, 2004, pp. 59–73. [16] E. Koutsoupias and C. Papadimitriou, “Worst-case equilibria,” in 16th Annual Symposium Theoretical Aspects Computer Science. STACS. Trier, Germany: Springer, 1999, pp. 404–413. [17] R. Fourer, D. M. Gay, and B. W. Kernighan, AMPL, A modeling language for mathematical programming. Duxbury Press, 2002. [18] B. Murtagh and M. Saunders, “MINOS 5.5 user’s guide,” Stanford University Systems Optimization Laboratory Technical Report SOL8320R, 1998. B. The equilibria of the N -player game Concerning the N -player scenario, our analysis aims at fully characterize the Nash equilibria of the game. We have shown that the equilibrium in which all players choose Pi = 0.5 is not the only one. We would like to prove conditions under which this equilibrium is the only, the best, or the worst among all the equilibria. R EFERENCES [1] R. Etkin, A. Parekh, and D. Tse, “Spectrum sharing for unlicensed bands,” IEEE Journal on Selected Areas in Communications, vol. 25, no. 3, pp. 517–528, 2007. [2] E. Altman, K. Avrachenkov, and A. Garnaev, “Transmission power control game with SINR as objective function,” in Network Control and Optimization. Springer Berlin, 2009, vol. 5425, pp. 112–120. [3] L. Grokop and D. Tse, “Spectrum sharing between wireless networks,” in IEEE INFOCOM, 2008, pp. 201–205. [4] M. Bennis, M. Le Treust, S. Lasaulce, M. Debbah, and J. Lilleberg, “Spectrum sharing games on the interference channel,” in International Conference on Game Theory for Networks. GameNets, Istanbul, Turkey, 2009, pp. 515–522. 790
© Copyright 2024 Paperzz