Congestion Games with Player-Specific Constants

Congestion Games with Player-Specific Constants⋆ Marios Mavronicolas1 , Igal Milchtaich2 , Burkhard Monien3, and Karsten Tiemann3,⋆⋆ 1 3 Department of Computer Science, University of Cyprus, 1678 Nicosia, Cyprus mavronic@cs.ucy.ac.cy 2 Department of Economics, Bar-Ilan University, Ramat Gan 52900, Israel milchti@mail.biu.ac.il Faculty of Computer Science, Electrical Engineering, and Mathematics, University of Paderborn, 33102 Paderborn, Germany {bm,tiemann}@uni-paderborn.de Abstract. We consider a special case of weighted congestion games with playerspecific latency functions where each player uses for each particular resource a fixed (non-decreasing) delay function together with a player-specific constant. For each particular resource, the resource-specific delay function and the playerspecific constant (for that resource) are composed by means of a group operation (such as addition or multiplication) into a player-specific latency function. We assume that the underlying group is a totally ordered abelian group. In this way, we obtain the class of weighted congestion games with player-specific constants; we observe that this class is contained in the new intuitive class of dominance weighted congestion games. We obtain the following results: Games on parallel links: – Every unweighted congestion game has a generalized ordinal potential. – There is a weighted congestion game with 3 players on 3 parallel links that does not have the Finite Best-Improvement Property. – There is a particular best-improvement cycle for general weighted congestion games with player-specific latency functions and 3 players whose outlaw implies the existence of a pure Nash equilibrium. This cycle is indeed outlawed for dominance weighted congestion games with 3 players – and hence for weighted congestion games with player-specific constants and 3 players. Network congestion games: For unweighted symmetric network congestion games with player-specific additive constants, it is PLS-complete to find a pure Nash equilibrium. Arbitrary (non-network) congestion games: Every weighted congestion game with linear delay functions and player-specific additive constants has a weighted potential. ⋆ This work was partially supported by the IST Program of the European Union under contract number IST-15964 (AEOLUS). ⋆⋆ Supported by the International Graduate School of Dynamic Intelligent Systems (University of Paderborn, Germany). L. Kučera and A. Kučera (Eds.): MFCS 2007, LNCS 4708, pp. 633–644, 2007. c Springer-Verlag Berlin Heidelberg 2007  634 M. Mavronicolas et al. 1 Introduction Motivation and Framework. Originally introduced by Rosenthal [15], congestion games model resource sharing among (unweighted) players. Here, the strategy of each player is a set of resources. The cost for a player on resource e is given by a latency function for e, which depends on the total weight of all players choosing e. In congestion games with player-specific latency functions, which were later introduced by Milchtaich [13], players are weighted and each player chooses her own latency function for each resource, which determines her own player-specific cost on the resource. These choices reflect different preferences, beliefs or estimates by the players; for example, such differences occur in multiclass networks or in networks with uncertain parameters. In this work, we introduce a special case of (weighted) congestion games with player-specific latency functions [13], which we call (weighted) congestion games with player-specific constants. Here, each player-specific latency function is made up of a resource-specific delay function and a player-specific constant (for the particular resource); the two are composed by means of a group operation. We will be assuming that the underlying group is a totally ordered abelian group (see, for example, [9, Chapter 1]). Note that this new model of congestion games restricts Milchtaich’s one [13] since player-specific latency functions are no longer completely arbitrary; simultaneously, it generalizes the weighted generalization of Rosenthal’s model [15] since it allows composing player-specific constants into each (resource-specific) latency function. For example, (weighted) congestion games with player-specific additive constants (resp., multiplicative constants) correspond to the case where the group operation is addition (resp., multiplication). We will sometimes focus on network congestion games, where the resources and strategies correspond to edges and paths in a (directed) network, respectively; network congestion games offer an appropriate model for some aspects of routing problems. In such games, each player has a source and a destination node and her strategy set is the set of all paths connecting them. In a symmetric network congestion game, all players use the same pair of source and destination; else, the network congestion game is asymmetric. The simplest symmetric network congestion game is the parallel links network with only two nodes. The Individual Cost for a player is the sum of her costs on the resources in her strategy. In a (pure) Nash equilibrium, no player can decrease her Individual Cost by unilaterally deviating to a different strategy. We shall study questions of existence of, computational complexity of, and convergence to pure Nash equilibria for (weighted) congestion games with player-specific constants. For convergence, we shall consider sequences of improvement and best-improvement steps of players; in such steps, a player improves (that is, decreases) and best-improves her Individual Cost, respectively. A game has the Finite Improvement Property [14] (resp., the Finite Best-Improvement Property, also called Finite Best-Reply Property [13]) if all improvement paths (resp., best-improvement paths) are finite. Both properties imply the existence of a pure Nash equilibrium [14]; clearly, the first property implies the second. Also, the existence of a generalized ordinal potential is equivalent to the Finite Improvement Property [14] (and hence it implies the Finite Best-Improvement Property and the existence of a pure Nash equilibrium as well). A weighted potential Congestion Games with Player-Specific Constants 635 [14] is a particular case of a generalized ordinal potential; an exact potential [14] is a particular case of a weighted potential. We observe that the class of (weighted) congestion games with player-specific constants is contained in the more general, intuitive class of dominance (weighted) congestion games that we introduce (Proposition 1). In this more general class, it holds that for any pair of players, the preference of some of the two players with regard to any arbitrary pair of resources necessarily induces an identical preference for the other player (Definition 2). State-of-the-Art. It is known that every unweighted congestion game has a pure Nash equilibrium [15]; Rosenthal’s original proof uses an exact potential [14]. It is possible to compute a pure Nash equilibrium for an unweighted symmetric network congestion game in polynomial time by reduction to the min-cost flow problem [3]. However, the problem becomes PLS-complete for either (arbitrary) symmetric congestion games [3] or asymmetric network congestion games where the edges of the network are either directed [3] or undirected [1]. Weighted asymmetric network congestion games with affine latency functions are known to have a pure Nash equilibrium [6]; in contrast, there are weighted symmetric network congestion games with non-affine latency functions that have no pure Nash equilibrium (even if there are only 2 players) [6,12]. Weighted (network) congestion games on parallel links have the Finite Improvement Property (and hence a pure Nash equilibrium) if all latency functions are non-decreasing; in this setting, [5] proves that a pure Nash equilibrium can be computed in polynomial time by using the classical LPT algorithm due to Graham [10] when latency functions are linear. (This is the well-known setting of related parallel links, which is equivalent to using the identity function for all delay functions in a weighted congestion game with multiplicative constants.) In the general case, it is strongly N P-complete to determine whether a given weighted network congestion game has a pure Nash equilibrium [2]. For weighted congestion games with (non-decreasing) player-specific latency functions on parallel links, there is a counterexample to the existence of a pure Nash equilibrium with only 3 players and 3 links [13]. This result is tight since such games with 2 players have the Finite Best-Improvement Property [13]. Unweighted congestion games with (non-decreasing) player-specific latency functions have a pure Nash equilibrium but not necessarily the Finite Best-Improvement Property [13]. The special case of (weighted) congestion games with player-specific linear latency functions (without a constant term) was studied in [7,8]. Such games have the Finite Improvement Property if players are unweighted [7], while there is a game with 3 weighted players that does not have it [7]. For the case of 3 weighted players, every congestion game with player-specific linear latency functions (without a constant term) has a pure Nash equilibrium but not necessarily an exact potential [8]. For the case of 2 links, there is a polynomial time algorithm to compute a pure Nash equilibrium [8]. A larger class of (incomplete information) unweighted congestion games with player-specific latency functions that have the Finite Improvement Property has been identified in [4]; the special case of our model where the player-specific constants are additive is contained in this larger class. 636 M. Mavronicolas et al. Contribution and Significance. We partition our results on congestion games with player-specific constants according to the structure of the strategy sets in the congestion game: Games on parallel links: – Every unweighted congestion game with player-specific constants has a generalized ordinal potential (Theorem 1). (Hence, every such game has the Finite Improvement Property and a pure Nash equilibrium.) The proof employs a potential function involving the group operation; the proof that this function is a generalized ordinal potential explicitly uses the assumption that the underlying group is a totally ordered abelian group. We remark that Theorem 1 does not need the assumption that the (resource-specific) delay functions are non-decreasing. Theorem 1 simultaneously broadens two corresponding state-of-the-art results for two very special cases: (i) each delay function is the identity function and the group operation is multiplication [7] and (ii) the group operation is addition [4]. We note that, in fact, the potential function we used is a generalization of the potential function used in [4] (for addition) to an arbitrary group operation. However, [4] applies to all unweighted congestion games. – It is not possible to generalize Theorem 1 to weighted congestion games (with player-specific constants): there is such a game with 3 players on 3 parallel links that does not have the Finite Best-Improvement Property – hence, neither the Finite Improvement Property (Theorem 2). To prove this, we provide a simple counterexample for the case of player-specific additive constants. – Note that Theorem 2 does not outlaw the possibility that every weighted congestion game with player-specific constants has a pure Nash equilibrium. Although we do not know the answer for the general case with an arbitrary number of players, we have settled the case with 3 players: every weighted congestion game with playerspecific constants and 3 players has a pure Nash equilibrium (Corollary 3). The proof proceeds in two steps. First, we establish that there is a particular best-improvement cycle whose outlaw implies the existence of a pure Nash equilibrium (Theorem 3). We remark that an identical cycle had been earlier constructed by Milchtaich for the more general class of weighted congestion games with player-specific latency functions [13, Section 8]. Second, we establish that this particular best-improvement cycle is indeed outlawed for the more specific class of dominance weighted congestion games (Theorem 4). Since a weighted congestion game with player-specific constants is a dominance weighted congestion game, the cycle is outlawed for weighted congestion games with player-specific constants as well; this implies the existence of a pure Nash equilibrium for them (Corollary 3). This implies, in particular, a separation of this specific class from the general class of congestion games with player-specific latency functions with respect to best-improvement cycles. We remark that Corollary 3 broadens the earlier result by Georgiou et al. [8, Lemma B.1] for congestion games with player-specific multiplicative constants and identity delay functions. Congestion Games with Player-Specific Constants 637 Network congestion games: Recall that every unweighted congestion game with player-specific additive constants has a pure Nash equilibrium [4]. Nevertheless, we establish that it is PLS-complete to compute one (Theorem 5) even for a symmetric network congestion game. The proof uses a simple reduction from the PLS-complete problem of computing a pure Nash equilibrium for an unweighted asymmetric network congestion game [3]. Arbitrary (non-network) congestion games: Note that Theorem 2 outlaws the possibility that every weighted congestion game with player-specific constants has the Finite Best-Improvement Property. Nevertheless, we establish that every weighted congestion game with player-specific constants has a weighted potential for the special case of linear delay functions and player-specific additive constants (Theorem 6). (Hence, every such game has the Finite Improvement Property and a pure Nash equilibrium). The proof employs a potential function and establishes that it is a weighted potential. For the special case of weighted asymmetric network congestion games with affine latency functions (which are not player-specific), the potential function we used reduces to the potential function introduced in [6] for the corresponding case. Theorems 1 and 3 suggest that the class of congestion games with player-specific constants provides a vehicle for reaching the limit of the existence of potential functions towards the direction of player-specific costs. 2 Framework and Preliminaries Totally Ordered Abelian Groups. A group (G, ⊙) consists of a ground set G together with a binary operation ⊙ : G × G → G; ⊙ is associative and allows for an identity element and inverses. The group (G, ⊙) is abelian if ⊙ is commutative. We will consider totally ordered abelian groups with a total order on G [9] which satisfies translation invariance: for all triples r, s, t ∈ G, if r ≤ s then r ⊙ t ≤ s ⊙ t. Examples of totally ordered abelian groups include (i) (R>0 , ·) under the usual number-ordering, and (ii) (R2 , +) under the lexicographic ordering on pairs of numbers. We will often focus on the case where G is R (the set of reals). Congestion Games. For all integers k ≥ 1, we denote [k] = {1, . . . , k}. A weighted congestion game with player-specific latency functions [13] is a tuple Γ = (n, E, (wi )i∈[n] , (Si )i∈[n] , (fie )i∈[n],e∈E ). Here, n is the number of players and E is a finite set of resources. For each player i ∈ [n], wi > 0 is the weight and Si ⊆ 2E is the strategy set of player i. For each pair of player i ∈ [n] and resource e ∈ E, fie : R>0 → R>0 is a non-decreasing player-specific latency function. In the unweighted case, wi = 1 for all players i ∈ [n]. In a (weighted) network congestion game (with player-specific latency functions), resources and strategies correspond to edges and paths in a directed network. In such games, each player has a source and a destination node, each of her strategies is a path from source to destination and all paths are possible. In a symmetric network congestion game, all players use the same pair of source and destination; else, the network congestion game is asymmetric. In the parallel links network, there are only two nodes; this gives rise to symmetric network congestion games. 638 M. Mavronicolas et al. Definition 1. Fix a totally ordered abelian group (G, ⊙). A weighted congestion game with player-specific constants is a weighted congestion game Γ with player-specific latency functions such that (i) for each resource e ∈ E, there is a non-decreasing delay function ge : R>0 → R>0 , and (ii) for each pair of a player i ∈ [n] and a resource e ∈ E, there is a player-specific constant cie > 0, so that for each player i ∈ [n] and resource e ∈ E, fie = cie ⊙ ge . In a weighted congestion game with player-specific additive constants (resp., playerspecific multiplicative constants), G is R and ⊙ is + (resp., G is R>0 and ⊙ is ·). The special case of weighted congestion games with player-specific constants where for all players i ∈ [n] and resources e ∈ E, cie = ǫ (the identity element of G) yields the weighted congestion games generalizing the unweighted congestion games introduced by Rosenthal [15]. So, (weighted) congestion games with player-specific constants fall between the weighted generalization of congestion games [15] and (weighted) congestion games with player-specific latency functions [13]. We now prove that, in fact, congestion games with player-specific constants are contained within a more restricted class of congestion games with player-specific latency functions that we introduce. Fix a weighted congestion game Γ with player-specific latency functions. Consider a pair of (distinct) players i, j ∈ [n] and a pair of (distinct) resources e, e′ ∈ E. Say that i dominates j for the ordered pair e, e′ if for every pair of positive numbers x, y ∈ R>0 , fie (x) > fie′ (y) implies fje (x) > fje′ (y). Intuitively, i dominates j for e, e′ if the decision of i to switch her strategy from e to e′ always implies a corresponding decision for j; in other words, j always follows the decision of i (to switch or not) for the pair e, e′ . Definition 2. A weighted congestion game with player-specific latency functions is a dominance (weighted) congestion game if for all pairs of players i, j ∈ [n], for all pairs of resources e, e′ ∈ E, either i dominates j for e, e′ or j dominates i for e, e′ . We prove: Proposition 1. A (weighted) congestion game with player-specific constants is a dominance (weighted) congestion game. Proof. Fix a pair of players i, j ∈ [n] and a pair of resources e, e′ ∈ E. We proceed by case analysis. Assume first that cie ⊙ cje′ ≥ cie′ ⊙ cje . We will show that j dominates i for e, e′ . Fix a pair of numbers x, y ∈ R>0 . Assume that fje (x) > fje′ (y) or cje ⊙ ge (x) > cje′ ⊙ ge′ (y). By translation-invariance, it follows that cie ⊙ cje ⊙ ge (x) > cie ⊙ cje′ ⊙ ge′ (y). The assumption that cie ⊙ cje′ ≥ cie′ ⊙ cje implies that cie ⊙ cje′ ⊙ ge′ (y) ≥ cie′ ⊙ cje ⊙ ge′ (y). It follows that cie ⊙ ge (x) > cie′ ⊙ ge′ (y) or fie (x) > fie′ (y). Hence, j dominates i for e, e′ . Assume now that cie′ ⊙ cje > cie ⊙ cje′ . We will show that i dominates j for e, e′ . Fix a pair of numbers x, y ∈ R>0 . Assume that fie (x) > fie′ (y) or cie ⊙ge (x) > cie′ ⊙ ge′ (y). By translation-invariance, it follows that cje ⊙ cie ⊙ ge (x) > cje ⊙ cie′ ⊙ ge′ (y). The assumption that cie′ ⊙ cje > cie ⊙ cje′ implies that cje ⊙ cie′ ⊙ ge′ (y) > cje′ ⊙ cie ⊙ ge′ (y). It follows that cje ⊙ ge (x) > cje′ ⊙ ge′ (y) or fje (x) > fje′ (y). Hence, i dominates j for e, e′ . ⊓ ⊔ Congestion Games with Player-Specific Constants 639 Profiles and Individual Cost. A strategy for player i ∈ [n] is some specific si ∈ Si . A profile is a tuple s = (s1 , . . . , sn ) ∈ S1 ×  . . . × Sn . For the profile s, the load δe (s) on resource e ∈ E is given by δe (s) = w . For the profile s, the i∈[n] | si ∋e i  Individual Cost of player i ∈ [n] is given by ICi (s) = e∈si fie (δe (s)) = e∈si cie ⊙ ge (δe (s)). Pure Nash Equilibria. Fix a profile s. A player i ∈ [n] is satisfied if she cannot decrease her Individual Cost by unilaterally changing to a different strategy; else, player i is unsatisfied. So, an unsatisfied player i can take an improvement step to decrease her Individual Cost; if player i is satisfied after the improvement step, the improvement step is called a best-improvement step. An improvement cycle (resp., best-improvement cycle) is a cyclic sequence of improvement steps (resp., best-improvement steps). A game has the Finite Improvement Property (resp., Finite Best-Improvement Property) if all sequences of improvement steps (resp., best-improvement steps) are finite; clearly, the Finite Improvement Property (resp., the Finite Best-Improvement Property) outlaws improvement cycles (resp., best-improvement cycles). Clearly, the Finite Improvement Property implies the Finite Best-Improvement Property. A profile is a (pure) Nash equilibrium if all players are satisfied. Clearly, the Finite Improvement Property implies the existence of a pure Nash equilibrium (as also does the Finite Best-Improvement Property), but not vice versa [14]. A generalized ordinal potential for the game Γ [14] is a function Φ : S1 × . . . × Sn → R that decreases when a player takes an improvement step. Say that a function Φ : S1 × . . . × Sn → R is a weighted potential for the game Γ [14] if there is a weight vector b = (bi )i∈[n] such that for every player k ∈ [n], for every profile s, and for every strategy tk ∈ Sk that transforms s to t, it holds that ICk (s)− ICk (t) = bk ·(Φ(s)− Φ(t)). If this even holds for the vector b with bi = 1 for all i ∈ [n], the function Φ is an exact potential [14]. A game has a generalized ordinal potential if and only if it has the Finite Improvement Property (and hence the Finite Best-Improvement Property and a pure Nash equilibrium) [14]. PLS(-complete) Problems. PLS [11] includes optimization problems where the goal is to find a local optimum for a given instance; this is a feasible solution with no feasible solution of better objective value in its well-determined neighborhood. A problem Π in PLS has an associated set of instances IΠ . There is, for every instance I ∈ IΠ , a set of feasible solutions F (I). Furthermore, there are three polynomial time algorithms A, B and C. A computes for every instance I a feasible solution S ∈ F(I); B computes for a feasible solution S ∈ F(I), the objectice value; C determines, for a feasible solution S ∈ F(I), whether S is locally optimal and, if not, it outputs a feasible solution in the neighborhood of S with better objective value. A PLS-problem Π1 is PLS-reducible [11] to a PLS-problem Π2 if there are two polynomial time computable functions F1 and F2 such that F1 maps instances I ∈ IΠ1 to instances F1 (I) ∈ IΠ2 and F2 maps every local optimum of the instance F1 (I) to a local optimum of I. A PLS-problem Π is PLS-complete [11] if every problem in PLS is PLS-reducible to Π. 640 M. Mavronicolas et al. 3 Congestion Games on Parallel Links We now introduce a function Φ : S1 × . . . × Sn → R with Φ(s) = e (s)  δ ge (i) ⊙ n  cisi . i=1 e∈E i=1 for any profile s. We prove that this function is a generalized ordinal potential: Theorem 1. Every unweighted congestion game with player-specific constants on parallel links has a generalized ordinal potential. Proof. Fix a profile s. Consider an improvement step of player k ∈ [n] to strategy tk , which transforms s to t. Clearly, ICk (s) > ICk (t) or gsk (δsk (s)) ⊙ cksk > gtk (δtk (t)) ⊙ cktk . Note also that δsk (t) = δsk (s) − 1 and δtk (t) = δtk (s) + 1, while δe (t) = δe (s) for all e ∈ E \ {sk , tk }. Hence, Φ(s) =  ge (i) ⊙ e∈E\{sk ,tk } i=1 = e (s)  δ e∈E\ e∈E\ cisi ⊙  cisi ⊙ \{k}  cisi ⊙ ge (i) ⊙  ge (i) ⊙  \{k} gtk (i) ⊙ gsk (δsk (s)) ⊙ cksk δtk (s) gsk (i) ⊙  gtk (i) ⊙ gtk (δtk (t)) ⊙ cktk i=1 δtk (t) δsk (t) e∈E\{sk ,tk } i=1 = Φ(t),  i=1 δe (t)  cisi ⊙ i∈[n]\{k} so that Φ is a generalized ordinal potential. gtk (i) ⊙ cksk δtk (s) gsk (i) ⊙ i=1 i∈[n]  i=1 i=1 δsk (s)−1  {sk ,tk } = gsk (i) ⊙ i=1 i∈[n] i=1   δsk (s)−1 ge (i) ⊙ {sk ,tk } >  i∈[n]\{k} i=1 e (s)  δ δtk (s) δsk (s) δe (s)   i=1 gsk (i) ⊙  gtk (i) ⊙ cktk i=1 ⊓ ⊔ Theorem 1 immediately implies: Corollary 1. Every unweighted congestion game with player-specific constants on parallel links has the Finite Improvement Property and a pure Nash equilibrium. We continue to prove: Theorem 2. There is a weighted congestion game with additive player-specific constants and 3 players on 3 parallel links that does not have the Finite Best-Improvement Property. Congestion Games with Player-Specific Constants 641 Proof. By construction. The weights of the 3 players are w1 = 2, w2 = 1, and w3 = 1. The player-specific constants and resource-specific delay functions are as follows: cie Link 1 Link 2 Link 3 Player 1 0 ∞ 5 0 ∞ Player 2 0 Player 3 ∞ 0 2 Link 1 Link 2 Link 3 ge (1) 1 2 1 13 2 ge (2) 8 ge (3) 14 ∞ 10 Notice that the profiles 1, 2, 3 and 3, 1, 2 are both Nash equilibria. Consider now the cycle 1, 1, 3 → 1, 1, 2 → 1, 2, 2 → 3, 2, 2 → 3, 2, 3 → 3, 1, 3 → 1, 1, 3 . The Individual Cost of the deviating player decreases in each of these steps: IC1 IC2 IC3 1, 1, 3 14 3 1, 1, 2 14 2 1, 2, 2 3, 2, 2 IC1 IC2 IC3 8 13 7 13 IC1 IC2 IC3 3, 2, 3 2 12 3, 1, 3 15 1 So, this is an improvement cycle. Furthermore, note that each step in this cycle is a bestimprovement step, so this is actually a best-improvement cycle. The claim follows. ⊓ ⊔ We continue to consider the special case of 3 players but for the general case of weighted congestion games with player-specific constants. We prove: Theorem 3. Let Γ be a weighted congestion game with player-specific latency functions and 3 players on parallel links. If Γ does not have a best-improvement cycle l, j, j → l, l, j → k, l, j → k, l, l → k, j, l → l, j, l → l, j, j (where l = j, j = k, l = k are any three links and w1 ≥ w2 ≥ w3 ), then Γ has a pure Nash equilibrium. We now continue to prove: Theorem 4. Every dominance weighted congestion game with 3 players on parallel links does not have an improvement cycle of the form l, j, j → l, l, j → k, l, j → k, l, l → k, j, l → l, j, l → l, j, j where l = j, j = k, l = k are any three links and w1 ≥ w2 ≥ w3 . Proof. Assume, by way of contradiction, that there is a dominance congestion game with such a cycle. Since all steps in the cycle are improvement steps, one gets for player 2 that f2j (w2 + w3 ) > f2l (w1 + w2 ) and f2l (w2 + w3 ) > f2j (w2 ). In the same way, one gets for player 3 that f3j (w3 ) > f3l (w2 + w3 ) and f3l (w1 + w3 ) > f3j (w2 + w3 ). We proceed by case analysis on whether 2 dominates 3 or 3 dominates 2 for j, l . Assume first that 2 dominates 3 for j, l . Then, the first inequality for player 2 implies that f3j (w2 + w3 ) > f3l (w1 + w2 ) ≥ f3l (w1 + w3 ) (since f3l is non-decreasing and w2 ≥ w3 ), a contradiction to the second inequality for player 3. Assume now that 3 dominates 2 for j, l . Then, the first inequality for player 3 implies that f2l (w2 +w3 ) < f2j (w3 ) ≤ f2j (w2 ) (since f2j is non-decreasing and w2 ≥ w3 ), a contradiction to the second inequality for player 2. ⊓ ⊔ Since dominance (weighted) congestion games are a subclass of (weighted) congestion games with player-specific latency functions, Theorems 3 and 4 immediately imply: 642 M. Mavronicolas et al. Corollary 2. Every dominance weighted congestion game with 3 players on parallel links has a pure Nash equilibrium. By Proposition 1, Corollary 2 immediately implies: Corollary 3. Every weighted congestion game with player-specific constants and 3 players on parallel links has a pure Nash equilibrium. 4 Network Congestion Games Theorem 5. It is PLS-complete to compute a pure Nash equilibrium in an unweighted symmetric network congestion game with player-specific additive constants. Proof. Clearly, the problem of computing a pure Nash equilibrium in an unweighted symmetric congestion game with player-specific additive constants is a PLS-problem. (The set of feasible solutions is the set of all profiles and the neighborhood of a profile is the set of profiles that differ in the strategy of exactly one player; the objective function is the generalized ordinal potential since a local optimum of this functions is a Nash equilibrium [14].) To prove PLS-hardness, we use a reduction from the PLS-complete problem of computing a pure Nash equilibrium for an unweighted, asymmetric network congestion game [3]. For the reduction, we construct the two functions F1 and F2 : F1 : Given an unweighted, asymmetric network congestion game Γ on a network G, where (ai , bi )i∈[n] are the source and destination nodes of the n players and (fe )e∈E are the latency functions, F1 constructs a symmetric network congestion game Γ ′ with n players on a graph G′ , as follows: – G′ includes G, where for each edge e of G, ge′ := fe and c′ie = 0 for each i ∈ [n]. – G′ contains a new common source a′ and a new common destination b′ ; for each ′ ′ player i ∈ [n], we add an edge (a′ , ai ) with g(a ′ ,a ) (x) := 0, ci(a′ ,a ) := 0, and i i c′k(a′ ,ai ) := ∞ for all k = i; in addition we add for each player i ∈ [n] an edge ′ ′ ′ (bi , b′ ) with g(b ′ (x) := 0, ci(b ,b′ ) := 0, and ck(b ,b′ ) := ∞ for all k = i. i i i ,b ) F2 : Consider now a pure Nash equilibrium t for Γ ′ . The function F2 maps t to a profile s for Γ (which, we shall prove, is a Nash equilibrium for Γ ) as follows: – Note first that for each player i ∈ [n], ti (is a path that) includes both edges (a′ , ai ) and (bi , b′ ) (since otherwise ICi (t) = ∞). Construct si from ti by eliminating the edges (a′ , ai ) and (bi , b′ ). It remains to prove that s = F2 (t) is a Nash equilibrium for Γ . By way of contradiction, assume otherwise. Then, there is a player k that can decrease her Individual Cost in Γ by changing her path sk to s′k . But then player k can decrease her Individual Cost in Γ ′ by changing her path tk = (a′ , ak ), sk , (bk , b′ ) to t′k = (a′ , ak ), s′k , (bk , b′ ). So, t is not a Nash equilibrium for Γ ′ . A contradiction. ⊓ ⊔ We remark that Theorem 5 holds also for unweighted symmetric network congestion games with player-specific additive constants and undirected edges since the problem of computing a pure Nash equilibrium for an unweighted, asymmetric network congestion game with undirected edges is also PLS-complete [1]. Congestion Games with Player-Specific Constants 643 5 Arbitrary Congestion Games We now restrict attention to weighted congestion games with player-specific additive constants cie and linear delay functions fe (x) = ae · x. This gives rise to weighted congestion games with player-specific affine latency functions fie (x) = ae · x + cie , where i ∈ [n] and e  ∈ E. For this case, we introduce a function Φ : S1 × . . . × Sn → R n with Φ(s) = i=1 e∈si wi · (2 · cie + ae · (δe (s) + wi )), for any profile s. For any pair of player i ∈ [n] and resource e ∈ E, define φ(s, i, e) = wi · (2 · cie + ae ·   (δe (s) + wi )), so that Φ(s) = ni=1 e∈si φ(s, i, e). We now prove that this function is a weighted potential: Theorem 6. Every weighted congestion game with player-specific affine latency functions has a weighted potential. Proof. Fix a profile s. Assume that player k ∈ [n] unilaterally changes to the strategy tk , which transforms s to t. Clearly, Φ(s) − Φ(t)    = φ(t, i, e) φ(s, i, e) − i∈[n] e∈ti i∈[n] e∈si  = φ(s, k, e) − e∈sk   φ(t, k, e) + e∈tk i∈[n]\{k}   φ(s, i, e) −  φ(t, i, e) e∈ti e∈si  We treat separately the first and the second part of this expression. On one hand,     φ(s, k, e) − φ(t, k, e) = φ(s, k, e) − φ(t, k, e) e∈sk = e∈tk  e∈sk \tk e∈tk \sk  wk (2 · cke + ae · (δe (s) + wk )) − e∈sk \tk On the other hand,      φ(t, i, e) = φ(s, i, e) − = ⎛  i∈ [n]\{k} =  ⎝ e∈sk \tk =  e∈ sk \tk = wk · e∈ti =si e∈si i∈[n]\{k} wk (2 · cke + ae · (δe (t) + wk )). e∈tk \sk  | e∈si  | e∈si  e∈sk \tk  ae · (δe (s) − wk ) − wk ·  e∈tk \sk (wi · ae · (δe (s) − δe (t)))+ i∈[n]\{k} (φ(s, i, e) − φ(t, i, e))   | e∈si  i∈[n]\{k} tk \sk | e∈si e∈tk \sk (φ(s, i, e) − φ(t, i, e)) i∈[n]\{k} e∈  ⎞ (φ(s, i, e) − φ(t, i, e))⎠ e∈si ∩(tk \sk ) (φ(s, i, e) − φ(t, i, e))+ i∈[n]\{k}  i∈[n]\{k} e∈si (φ(s, i, e) − φ(t, i, e))+ e∈si ∩(sk \tk )   (wi · ae · (δe (s) − δe (t))) ae · (δe (t) − wk ) . 644 M. Mavronicolas et al. Putting these together yields that Φ is a weighted potential with weight vector b having 1 , i ∈ [n]. ⊓ ⊔ bi = 2w i Theorem 6 immediately implies: Corollary 4. Every weighted congestion game with player-specific affine latency functions has the Finite Improvement Property and a pure Nash equilibrium. References 1. Ackermann, H., Röglin, H., Vöcking, B.: On the Impact of Combinatorial Structure on Congestion Games. In: Proceedings of the 47th Annual IEEE Symposium on Foundations of Computer Science, pp. 613–622. IEEE Computer Society Press, Los Alamitos (2006) 2. Dunkel, J., Schulz, A.: On the Complexity of Pure-Strategy Nash Equilibria in Congestion and Local-Effect Games. In: Spirakis, P.G., Mavronicolas, M., Kontogiannis, S.C. (eds.) WINE 2006. LNCS, vol. 4286, pp. 50–61. Springer, Heidelberg (2006) 3. Fabrikant, A., Papadimitriou, C.H., Talwar, K.: The Complexity of Pure Nash Equilibria. In: Proceedings of the 36th Annual ACM Symposium on Theory of Computing, pp. 604–612. ACM Press, New York (2004) 4. Facchini, G., van Megan, F., Borm, P., Tijs, S.: Congestion Models and Weighted Bayesian Potential Games. Theory and Decision 42, 193–206 (1997) 5. Fotakis, D., Kontogiannis, S., Koutsoupias, E., Mavronicolas, M., Spirakis, P.: The Structure and Complexity of Nash Equilibria for a Selfish Routing Game. In: Widmayer, P., Triguero, F., Morales, R., Hennessy, M., Eidenbenz, S., Conejo, R. (eds.) ICALP 2002. LNCS, vol. 2380, pp. 123–134. Springer, Heidelberg (2002) 6. Fotakis, D., Kontogiannis, S., Spirakis, P.: Selfish Unsplittable Flows. Theoretical Computer Science 348(2–3), 226–239 (2005) 7. Gairing, M., Monien, B., Tiemann, K.: Routing (Un-)Splittable Flow in Games with PlayerSpecific Linear Latency Functions. In: Bugliesi, M., Preneel, B., Sassone, V., Wegener, I. (eds.) ICALP 2006. LNCS, vol. 4051, pp. 501–512. Springer, Heidelberg (2006) 8. Georgiou, C., Pavlides, T., Philippou, A.: Network Uncertainty in Selfish Routing. In: CDROM Proceedings of the 20th IEEE International Parallel & Distributed Processing Symposium, p. 105. IEEE Computer Society Press, Los Alamitos (2006) 9. Goodearl, K.R.: Partially Ordered Abelian Groups with Interpolation. American Mathematical Society (1986) 10. Graham, R.L.: Bounds on Multiprocessing Timing Anomalies. SIAM Journal of Applied Mathematics 17(2), 416–429 (1969) 11. Johnson, D.S., Papadimitriou, C.H., Yannakakis, M.: How Easy is Local Search? Journal of Computer and System Sciences 37(1), 79–100 (1988) 12. Libman, L., Orda, A.: Atomic Resource Sharing in Noncooperative Networks. Telecommunication Systems 17(4), 385–409 (2001) 13. Milchtaich, I.: Congestion Games with Player-Specific Payoff Functions. Games and Economic Behavior 13(1), 111–124 (1996) 14. Monderer, D., Shapley, L.S.: Potential Games. Games and Economic Behavior 14(1), 124– 143 (1996) 15. Rosenthal, R.W.: A Class of Games Possessing Pure-Strategy Nash Equilibria. International Journal of Game Theory 2, 65–67 (1973)