Communication in Bayesian games: Overview of work on implementing mediators in game theory

Transcription

1 Communication in Bayesian games: Overview of work on implementing mediators in game theory Françoise Forges June 2010 This is a slightly enriched version of the slides of my talk at the workshop Decentralized Mechanism Design, Distributed Computing and Cryptography organized by I. Abraham, D. Gerardi and J. Halpern, Princeton, June 3-4, The main reference is Forges (2009). CEREMADE et LEDa, Université Paris-Dauphine, francoise.forges@gmail.com 1

2 Plan of the talk 1. Basic de nitions: Bayesian game, cheap talk, mediator, communication equilibrium outcome (CEO), implementation of CEOs without a mediator 2. From in nitely repeated games to cheap talk 3. Cheap talk in sender-receiver games 4. From mental poker to cheap talk 5. Some representative implementation results 6. Topics for future research 2

3 1 Basic De nitions A Bayesian game is described by the parameters N; (T i ) i2n ; p; (A i ) i2n ; (u i ) i2n, where - N is the set of players - T i is the ( nite) set of types of player i, i 2 N - p is a probability distribution over T = Q j2n T j - A i is the ( nite) set of actions of player i; let A denote the action space: A = Q j2n Aj. - u i : T A! R is the (von Neumann-Morgenstern) utility function of player i is played as follows: A move of nature chooses t = (t j ) j2n according to p Player i is only informed of his own type t i The players choose simultaneously an action. In order to give a de nition of cheap talk in the Bayesian game, we rst focus on interim, bounded cheap talk extensions of, which take place for a nite number of stages S, after the move of nature: At the beginning of stage s = 1; 2; :::; S, every player i selects a single, non-anonymous, message (from a nite set), which is delivered to a subset of players at the end of stage s; messages from di erent players are sent simultaneously. Players perfectly recall all their past messages. At stage S + 1, the players simultaneously choose an action as in and are rewarded as in, independently of the exchanged messages. A cheap talk extension of generates a new ( extended ) game, to which the standard implicit common knowledge assumption applies. Remark: the expression cheap talk (or plain conversation, or direct communication, etc.), has been used, more or less formally, to cover various ways in which players can exchange messages. Here, a player can only 3

4 send one message at every stage, but this message can be received by several players at the same time and yet not be public ( partial broadcast ). Alternatively, one could for instance just assume that there is a secure communication channel between any pair of players. In the description above, messages are delivered without delay. Let us now introduce communication with a mediator in the Bayesian game (see Forges (1986) and Myerson (1986)). We start with the simplest conceivable de nition. For every N tuple of types t = (t j ) j2n, let q(:jt) be a probability distribution over the action space A and let q = (q(:jt)) t2t. The interpretation is that, after the move of nature in, a mediator invites every player i, i 2 N, to report a type r i, then selects an N-tuple of actions a according to q(:jr), r = (r i ) i2n, and privately recommends a i to player i. This de nes a mediated game. A system of type dependent probability distributions q is a communication equilibrium outcome (CEO) of if in the mediated game, none of the players can gain by unilaterally lying on his type or deviating from the recommended action (i.e., to be honest and obedient is a Nash equilibrium in the mediated game): X X p(t i jt i ) q(ajt)u i (t; a) X X p(t i jt i ) q(ajr i ; t i )u i (t; i (a i ); a i ) t i 2T i a2a t i 2T i a2a 8i 2 N; 8t i ; r i 2 T i ; 8 i : A i! A i This apparently restrictive de nition is justi ed by the canonical representation of communication equilibria (which generalizes the revelation principle in mechanism design, see, e.g., Myerson (1982)). In order to state it, let us consider an arbitrary ex ante and interim cheap talk extension of. By ex ante, we mean that the players can also exchange messages before the move of nature in. Let us add a mediator to the cheap talk extension, namely, a set of inputs Is i and a set of outputs Os i for every player i at every stage s, together with transition probabilities to choose N tuples of outputs at stage s as a function of all past inputs and outputs. A general mediated extension of is de ned as follows: at every stage, before exchanging cheap messages, the players simultaneously send an input to the mediator 4

5 and then receive a private output from the mediator. We can now state a basic lemma: Canonical representation of communication equilibria: The set of all Nash equilibrium outcomes of all general mediated extensions of a Bayesian game coincides with the set of CEOs of : The communication equilibrium extends the concept of correlated equilibrium, which was originally de ned by Aumann (1974, 1987) for games with complete information, in which j T i j= 1, i 2 I. The basic lemma applies. More precisely, let a correlated equilibrium outcome be de ned in this case as a probability distribution over A. The set of all Nash equilibrium outcomes of all general mediated extensions of a game G with complete information coincides with the set of correlated equilibrium outcomes of G. More generally, correlated equilibria also account for ex ante cheap talk in the Bayesian game (which can be represented in strategic form, with strategy sets A T i i, i 2 N); typically, a mediator recommends then a complete strategic plan (in A T i i ) to every player i 2 N before the beginning of the game (i.e., even before the move of nature). In particular, we can study correlated equilibria in an interim cheap talk extension of the Bayesian game. However some game theorists (e.g., Myerson (2004)) argue that a Bayesian game has no ex ante stage. We can now formulate some natural questions: Implementation problem: Can we achieve all CEOs of as (possibly re ned, e.g., sequential ) Nash (or correlated) equilibrium outcomes of a cheap talk extension of? Variant: Can we achieve a speci c class of CEOs of as (possibly re ned ) Nash (or correlated) equilibrium outcomes of a cheap talk extension of? Characterization problem: Can we characterize all (possibly re ned ) Nash (or correlated) equilibrium outcomes of a speci c cheap talk extension of? 5

6 2 From in nitely repeated games to cheap talk In the late sixties, Aumann and Maschler introduced the model of in nitely repeated games with incomplete information in order to study unmediated communication in strategic games. This seminal work rst appeared as a (rather con dential) report to the U.S. Arms Control and Disarmament Agency (see Aumann and Maschler (1995)). Extending Aumann and Maschler s model, Aumann, Maschler and Stearns (1968) consider an in nitely repeated two-person game 1 in which player 1 learns his type t 1 2 T 1 at stage 0, jt 2 j = 1 and both players simultaneously choose actions at every stage 1, 2,...The players observe each other s action but not their stage utilities. In the game 1, information transmission is cheap as far as utilities are not discounted: the players can use as many stages as they wish for communication purposes. Simultaneous moves allow for jointly controlled lotteries, a device which will turn out to be crucial later on to implement mediators by cheap talk. A major insight in Aumann, Maschler and Stearns (1968) is that several stages of information transmission make a di erence in the sense that the set of Nash equilibrium outcomes that the players can achieve strictly increases with the number of communication stages. But the game 1 is a complex one, since in addition to cheap talk, the game allows for long term cooperation, as in the so-called Folk theorem. S. Hart (1985), together with Aumann and Hart (1986), provides a full characterization of all Nash equilibrium outcomes of 1. In particular, building on a geometric example of Aumann and Hart (1986), Forges (1984) demonstrates that more equilibrium outcomes can be achieved with an unbounded number of stages of information transmission. The previous results gave rise to remakes, namely, they were reformulated in terms of Nash equilibrium outcomes of long cheap talk extensions of a (one-shot) Bayesian game. Forges (1990a) considers a job market example with the same basic features as in Forges (1984). Aumann and Hart (2003) characterize the Nash equilibrium outcomes of in nite cheap talk extensions of any Bayesian game in which jnj = 2, jt 2 j = 1. Coming back to the in nitely repeated game, Forges (1985) proposes a characterization of all correlated equilibria of 1 under the further assumption that player 1 s actions have no impact on utilities, i.e., that u i (t 1 ; a 1 ; a 2 ) 6

7 u i (t 1 ; a 2 ), i = 1; 2. 1 then amounts to a sender-receiver game. As a direct by-product of Forges (1985), one gets the following result on the implementation of a mediator by cheap talk: Proposition 1: Let be a two-person Bayesian game with jt 2 j = ja 1 j = 1. Every communication equilibrium outcome (CEO) of can be achieved as a correlated equilibrium of an interim cheap talk extension of, in which the informed player sends a single message from a nite set to the uninformed player. Forges (1988a) extends the result to the case of multiple senders; Forges (1988b) characterizes the correlated equilibria in all games 1 in which jn j = 2 and jt 2 j = 1, i.e., allowing player 1 to make utility relevant decisions. 3 Cheap talk in sender-receiver games Crawford and Sobel (1982) (and, independently, Green and Stokey (1980/ 2007)) considered a speci c, economically relevant, class of sender-receiver games, in which T 1 = A 2 = [0; 1]. They characterized, as partitional equilibrium outcomes, all Nash equilibrium outcomes of the one-shot interim cheap talk extension of, in which the informed player sends a single message to the uninformed player. The main message at the time was that Cheap talk matters!, which came as a surprise to economists who, in the light of e.g. Spence (1973), considered that signalling costs were crucial to fruitful information transmission. Krishna and Morgan (2004) show that several stages of information transmission make a di erence in Crawford and Sobel s model by exhibiting a three-stage cheap talk equilibrium which Pareto-improves on single stage ones. Goltsman, Hörner, Pavlov and Squintani (2009) provide necessary and su cient conditions for cheap talk implementation of e cient CEOs. Blume (2010) establishes that e cient CEOs can be implemented as correlated equilibria of the one-shot cheap talk game (relying on a construction which di ers from Forges (1985)). 7

8 4 From mental poker to cheap talk In , Imre Bárány was visiting CORE (Belgium) and presented his paper Mental poker with Z. Füredi. J.-F. Mertens suggested that the techniques used in that paper could be useful to implement correlated equilibria by plain conversation between su ciently many players. Proposition 2 (Bárány (CORE DP 1987, 1992)): Let G be a strategic form game with n 4 players. Every rational (i.e., in Q) correlated equilibrium outcome of G can be achieved as a Nash equilibrium of a cheap talk+pvr (ex ante) extension of G with nitely many stages. In Bárány s original statement, cheap talk is enriched by possible public veri cation of the record (PVR): at every stage, every player can ask for the revelation of all exchanged messages. But this assumption is in fact not necessary, thanks to logical signatures (Ben-Or, Goldwasser and Widgerson (1988), Vida (2007)). Observe that the implementation is in Nash, not sequential, equilibrium; indeed, in Bárány s construction, players need not optimize out of the equilibrium path. 5 Some (further) representative implementation results The following result is in the same vein as proposition 1: Proposition 3 (Forges (1990b)): Let be a Bayesian game with n 4 players. Every communication equilibrium outcome (CEO) of can be achieved as a ( sequential ) correlated equilibrium of an interim two stage cheap talk extension of. Several de nitions of sequential correlated equilibrium are conceivable; in the previous statement, we just mean that the construction does not involve any irrational punishments. The same remark already applied to proposition 1. By combining the variant of proposition 2 (without PVR), the previous proposition 3 and Gerardi (2000), we get the following 8

9 Corollary: For n 4, every rational CEO can be implemented as a Nash equilibrium with interim cheap talk in nitely many stages. A direct application of proposition 2 and proposition 3, as in Forges (1990b), leads the players to perform the cheap talk phase achieving correlation at the ex ante stage. Gerardi (2000) modi es the construction to guarantee that cheap talk only takes place at the interim stage. The corollary is consistent with recent results, like Abraham, Dolev, Gonen and Halpern (2006) and Abraham, Dolev and Halpern (2008), which will be presented in more details later in the workshop. The corollary inherits the main drawback of proposition 2, namely that the implementation of the correlation phase makes use of possibly nonrational punishments. The problem disappears once there are at least 5 players: Proposition 4 (Gerardi (2004)): Let be a Bayesian game with n 5 players. Every rational (i.e., in Q), full support, CEO of can be achieved as a sequential Nash equilibrium of an interim cheap talk extension of with nitely many stages. The solution to the implementation problem in this proposition satis es three essential requirements: (essentially) all CEO s are implemented, cheap talk takes place exclusively at the interim stage and the implementation is achieved with an appropriate re nement of Nash equilibrium (see Gerardi and Myerson (2007) for a detailed analysis of sequential CEOs, in particular for the role of the full support assumption). The assumption of 5 players in proposition 4 is of course a strong one. Possibility results are also available for n 4 players: Proposition 5 (Ben Porath (2003+corrigendum 2006)): Let be a Bayesian game with n 4 players. Let q be a rational (i.e., in Q), full support, CEO of which interim strictly dominates a Nash equilibrium of : 8i 2 N; 8t i 2 T i : U i (q j t i ) > U i ( j t i ) q can be achieved as a sequential Nash equilibrium of an interim cheap talk extension of with nitely many stages. In the previous statement, the restriction on the CEO s to be implemented is extremely useful, since allows to credibly punish all players, whatever 9

10 their types, at the same time (in, i.e., at the prior probability distribution p). The proofs of propositions 3, 4 and 5 all make use of the number of players to rely on majority rule at some point: a receiver expects the same message from three di erent senders and decides by majority rule; this makes useless any sender s unilateral deviation. Ben Porath (2003) states proposition 5 for n = 3, but, as explained in Ben Porath (2006), for n = 3, an additional assumption is required. Before we go on with n = 3, let us mention that for n 4, we can still hope for possibility results that would not put any restriction on the implementable CEO s: Conjecture (Vida (2010, personal communication)): for n 4, every CEO can be implemented as a sequential Nash equilibrium with interim cheap talk if messages can be chosen from a continuum. When the number of players is n = 3, several impossibility results indicate that satisfactory implementation require careful assumptions (see Bárány (1992), Forges (1990b), Ben Porath (2003), Abraham, Dolev, Gonen and Halpern (2006), Abraham, Dolev and Halpern (2008), Vida (2007)). For instance, Ben Porath (2006) shows that proposition 5 holds for n = 3 if it is understood that cheap talk allows for an appropriate form of PVR (every player i can safely deposit at stage t a message to be publicly revealed at stage t + s). Other remedies are: cheap talk using a continuum of messages (Vida (2010, personal communication)), cheap talk with a possibly in nite number of rounds (see Abraham, Dolev and Halpern (2008) and the talks by Joe Halpern and Ittai Abraham in this workshop). For n = 2 players, there are even more impossibility results. For instance, with complete information, there is no hope to implement more than lotteries over Nash equilibrium outcomes. Proposition 6 (Vida (2007)): Let be a Bayesian game with n 2 players; essentially every CEO of can be implemented as a correlated equilibrium with interim, long, a.s. nite, cheap talk. Ben Porath (1998) and R.V. Krishna (2007) propose implementation results for n = 2 players by relaxing the rules of cheap talk so as to allow the players to make use of urns or envelopes. Dodis, Halevy and Rabin (2000) and Urbano and Vila (2002, 2004) rely on cryptography. 10

11 6 Topics for future research In nitely repeated games with incomplete information: as pointed out in section 2, this model was conceived to study cheap information transmission at the same time as cooperation but was mostly studied for two person games with lack of information on one side. An important feature of this framework is that it xes the rules of the implicit cheap talk allowed to the players (which can be seen as an advantage or a drawback). Another important property of this model is that it implicitly captures the e ects of commitment. Just to mention a possible recent interest in the topic: Golosov, Skreta, Tsyvinski and A. Wilson (2009) consider a repeated version of Crawford and Sobel (1882) s game. Strong (or coalition-proof) correlated/communication equilibria: in this talk, we focused on communication protocols which resist to every player s unilateral deviations (captured by Nash equilibrium). Starting with Aumann s strong Nash equilibrium, game theory has developed tools to deal with deviations by coalitions of players. Some papers (Milgrom and Roberts (1996), Moreno and J. Wooders (1996), I. Ray (1996)) proposed a de nition of coalition-proof correlated equilibria, while Einy and Peleg (1995) made an attempt at de ning coalitionproof communication equilibria. A main issue is to determine the cheap talk extension in which to apply game-theoretical solution concepts dealing with possible deviations by coalitions: the revelation principle typically does not survive in this environment. More generally, the deviations by coalitions are handled with a very speci c scenario in mind. In computer science, it is customary to design protocols that are robust to deviations by coalitions of a su ciently small size (see Abraham, Dolev, Gonen and Halpern (2006), Abraham, Dolev and Halpern (2008) and the talks by Joe Halpern and Ittai Abraham in this workshop). Mediation with commitment: in this talk, mediated communication has only been used to correlate strategies and to exchange information, but it can also serve as a commitment device. Moulin and Vial (1978) is an early paper using this idea. Recent papers on the topic are Monderer and Tennenholtz (2009), Ashlagi, Monderer and Tennenholtz (2009). 11

12 References [1] Abraham, I., D. Dolev, R. Gonen and J. Halpern (2006), Distributed computing meets game theory: robust mechanisms for rational secret sharing and multiparty computation, 25th ACM Symposium on Principles of Distributed Computing. [2] Abraham, I., D. Dolev and J. Halpern (2008), Lower bounds on implementing robust and resilient mediators, Fifth IACR Theory of Cryptography Conference. [3] Ashlagi, I., D. Monderer and M. Tennenholtz (2009), Mediators in position auctions, Games and Economic Behavior 67, [4] Aumann, R. J. (1974), Subjectivity and correlation in randomized strategies, Journal of Mathematical Economics 1, [5] Aumann, R. J. (1987), Correlated equilibrium as an expression of Bayesian rationality, Econometrica 55, [6] Aumann, R. J., Maschler, M. and Stearns, R. (1968), Repeated games with incomplete information: an approach to the nonzero sum case, Reports to the U.S. Arms Control and Disarmament Agency, ST-143, Chapter IV, [7] Aumann, R. J. and Maschler M. (1995), Repeated Games of Incomplete Information, Cambridge: M.I.T. Press. [8] Aumann, R. J. and S. Hart (1986), Bi-convexity and bi-martingales, Israel Journal of Mathematics 54, [9] Aumann, R. J. and S. Hart (2003), Long cheap talk, Econometrica 71, [10] Bárány, I. (1992), Fair distribution protocols or how players replace fortune, Mathematics of Operations Research 17, [11] Bárány, I. and Z. Füredi (1983), Mental poker with three or more players, Information and Control 59,

13 [12] Ben-Or M., S. Goldwasser and A. Widgerson (1988), Completeness theorems for non-cryptographic fault-tolerant distributed computation (extended abstract), Proceedings 20 STOC ACM, [13] Ben-Porath, E. (1998), Correlation without mediation: expanding the set of equilibrium outcomes by cheap pre-play procedures, Journal of Economic Theory 80, [14] Ben-Porath, E. (2003), Cheap talk in games with incomplete information, Journal of Economic Theory 108, [15] Ben-Porath, E. (2006), A correction to Cheap talk in games with incomplete information, mimeo, Hebrew University of Jerusalem. [16] Blume, A. (2010), Correlated equilibria in sender-receiver games, mimeo, University of Pittsburgh. [17] Crawford, V. and J. Sobel (1982), Strategic information transmision, Econometrica 50, [18] Dodis, Y., S. Halevi and T. Rabin (2000), A cryptographic solution to a game theoretic problem, CRYPTO 2000: 20th International Cryptology Conference, Springer-Verlag, [19] Forges, F. (1984), Note on Nash equilibria in in nitely repeated games with incomplete information, International Journal of Game Theory 13, [20] Forges, F. (1985), Correlated equilibria in a class of repeated games with incomplete information, International Journal of Game Theory 14, [21] Forges, F. (1986), An approach to communication equilibria, Econometrica 54, [22] Forges, F. (1988a), Can sunspots replace a mediator? Journal of Mathematical Economics 17, [23] Forges, F. (1988b), Communication equilibria in repeated games with incomplete information, Mathematics of Operations Research, 13,

14 [24] Forges, F. (1990a), Equilibria with communication in a job market example, Quarterly Journal of Economics 105, [25] Forges, F. (1990b), Universal mechanisms, Econometrica 58, [26] Forges, F. (2009), Correlated equilibria and communication in games, Encyclopedia of Complexity and Systems Science, R. Meyers, ed., Springer. [27] Gerardi, D. (2000), Interim pre-play communication, mimeo, Yale University. [28] Gerardi, D. (2004), Unmediated communication in games with complete and incomplete information, Journal of Economic Theory 114, [29] Gerardi, D. and R. Myerson (2007), Sequential equilibria in Bayesian games with communication, Games and Economic Behavior 60, [30] Golosov, M, V. Skreta, A. Tsyvinski and A. Wilson (2009), Dynamic strategic information transmission, mimeo. [31] Goltsman, M., J Hörner, G. Pavlov and F. Squintani (2009), Arbitration, mediation and cheap talk, Journal of Economic Theory 144, [32] Green, J. and N. Stokey (2007), A two-person game of information transmission, Journal of Economic Theory 135, (H.I.E.R. Discussion Paper, Harvard University 1980). [33] Hart, S. (1985), Nonzero-sum two-person repeated games with incomplete information, Mathematics of Operations Research 10, [34] Krishna, R. V. (2007), Communication in games of incomplete information: two players, Journal of Economic Theory 132, [35] Krishna, V. and J. Morgan (2004), The art of conversation: eliciting information from experts through multistage communication, Journal of Economic Theory 117,

15 [36] Milgrom, P., and Roberts, J. (1996), Coalition-proofness and correlation with arbitrary communication possibilities, Games and Economic Behavior 17, [37] Monderer, D. and M. Tennenholtz (2009), Strong mediated equilibrium, Arti cial Intelligence 173, [38] Moreno, D., and J. Wooders (1996), Coalition-proof equilibrium, Games and Economic Behavior 17, [39] Moulin H. and J.P. Vial (1978), Strategically zero-sum games: the class of games whose completely mixed equilibria cannot be improved upon, International Journal of Game Theory [40] Myerson, R. (1982), Optimal coordination mechanisms in generalized principal-agent problems, Journal of Mathematical Economics, 10, [41] Myerson, R. (1986), Multistage games with communication, Econometrica 54, [42] Myerson, R. (2004), Harsanyi s games with incomplete information, Management Science 50, [43] Ray, I. (1996), Coalition-proof correlated equilibrium: a de nition, Games and Economic Behavior 17, [44] Spence, A. M. (1973), Job market signalling, Quarterly Journal of Economics 87, [45] Urbano, A. and Vila, J. (2002), Computational complexity and communication: coordination in two-player games, Econometrica 70, [46] Urbano, A. and Vila, J. (2004a), Computationally restricted unmediated talk under incomplete information, Economic Theory, 23, [47] Vida, P. (2007), From communication equilibria to correlated equilibria, mimeo, University of Vienna. 15