* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Download Mechanism Design
Survey
Document related concepts
Transcript
Recent contributions to Implementation/Mechanism Design Theory E. Maskin Mechanism Design/Implementation • part of economic theory devoted to “reverse engineering” • usually we take mechanism, game, or economy as given – try to predict the outcomes it generates in equilibrium • in MD, we (the “planner”) start with outcome(s) a we want as a function of underlying state (describes utility utility functions ui , , technologies, endowments, etc.) social choice function f : A – difficulty: we may not know state – try to design a mechanism (game, outcome function, tax schedule) whose equilibrium outcomes same as that prescribed by social choice function mechanism implements SCF 2 • Goes back (at least) to 19th century Utopians – can one design “humane” alternative to laissezfaire capitalism? • Socialist Planning Controversy 1920s-40s – can one construct a centralized planning mechanism that replicates or improves on competitive markets? O. Lange and A. Lerner: yes L. von Mises and F. von Hayek: no – brought to fore 2 major themes incentives information 3 Formal mechanism-design theory dates from 3 papers • L. Hurwicz (1960) – introduced basic concepts • mechanism • informational decentralization • informational efficiency • W. Vickrey (1961) – exhibited a particular but important mechanism: 2nd price auction • J. Mirrlees (1971) – developed standard analytic techniques – derived standard properties (e.g., “no distortion at top”) 4 Since then, field has expanded dramatically • vast literature, ranging from – very general possible outcomes abstract set of social alternatives (at least 10 major survey articles and books in last dozen years or so) – quite particular design of bilateral contracts between buyer and seller (several recent books on contract theory, including BoltonDewatripont (2005) and Laffont-Martimort (2002)) design of auctions for allocating a good among competing bidders (several recent books - - Krishna (2002), Milgrom (2004), Klemperer (2004)) – far too much recent work to survey properly here – will pick 3 specific developments (both general and particular) 5 • interdependent values in auction design • robustness of mechanisms • indescribable states, renegotiation and incomplete contracts 6 Interdependent values in auction design • seller has 1 good • n potential buyers • how to allocate good efficiently? (to buyer who values good the most) i.e., how to implement SCF that selects efficient allocation 7 In private values case (each buyer’s valuation is independent of others’ information), Vickrey (1961) answered question: • 2nd price auction is efficient – buyers submit bids – winner is high bidder – winner pays 2nd highest bid • if vi is buyer i’s valuation, optimal for him to bid bi vi • winner will have highest valuation 8 What if values are interdependent? • each buyer i gets private signal si (onedimensional) • buyer i’s valuation is vi si , si • buyer i no longer knows own valuation – so can’t bid valuation in equilibrium – might bid expected valuation, but this not enough for efficiency: might have Es vi si , si Es v j s j , s j but vi si , si v j s j , s j i j 9 • consider auction in which – each buyer i makes contingent bid bi vi i's bid if other buyers' valuations revealed to be vi – calculate fixed point v1 , , vn such that for all i vi bi v i – winner is buyer i such that () vi max v j j i – winner pays bi v i max vj , where vj b j v j • j i ji v If (i) vi 0 and (ii) vi j whenever vi v j and i j si si si then, in equilibrium buyer i with signal si bids true contingent valuation: () bi v i si , s i vi si , si for all si – idea: i pays lowest constant bid max b j v j that would win • • • j i – in ordinary 2nd price auction, i pays lowest bid from , true valuations are fixed point from , winner has highest valuation auction efficient 10 • dynamic auctions like English auction easier on buyers than one-shot mechanisms like 2nd –price auction • Perry and Reny (2005) develop dynamic auction – even works for multiple goods, if substitutes • open problem: How to handle multiple goods with complementarities in dynamic auction • Jehiel and Moldovanu (2002) and Jehiel, Moldovanu, Meyer-Ter-Vehn, and Zame (2005) establish fundamental limitation on how far one can go with multiple goods and multidimensional signals 11 Robust Mechanism Design auction in which buyer i bids bi vi is “robust” or “independent of detail” in sense that • it doesn’t matter whether auction designer knows buyers’ signal spaces or functional forms vi si , si • it doesn’t matter what buyer i believes about the distribution of s i – optimal for buyer i to set bi vi si , si vi si , si for all si regardless of i’s belief about s i – i.e., bidding truthfully is an ex post equilibrium (remains equilibrium even if i knows s i ) 12 Why is robustness important? • common in Bayesian mechanism design to identify buyer i’s possible types with his possible preferences (common more generally than justified) set of possible types set of possible preferences • but this has extreme implication: if you know i’s preferences, know his beliefs over other’s types – no reason why this should hold – very strong consequences: in auction model above, if signals correlated, auctioneer can attain efficiency and extract all buyer surplus, even without conditions such as vi v j si si (Crémer and McLean (1985)) – As Neeman (2001) and Heifetz and Neeman (2004) show, CrémerMcLean result goes away for suitably rich type spaces (preference corresponds to multiple possible beliefs) • no reason why auction designer should know what buyers’ type spaces are (how preferences related to beliefs) 13 Given SCF f : , can we find mechanism for which, regardless of type space associated with preference space , there always exists f-optimal equilibrium? (robust partial implementation) • sufficient condition: f partially implementable in ex post equilibrium, i.e., there exists mechanism that always has f-optimal ex post equilibrium (may be other equilibria) – ex post equilibrium reduces to dominant strategy equilibrium with private values • interestingly, Bergemann and Morris (2004) show that condition not necessary 14 But ex post partial implementability is necessary for robust partial implementation if • outcome space takes form X Y1 "common" Yn "private transfers" outcome (public good) agent i cares just about x, yi • satisfied in above auction model (and, more generally, in quasilinear models) 15 • So far have concentrated on partial implementation (not all equilibria have to be f-optimal) • But unless planner sure that agents will play f-optimal equilibrium, more appropriate concept is full implementation: all equilibria of mechanism must be f-optimal 16 • key to full implementation is some species of monotonicity – full implementation in Nash equilibrium (agents have complete information) requires standard monotonicity: for all, , if, for all j, then u j f , u j b, implies u j f , u j b, for all b f f equivalently: for all : and for which f f (here ), there exist i and a A such that ui f , ui a, and ui a, ui f , analogous condition for Bayesian implementation (agents have incomplete information) - - Postlewaite and Schmeidler (1986), Palfrey and Srivastava (1989), Jackson (1991) 17 • standard monotonicity: for all : and for which f f there exist i and a A such that ui f , ui a, and ui a, ui f , • ex post monotonicity key to ex post full implementabilty (Bergemann and Morris 2005): for all 1 , , n j : j j and i, i for which f f , there exist i and a such that ui f i , i i , i , i i ui a, i , i i for all i and ui a, ui f , 18 • in economic settings with n > 3, f is ex post fully implementable if and only if it satisfies ex post monotonicity and ex post incentive compatibility ui f , ui f i , i , with all i, i , • ex post equilibrium is refinement of Nash equilibrium but ex post monotonicity doesn’t imply standard monotonicity (nor is it implied) – although ex post equilibrium is more demanding solution concept, makes ruling out equilibria easier 19 • Notable SCF where ex post monotonicity but not standard monotonicity satisfied: efficient generalized second-price allocation rule in interdependent values auction model when n > 3 • ex post monotonicity satisfied because truthful equilibrium is unique ex post equilibrium – Berliun (2003) shows that hypothesis n > 3 is important: there exist inefficient ex post equilibria in case n = 2. • efficient second-price allocation rule not standardly monotonic: if lower losers’ signal values, monotonicity requires that same allocation still chosen - - but winner’s payment will fall 20 • But ex post full implementation not quite enough – ensures ex post equilibria optimal – but other equilibria could be nonoptimal • really need robust full implementation: f : implementable by mechanism such that, regardless of type space associated with Θ, all equilibria are f-optimal 21 • robust monotonicity is key: for all there exist i and i such that f i,i f i,i for all i and for all i and all probability distributions i on i1 i there exists a A such that ui f i , i i , i , i i ui a, i , i i for all i and i i ui a, i ,i i i ui f i , i , i , i i i – stronger than both ex post monotonicity and standard monotonicity – together with ex post incentive compatibility, necessary and sufficient for robust full implementation in economic environments • For n > 3, generalized 2nd price allocation rule robustly fully virtually implementable as long as not “too much” interdependence, i.e., vi not too big s j 22 • so far, “robustness” requirement pertains to mechanism designer – may not know agents’ type spaces • also recent contributions in which robustness pertains to agents playing mechanism 23 • large literature considering implementation in various refinements of Nash equilibrium – allows implementation of nonmonotonic SCFs • any species of Nash equilibrium entails that agents have common knowledge of preferences, i.e., state • but what if agents are (slightly ) uncertain about ? – which SCFs are robust to this uncertainty? – answer depends on nature of uncertainty • for a natural form of uncertainty, only monotonic SCFs robustly implementable in this sense (Chung and Ely 2003) 24 Example (Jackson and Srivastava (1996)) n2 A a, b, c 1 c a b , 2 a b c f implemented in undominated Nash equilibrium by 1 c a b f a 2 a c b f c m2 m2 m1 a a m1 b c not monotonic m1, m2 unique u.d. equil in m1, m2 unique u.d. equil in • m2 is dominated in state if player 2 sure that state is • If 2 attaches small probability to , m2 not dominated , so mechanism no longer implements f • in fact, no mechanism can implement f because nonmonotonic 25 Open problems: (1) Implications of ex post monotonicity and robust monotonicity for applications (2) implications of other sorts of uncertainty for implementability 26 Indescribable States, Renegotiation and Incomplete Contracts • incomplete contracts literature studies how assigning ownership (or control) of productive assets affects efficiency of outcome • For efficiency to be in doubt, must be some constraint on contracting (i.e., on mechanism design) 27 • In this literature, constraint is incompleteness of contract – contract not as fully contingent on state of world as parties would like • Reason for incompleteness – parties plan to trade some good in future – do not know characteristics of good (state) at the time of contracting (although common knowledge at time of trade) – contract cannot even describe set of possible states (too vast) – so contract cannot be fully contingent 28 Nevertheless, we have: Irrelevance Theorem: If parties are risk averse and can assign probability distribution to their future payoffs, then can achieve same expected payoffs as with fully contingent contract (even though cannot describe possible states in advance) 29 Idea: • design contracts that specify payoff contingencies • later, when state of world realized, can fill in physical details • possible problem: incentive compatibility – will it be in parties interest to specify physical details truthfully? – but if different states different preferences, can design mechanism to ensure incentive compatibility • make mechanism part of contract 30 Where does risk aversion come in? Answer: helps with incentive compatibility • if parties are supposed to play m1 , m2 in but 1 plays m1 instead, must be punished • but if m1, m2 is equilibrium play in , then not clear from m1, m2 who has deviated • resolution: punish them both with inefficient outcome a. 31 But what if parties can renegotiate outcome ex post ? • not an issue when designer is third party; here parties themselves design contract • why settle for a ? • will renegotiate a to get something Pareto optimal • renegotiation interferes with effective punishment: can’t punish both parties • in Segal (1999) and Hart and Moore (1999), renegotiation is so constraining that mechanisms are useless Risk aversion • Pareto frontier (in utility space) is strictly concave • so if randomize between 2 Pareto optimal points, generate point in interior (bad outcome) • so can punish both parties after all. 32 Why isn’t randomization renegotiated away? • randomization occurs only out of equilibrium • ex ante, parties have no incentive to renegotiate (expect equilibrium – have incentive not to renegotiate: renegotiation interferes with getting equilibrium outcome • what about renegotiation ex post? – create randomizing device so that as soon as a party deviates from equilibrium, randomization realized – no time to renegotiate 33 Open problem: How to provide fully satisfactory foundation for incomplete contracts? • possible answer: bounded rationality (inability to perform dynamic programming) 34