Extensive games with perfect information ECON5200 Advanced microeconomics Lectures in game theory Fall 2009, Part 2 05.10.2009 G.B. Asheim, ECON5200-2 1 Incomplete information Imperfect information Complete information Perfect information Lecture 1 Strategic games Static Application: Bargaining games Dynamic Lecture 2 Extensive games (multi-stage games) First: Extensive games with perfect information and without simultaneous moves 05.10.2009 Lecture 1 Bayesian games Lecture 3 Extensive games (the general case) Application: Repeated games Then: Extensive games with perfect information and with simultaneous moves (multi-stage games) 2 G.B. Asheim, ECON5200-2 Definition (89.1) of an extensive game w/perf. info. N = {1, 2} 1 (The set of players) H = {, T, B, (T, L), (T, R), (B, L'), (B, R')} T (The set of histories) 2 Z {(T, L), (T, R), (B, L'), (B, R')} (The set of terminal histories) H\Z {, T, B} B 2 L R L R 1 2 1 1 2 1 0 0 P() 1, P(T) 2, P(B) 2 (Player function) 1(T, L) 1, 1(T, R) 1, 1(B, L') 2, 1(B, R') 0 (Payoff 2(T, L) 2, 2(T, R) 1, 2(B, L') 1, 1(B, R') 0 functions) A() {a| (, a) H} {T, B} (Action sets) A(T) {a| (T, a) H} {L, R} A(B) {a| (B, a) H} {L, R} 05.10.2009 G.B. Asheim, ECON5200-2 3 1 Definition 92.1 A strategy for player i Strategy N in an extensive game w/perf. info. 1 is a function that assigns an action in A(h) to each non-terminal history B T 2 2 h H\Z for which P(h) i. L Strategies for player 1: T, B 1 Strategies for player 2: (L, L), (L, R), (R, L), (R, R), 2 R L R 1 1 2 1 0 0 What is the difference between an action and a strategy? A strategy is a rule of action. In a simultaneous-move game actions cannot be conditioned on anything. Strategy = action in such games. 05.10.2009 4 G.B. Asheim, ECON5200-2 Every strategy profile determines a termi1 nal history (an outcome) O(T, (L, L')) (T, L) O(T, (L, R')) (T, L) O(T, (R, L')) (T, R) O(T, (R, R')) (T, R) O(B, (L, L')) (B, L) O(B, (L, R')) (B, R) O(B, (R, L')) (B, L) O(B, (R, R')) (B, R) B T 2 2 L R L R 1 2 1 1 2 1 0 0 Let Si denote the strategy set for player i. Definition 92.1 A Nash equilibrium of an extensive game w/perf. info. is a strategy profile s such that, for each i N, si S i , i (O( si , si )) i (O( si , si )). 05.10.2009 5 G.B. Asheim, ECON5200-2 Let Si denote the strategy set for player i. Define ui by ui(s) i(O(s)) for each si. 1 Definition 92.1 The strategic T form of an extensive game 2 w/perf. info. is the strategic L R game G N, (Si), (ui). 1 Lemma A strategy profile 2 s is a Nash equilibrium L L of an extensive game 1 , 2 T if and only if it is a 2 ,1 B Nash equilibrium of the strategic form of . 05.10.2009 G.B. Asheim, ECON5200-2 1 1 LR B 2 L R 2 1 0 0 R L 1, 2 1, 1 0, 0 2, 1 RR 1, 1 0, 0 6 2 Definition 97.1 The subgame of Subgame 1 an extensive game w/perf. info. that follows the history h H/Z B T is the extensive game where 2 2 H|h is the set of histories h for which (h, h) H, Z|h is the set of histories h for which (h, h) Z, P|h is defined by P|h(h) P(h, h) for all h H|h\Z|h, i|h(h) i|h (h, h) for all h Z|h. 05.10.2009 L R L R 1 2 1 1 2 1 0 0 N = {1, 2} H|T = {, L, R} Z|T {L, R} H|T\Z|T {} P() 2 1|T(L) 1, 1|T (R) 1 2|T (L) 2, 2|T (R) 1 7 G.B. Asheim, ECON5200-2 Subgame-perfect equilibrium Let Si|h denote the strategy set for player i in the game (h). Definition 97.2 The subgame perfect equilibrium of an extensive game w/perf. info. is a strategy profile s such that for each player i N and each h H\Z for which P(h) = i, si Si |h , 1 B T 2 2 L R L R 1 2 1 1 2 1 0 0 (B,(L,L)) is the unique subgame-perfect equil. (B,L) is the unique subgame-perfect equil. i (Oh ( si |h , si |h )) i (Oh ( si |h , si )). outcome 05.10.2009 G.B. Asheim, ECON5200-2 8 Lemma A strategy profile s is a subgame perfect equilibrium of an extensive game w/perf. info. if and only if s|h is a Nash equilibrium of (h) for all h H\Z. If the set of histories is finite, then the extens. game is finite. Proposition 99.2 Every finite extensive game w/perf. info. has a subgame perfect equilibrium. Proof is based on backward induction. Lemma 98.2 (The one deviation property) Let be a finite horizon game w/perf. info. A strategy profile is a subgame perfect equilibrium for if and only if each player i N and each h H\Z for which P(h) = i, player i cannot gain by changing his action only at h. 05.10.2009 G.B. Asheim, ECON5200-2 9 3 Definition of an extensive game w/simultan. moves (multi-stage games) N = {1, 2} (The set of players) H = {, T, C, (C,(B,B)), (C,(B,S)), (C,(S,B)), (C,(S,S))} 1 T C B S 2 3 , 1 0 , 0 B (The set of histories) 2 S 0, 0 1, 3 Z {T, (C,(B,B)), (C,(B,S)), (C,(S,B)), (C,(S,S))}(The set of terminal histories) H\Z {, C} P() 1, P(C) {1, 2} (Set-valued player function) 1(T) 2, 1(C,(B,B)) 3, 1(C,(B,S)) 0, 1(C,(S,B)) 0 , 1(C,(S,B)) 1 2(T) 2, 2(C,(B,B)) 1, 2(C,(B,S)) 0, 2(C,(S,B)) 0 , 2(C,(S,B)) 3 A() {a| (, a) H} {T, C} (Action sets) A(C) {a| (C, a) H} A1(C) A2(C) {B, S} {B, S} 05.10.2009 G.B. Asheim, ECON5200-2 10 Does actual behavior conform to subgame-perfection? The ultimatum game The take-it-or-leave game Game of trust (Game of punishment) The paradox of backward induction: Why should a player conform to backward induction at decision nodes where he/she knows that an earlier player has deviated from backward induction? 05.10.2009 G.B. Asheim, ECON5200-2 11 Results Backward induction identifies a unique strategy profile in a finite perfect information game without simultaneous moves and with no payoff ties. Such a strategy profile is the unique subg.-perf. equil. Observation Backward induction generalizes rationalizability to perfect information games w/out simultaneous moves. Questions How to define rationalizability for more general classes of extensive games. How to define equilibrium for extensive games so that equilibrium implies backw. ind. in perf. info. games. 05.10.2009 G.B. Asheim, ECON5200-2 12 4 Relationship between equilibrium concepts Proper equilibrium Myerson (1978) Nash equilibrium Extensive form perfect equil. Strategic form perfect equil. Quasi-perfect equilibrium Selten (1975) van Damme (1984) Weak sequential equilibrium Sequential equilibrium Reny (1992) Kreps & W. (1982) 05.10.2009 Selten (1975) Subg.-perf. equil. = Selten (1965) in multi-stage games 13 G.B. Asheim, ECON5200-2 Relationship between rationalizability concepts 1 11 2 2 2120 212 4 1 Rationalizability 22 3 0 111 110 Bernh. (1984) Pearce (1984) 0 1 003-1 -1 3 3 Schuhm. (1999) Permissibility Quasi-perfect rationalizability 0 (1994) Börgers 0 (1992) Brandenb. Dek. & Fud. (1990) Asheim & P (2005) Weak sequential rationalizability Sequential rationalizability Ben-Porath (1997) Dekel et al. (1999) 05.10.2009 1 T Extensive form rationalizability 2 2 Pearce (E.trica, 1984) and Battigalli & Sinichalchi (JET, 2002) S B T 2, 2 2, 2 CB 3, 1 0, 0 CS 0, 0 1, 3 Iterated elimination of weakly dominated strategies 14 G.B. Asheim, ECON5200-2 Forward induction 05.10.2009 Proper rationalizability C B S B 3, 1 0, 0 S 0, 0 1, 3 S B T 2, 2 2, 2 CB 3, 1 0, 0 CS 0, 0 1, 3 Iterated elimination of choice sets under full admissibility Asheim & Dufwenberg (EJ, 2003) G.B. Asheim, ECON5200-2 15 5 Incomplete information Imperfect information Complete information Perfect information Lecture 1 Strategic games Static Dynamic Lecture 2 Extensive games (multi-stage games) First application: Bargaining games Second application: Repeated games Extensive games with perfect information and with simultaneous moves (multi-stage games) Extensive games with perfect information and without simultaneous moves 05.10.2009 Lecture 1 Bayesian games Lecture 3 Extensive games (the general case) 16 G.B. Asheim, ECON5200-2 Bargaining games of alternating offers Version: Split-the-pie w/linear payoff function N = {1, 2} Players: Histories: I or (x0, R, x1, R, … , xt, R) (x0, R, x1, R, … , xt) (x0, R, x1, R, … , xt, A) (x0, R, x1, R, … ) II III IV Terminal histories: Histories of type III and IV Player fn.: P(h) = 1 for hist. of type I and II if h = or t is odd. P(h) = 2 for hist. of type I and II if t is even. Payoff fn.: 1(h) = 1txt if h Z and h is finite, where 0 1t 1. 2(h) = 2t (1xt) if h Z and h is finite, where 0 2t 1. 1(h) = 2(h) = 0 if h Z and h is infinite. 05.10.2009 17 G.B. Asheim, ECON5200-2 2 1 x 0 2 Player 1 is in the same situation in periods 0 and 2. A x0 , 1 x0 R 2 xt is 1’s share in period t. is a com- x1 1 R A x1 , (1 x1 ) 2 x2 2 1 1 Rubinstein’s model mon dis 2 x 2 , 2 (1 x 2 ) bargaining count factor 05.10.2009 G.B. Asheim, ECON5200-2 18 6 Proposition 122.1 The bargaining game of alternating offers (split-the-pie w/linear payoff functions) has a unique subgame perfect equilibrium. This equilibrium is characterized by: 1 1 2 x 1 demands x t 1 1 2 x 2 A 1-x 1 (1 2 ) R t 1 accepts if and only if x 2 1 1 2 x 1 1 1 (1 2 ) t 1 A 1 2 offers x 1 1 2 1 1 2 R 1 (1 1 ) 2 accepts if and only if 1 x t 2 1 1 2 0 0 0 1 05.10.2009 1x1 2(1-x1 ) 19 G.B. Asheim, ECON5200-2 Properties of the subgame-perfect equilibrium Efficiency. Agrement is reached at once. Stationary strategies. Each player proposes the same each time he offers, and the uses the same criterion to decide whether to accept an offer. BUT: Note that it is not imposed a priori that strategies must be stationary; this is a result of the analysis. First-mover advantage. Player 1 (who moves first) receives the biggest share: 1 2 1 1 If 1 2 , 1 (O ( s )) 1 1 2 1 2 Increased impatience reduces a player’s share of the pie. 05.10.2009 20 G.B. Asheim, ECON5200-2 Infinitely repeated games Prisoners’ dilemma D D 1, 1 C 3, 0 C 0, 3 2, 2 u2 3 2 1 Can repetition discipline 1 the players to cooperate? Deviating yields a short run gain. Deviating yields a loss of reputation that undermines future cooperation. 05.10.2009 G.B. Asheim, ECON5200-2 2 u1 3 Yes, if gain now PV of future loss 21 7 ● Definition 137.1 Let G N, (Ai), (ui) be a strategic game; let A i N Ai where, for each i N, Ai is compact and ui is continuous. A -discounted infinitely repeated game of G is an extens. game w/perfect information and simultan. moves, where H t0 At ( A0 {} contains only the empty history) Z A (the set of terminal histories) P(h) N for each nonterminal history h H ● ui (( a t )) (1 ) t1 t 1ui (a t ) ; (at) A and 0 1. Feasible payoff profile for G: v is such that there exists (a)aA with a 0 and aAa 1 s.t. v aA a u(a) . 05.10.2009 22 G.B. Asheim, ECON5200-2 (Average) payoff in a repeated game (1 )t 1 t 1vt Discounting p: Probability that the game ends. r: Rate of time preference. 1 p 1 r Why multiply with (1 )? Suppose that vt v for all t. Then (1 )t 1 t 1vt (1 )(1 2 )v v Special case 1: v1 v , but vt 0 for all t 1. (1 )t 1 t 1vt (1 )v Special case 2: v1 0 , but vt v for all t 1. (1 )t 1 t 1vt v 05.10.2009 G.B. Asheim, ECON5200-2 23 Strategies as machines Example: Simple strategy profiles (Abreu, 1988) Consider the following paths: (a (0)t ) a(0)1 , a (0) 2 , a(0) 3 , a (0) 4 , (a (1)t ) a(1)1 , a(1) 2 , a (1)3 , a(1) 4 , (a (n)t ) a(n)1 , a (n) 2 , a(n)3 , a (n) 4 , Rules: Start with (a(0)t). If i N (and only i) deviates from (a(j)t), then (re)start (a(i)t). Write: 05.10.2009 (a (0)t ), (a (1)t ), , (a (n)t ) G.B. Asheim, ECON5200-2 24 8 Definition 153.1 (The one-deviation property) A strategy profile is a subgame perfect equilibrium for the -discounted infinitely repeated game of G if and only if there is no history after which a player can gain by deviating for a single period. Examples of infinitely repeated games Proof: (Only if ) Trivial. (If ) If there is a profitable multi-period deviation of infinite duration, then there is a profitable multi-period deviation of finite duration. (This is due to the discounting: What happens in the distant future is of little importance, implying the deviation is profitable even if truncated.) But if there is a profitable multi-period deviation of finite duration, then there is a profitable one-period deviation (see proof of Lemma 98.2). 05.10.2009 G.B. Asheim, ECON5200-2 25 u2 Set of subgame Lemma 153.3 The set of payoff perfect equilibrium profiles that can be realized in subg. payoffs perf. equilibria for the -disc. infinitely repeated game of G is compact. m(2) Corollary Consider the -disc. u1 m(1) infinitely repeated game of G. For each player i N, there is (a) a minimal payoff m(i) that can be realized as a subg. perf. equilibr. payoff, and (b) a subg. perf. equilibr. path (p(i)t) s.t. ui (( p (i) t )) m(i) Proposition 154.1 Consider the -disc. inf. repeated game of G. Then (at) is a subg. perf. equilibr. path if and only if the simple strategy profile (a(0) t ), ( p(1) t ), , ( p (n)t ) is a subgame perfect equilibrium. 05.10.2009 G.B. Asheim, ECON5200-2 26 Let G N, (Ai), (ui) be a strategic game. Player i's minmax payoff in G, vi, is the lowest payoff that the other players can force upon i: vi min max ui ( ai , ai ) a i Ai ai Ai . A payoff profile w in G is enforceable if wi vi for all i N. A payoff profile w in G is strictly enforceable if wi vi for all i N. If a A is an outcome of G for which u(a) is (strictly) enforceable, then we refer to a as a (strictly) enforceable outcome of G. 05.10.2009 G.B. Asheim, ECON5200-2 27 9 A Folk Theorem Proposition 151.1 Let a be a strictly enforceable outcome of G. Assume that there is a collection (a(i))iN of strictly enforceable outcomes of G such that for each player i N we have u2 ui(a) ui(a(i)) and u (a (1)) ui(a(j)) ui(a(i)) for all j i. Then there exists u (a ) 1 such that for all u(b , p ) u (a (2)) there is a subg. perf. equilibrium u1 of the -discounted infinitely u( p , b ) repeated game of G that generates the path (at) where at a for all t. 1 2 1 05.10.2009 G.B. Asheim, ECON5200-2 2 28 10