GAME THEORY 1) Strategic Form or Normal Form Game 2) Extensive Form Game Normal (Strategic) Form Game Provide a ‘reduced’ summary of a game. Normal form game is defined by - a set of players - a set of strategies Common Knowledge - a set of payoffs } Description is common knowledge: each knows that another person knows that he knows and so on ……….that he knows …..(e.g. the sets of players, strategies and payoffs) Players are fully rational, maximize expected utility given subjective beliefs and beliefs are modified when new information arrives (e.g. according to Bayes’ law) Example: Matching Pennies Column Head Tail Head Row Tail 1 , -1 -1 , 1 -1 , 1 1 , -1 No Nash Equilibrium in the game The entries in each cell indicate payoffs. The first number is Row’s payoff, the second is Column’s payoff. Example: The Prisoner’s Dilemma game Column Cooperate Defect Cooperate Row Defect 0,0 1 , -5 -5 , 1 -2 , -2 Each player has an incentive to defect regardless of what the other does → Defect is a dominant strategy Strategy profile (Defect, Defect) is the unique Nash equilibrium in pure strategy. However, cooperation gives higher payoffs Note: the first element in a vector of strategies denote the first player’s (Row’s) strategy, the second element denotes the second player’s (Column’s) strategy. Solution Concepts Pure strategies: Strategies in which a choice is made with probability of 1 Mixed strategies: Random strategies i.e. choices are made with probability > 0 and < 1 If R is the set of strategies available to Row, the set of mixed strategies is the set of all probability distributions over R where the probability of playing strategy r in R is pr (similarly for Column) If Row’s subjective probability distribution over Column’s set of strategies is ΠC, πc is the prior probability that Column’s choice is c, and Column’s subjective probability over Row’s set of strategies is ΠR, πr is the prior probability that Row’s choice is r Note that ∑πc = 1 and ∑πr =1 c∈C r∈R - Row’s objective is to choose pr for each r to maximize Row’s expected payoff ∑ ∑ pr πcur (r,c) r∈R c∈C - Column’s objective is to choose pc for each c to maximize Column’s expected payoff ∑ ∑ pcπr uc (r, c) c∈C r∈R Nash Equilibrium Consistency Requirement: each player’s belief about the other player’s choices coincides with the actual choices the other player intends to make. Rational Expectations: Expectations coincide with actual frequencies. Definition: Nash Equilibrium A Nash equilibrium consists of vector of prior belief probabilities (πr, πc) over strategies and vector of probabilities (pr, pc) of choosing strategies such that 1. The beliefs are correct: pr = πr and pc = πc for all r and c and 2. Each player is choosing probability pr or pc (Row player is choosing pr for all r and Column player is choosing pc for all c) so as to maximize his own expected utility given his belief. Alternatively, with two strategies for each player, a Nash equilibrium is a pair ( p̂r , p̂c ) such that Eu r ( pˆ r , pˆ c ) ≥ Eu r ( pr , pˆ c ) for every p r ∈[0,1] Euc ( pˆ c , pˆ r ) ≥ Euc ( pc , pˆ r ) for every pc ∈[0,1] *Here the distinction between beliefs and actual frequencies is blurred. Definition: Nash equilibrium in pure strategies A Nash equilibrium in pure strategies is a pair * * (r , c ) such that ur(r*, c*) ≥ ur(r, c*) for all Row’s strategies r and uc(r*, c*) ≥ ur(r*, c) for all Column’s strategies c In a Nash equilibrium, no player has incentive to deviate unilaterally from a Nash equilibrium strategy. → A ‘Rest’ Point Example: Battle of the sexes Ruth Top (Opera) Bottom (Football) Chris Left (Opera) Right (Football) 2, 1 0, 0 0, 0 1, 2 Two Nash equilibria (Top, Left) and (Bottom, Right) in pure strategies. Finding the solution by maximization Row’s problem is to maximize expected payoff Max pt, pb pt [2pl + 0pr] + pb [0pl + 1pr] such that pt + pb =1 pt ≥ 0 and pb ≥ 0 Lagrangian L = 2ptpl + pbpr − λ (pt + pb − 1) − µ1 pt − µ2 pb Differentiate wrt. pt and pb yield ∂L = 2pl − λ − µ1 = 0 ∂pt ∂L = 2pr − λ − µ2 = 0 ∂pt Since we already know the Nash strategy profiles in pure strategies. We will consider only the case where pt > 0 and pb > 0 (i.e. µ1 = 0 and µ2 = 0) → pl = 1/3 and pr = 2/3 Similarly, can solve Column’s maximization problem → pt = 2/3 and pb = 1/3 Diagram: Best response functions and equilibria Interpretation of Mixed Strategies - Some games have strategies that are more sensible as mixed strategies than others - For any mixed strategy equilibrium, if one player believes the other will play the equilibrium mixed strategies, then he is indifferent as to whether he plays any other mixed or pure strategies as long as the set of strategies that form a basis for that strategy belongs to that of his equilibrium mixed strategy. - Can interpret mixed strategies as being random chance of meeting a person from a population who plays a pure strategy but that strategy is not known. - Alternatively, individuals’ choices may be deterministic e.g. depend on moods but there is still uncertainty for other players. Repeated Game - A repeated game is a repetition of a one-shot game - The strategy space of the repeated game is much larger than that of a one-shot game Example: Prisoners’ Dilemma In the long run, it is in the best interest of both players to get (Cooperate, Cooperate) as a solution. Finitely Repeated Game Last round: They are essentially playing a one-shot game. Hence, (Defect, Defect) Next to last: Still no long-run benefit to encourage cooperation. Again, (Defect, Defect) Same logic applies to other previous rounds *Hence, in all periods, (Defect, Defect) is the solution. Infinitely Repeated Game Consider a trigger strategy (or punishment strategy) which specifies Cooperate on the current move unless the other player defected on any previous rounds Assume that each player discount their future payoff by 1+ρ, ρ is the discount factor - Suppose, they cooperate up to move at period T. At T, if a player decides to defect, he gets 1 and a stream of payoff -2 forever 1 Expected payoff from deviation = 1 − 2 ( ρ ) - Suppose, they cooperate up to move at period T. At T, if a player decides to defect, he gets 1 and a stream of payoff -2 forever Expected payoff from cooperation = 0 1 Continue to choose Cooperate if 0 > 1 − 2 ( ρ ) i.e. ρ > ½ Refinements of Nash Equilibrium - Nash (1950) shows that with finite number of players and finite number of pure strategies, a Nash equilibrium exists (at least in mixed strategy)1 - But we may have multiple Nash equilibria. - Refinements of Nash equilibrium are criteria used to choose among Nash equilibria. Dominant Strategies Let r1 and r2 be two of Row’s strategies. r1 strictly (weakly) dominates r2 if payoff of r1 is strictly larger for all choices Column might make (at least as large for all choices and strictly larger for some choices) A Dominant Strategy Equilibrium is a choice of strategy by each player such that each strategy dominates every other strategies. Elimination of Dominated Strategies Given a game, one first eliminates all strategies that are dominated and then calculate the Nash equilibria. Row Top Bottom h 1 Left 2, 2 2, 0 Column Right 0, 2 1, 1 Also, with infinite continuum of strategies, a Nash can be shown to exist in pure strategy. Two Nash equilibria (Top, Left) and (Bottom, Right) in pure strategies. If Row assumes that Column will never play his weakly dominated strategy then the only equilibrium for the game is (Bottom, Right). Although, elimination of strictly dominated strategies is generally acceptable, elimination of weakly dominated strategies may not be sensible. Sequential Game Some choices are made sequentially and players may know other players’ choice before making his own. Column Left Right 1 , 10 1, 5 Row Top 0, 0 2, 1 Bottom h Two pure Nash: (Top, Left) and (Bottom, Right) Extensive Form Game Provides ‘extended’ description of the game. Often represented by a game tree indicating the choices that each player can make at each stage of the game. 1, 10 Left Column Top 1, 5 Row moves first, Right Column moves after Row observing Row’s move. Left 0, 0 Bottom Column Right 2, 1 Each decision point is a node A subgame consists of a node and all strategies and payoffs available from then on. A Subgame Perfect Equilibrium (SPE) is Nash equilibrium in every subgame. We can usually solve for SPE using ‘backward induction’ Note that (Top, Left) is eliminated. Extensive form can be used to model situations where some moves are sequential, some are simultaneous. Information Set is the set of all nodes that cannot be differentiated by the agent. 1, 10 Left Column Top Right 1, 5 Left 0, 0 Right 2, 1 Row Bottom Column cannot tell where he is, as if we are in simultaneous move game Column Example: A simple bargaining model - N- stage game with 2 players A and B who have $1 to divide between them - 1st stage, A makes an offer, B can accept or wait - 2nd stage, B makes an offer, A can accept or wait - And so on to Nth stage Solving for SPE Suppose that A’s and B’s discount rates are α and β If N=3, at last stage A offers B $0 and A gets $1. Hence, at 2nd stage B can offer A $α and gets $(1- α). At 1st stage, A offers B $β(1- α) and takes $(1- β(1- α)) for himself. Game with Incomplete Information Players may not have complete information about other players, e.g. valuation of good, preferences, etc. Agent’s type describes all uncertainty that one agent may have about another. A Bayes’ Nash equilibrium is a set of strategies for each type of player such that the expected value of each type of player is maximized given the strategies of other players. Example: Battle of the sexes (modified) - Ruth knows Chris’ preference but Chris does not know hers. - Ruth may like to be with him (type 1) or to be alone (type 2). - Chris attaches a probability ρ to Ruth being a type 1 - Ruth knows Chris’s estimate of her preference (know ρ) This is assumption of ‘common knowledge’ - If Ruth is type 1, her preference is as before - If she is type 2, she gets 0 if she goes anywhere with Chris and gets 1 if goes alone to football and 2 if goes alone to opera. - Chris’ preference is as before Solving for pure strategy Bayes-Nash equilibrium First, turn incomplete information game to imperfect information game (Draw tree) The set of pure strategies for Chris is {O, F} For Ruth is {(O,O), (O,F), (F,O), (F,F)} (1st entry for type 1) 1) Suppose Chris plays F, in a best response, Ruth plays F if she is type 1 and O if she is type 2. i.e. (F,O) Chris’ expected payoff is 2ρ + 0(1- ρ) = 2ρ For this profile ((F,O), F) to be a Bayes’ Nash equilibrium, Chris must actually prefer F to O i.e. (expected payoff from F) 2ρ ≥ 1- ρ (expected payoff from O) → ρ ≥ 1/3 2) If Chris plays O, Ruth will play (O,F). For ((O,F), O) to be a Bayes’ Nash, Chris must actually prefer O to F, that is we must have 1ρ + 0(1- ρ) ≥ 0ρ + 2(1- ρ) → ρ ≥ 2/3 Two pure-strategy Bayes-Nash equilibria when ρ ≥ 2/3. i.e. ((F,O), F) and ((O,F), O) One pure-strategy Bayes-Nash when 1/3 ≤ ρ < 2/3, i.e. ((F,O), F) No pure-strategy Bayes-Nash when ρ < 1/3 Discussion Different beliefs about distribution of various types lead to different optimal behabiour. Ladyard (1986): nearly any pattern of behaviour can be supported by some pattern of beliefs.