Lecturer Parmar BaG NUS Econ Chapter 6 Game Theory a 10. Games in strategie form. 2. Dominance and iterated dominance. The Prisoners’ Dilemma. 3. Weak dominance. 4. Nash Equilibrium in Pure Strategies. 5. Nash Equilibrium in Mixed Strategies. 6. 1. 8 9. Games in extensive form, ". Refinement of Nash Equilibria - Subgame Perfect Equilibria, . Games of Incomplete Information. Repeated Games and the Folk Theorem. Some game-theoretic models of oligopoly. References L Jehle & Reny, ibid, ch 9 Dna, A. AND S. SKEATH (1999): Games of Strategy. Norton. (A non- technical introduction to strategic interaction with lots of interesting ex- amples) FUDENBERG, DREW AND JEAN TrROLE (1991): Game Theory. MIT Press. Gipzons, Roserr (1992): A Primer in Game Theory. Harvester Wheat- sheaf. Kreps, Davin (1990): A Course in Microeconomic Theory. Princeton University Press. 27Birkbeck Economics 28 OsnorNe, MARTIN AND ARIEL RUBINSTEIN (1994): A Course in Game Theory. MIT Press. ‘Watson, J. (2002): Strategy. Norton. 6.1 Games in Normal (or Strategic) Form An n-person game in strategic form (or, normal form) has 3 essential elements 1. A finite set of players 1 = {1,2,...,n} 2. For each player i, a finite set of strategies S;. Let s = (s1,82,..-8n) denote an n-tuple of strategies, one for each player. This n-tuple is called a strategy combination or strategy profile. The set $ = 5; x Sx... Sq denotes the set of n-tuple of strategies. 3. For each player i, there is a payoff function P;: S + R, which associates with each strategy combination (81,82)... 8), & payofl Pi(s),82,.-.,$n) for player i. Since we have one such function for each player i, in all we have n such functions. Note: If the typical player is denoted by i, we sometimes denote all other players (her ‘opponents’) by the (vector) —i. Hence, a typical strategy profile is denoted as (5;, 3-3). 6.2 Dominance and Iterated Dominance Definition 7 The (pure) strategy si is (strictly) dominated for player i if there exists 6} € S; such that u(s},3«) > wi(si,8-1) 8s. If, in a particular game, some player has a dominated strategy, it is reasonable to expect that the player will not use that strategy. Prisoners’ Dilemma The Prisoners’ Dilemma game below is an example of a game where a single round of elimination of dominated strategies allows us to solve the game. Player 2 Confess Not Confess Confess [-5,-5 0-8 Player 1 Not Confess| -8,0 -1,-1Birkbeck Economies 29 How would you play this game? In general there may be successive stages of elimination, This method of narrow- ing down the set of ways of playing the game is described as iterated dominance. If in some game, all strategies except one for each player can be eliminated on the criterion of being dominated (possibly in an iterative manner), the game is said to be dominance solvable. Player 2 Left Middle Right Top[ 43 | 2,7 | 04 Phyer1 Middie[ 5,5 | 5,1 | -4-2 Bottom| 3,5 | 1,5 | -1,6 ‘We can eliminate dominated strategies iteratively as follows. 1. For player 1, Bottom is dominated by Top. Eliminate Bottom. 2. In the remaining game, for player 2, Right is dominated by Middle, Elim- inate Right. 3, In the remaining game, for player 1, Top is dominated by Middle. Elimi- nate Top. 4, In the remaining game, for player 2, Middle is dominated by Left. Elimi- nate Middle. This gives us (Middle,Left) as the unique equilibrium. 6.3 Weak Dominance Definition 8 The (pure) strategy s; is weakly dominated for player i if there exists 8; € Sy such that us(sh.s:) > wi(si,8-i) Vs-i, with strict inequality holding for some si. Player 2 Left Right Top[ 51 | 40 Plyert Middle! 6,0 | 3,1 Bottom| 64 | 4,4 Here, for player 1, Middle and ‘Top are weakly dominated by Bottom. Eliminate Middle and Top. The equilibria are (Bottom, Left) and (Bottom, Right).Birkbeck Economics 30 6.4 Nash Equilibrium However, for many games the above criteria of dominance or weak dominance are unhelpful - none of the strategies of any player might be dominated or weakly dominated. ‘The following is the central solution concept in game theory. Definition 9 (Nash Equilibrium in Pure Strategies) A strategy profile (37, 8°) is a Nash equilibrium if for each player i, wis}, s%s) > u(sirsts) ¥ si € Si. Player 2 Left Middle Right Top[ 04 | 40 | 53 Plyer1 Middle[ 4,0 | 04 | 53 Bottom [3,5 | 35 | 66 ‘The only Nash Equilibrium in this game is (Bottom, Right). A Nash equilibrium is a strategy combination in which each player chooses a best response to the strategies chosen by the other players. In the Prisoners’ Dilemma, the case in which each prisoner confesses is a Nash equilibrium. (If there is dominant strategy equilibrium, it must be a Nash equilibrium as well) In general, we can argue that if there is an obvious way to play the game, this must lead to a Nash equilibrium. Of course, there may exist more than one Nash equilibrium in the game, and hence the existence of a Nash equilibrium does not imply that there is an ‘obvious way to play the game’. How to look for Nash Equilibria in simple games? Consider the following game, known as the “battle of the sexes.” Wile Football Opera Football [2,1 | -1-1 Husbond Opera | -1-1 12 In essence, we must examine all strategy combinations, and for each one, check to see if the Nash equilibrium conditions are satisfied. Consider the Battle of the Sexes depicted in the figure above. a, Start with the strategy combination (Football, Football).Birkbeck Economics 31 (a1) Look at the payoffs from the husband’s viewpoint. If the wife goes to the football match, is football optimal for him? Yes, because 2 > —1. (a2) Now look at the payoffs from the wife's viewpoint. If the husband goes to the football match, is football optimal for her? Yes, because 1 > ~1. Since the answer is ‘yes’ in both (al) and (a2), (Football, Football) is a Nash equilibrium. b, Next, consider the strategy combination (Football, Opera). (b1) Look at the payofs from the husband’s viewpoint. If the wife goes to the Opera, is football optimal for him? No, because by going to the football he gets -1, and he could do better by going along to the Opera which would fetch 1, For this strategy combination, the Nash equilibrium condition does not hold for the husband. (b2) We could look at this strategy combination from the wife's viewpoint, but given that the Nash condition does not hold in (b1), we need not really bother. In short,(Football, Opera) is not a Nash equilibrium. ¢. Next, consider the strategy combination (Opera, Opera) Checking (cl) and (c2), this turns out to be a Nash equilibrium. 4, Next, consider the strategy combination (Opera, Football) This is not a Nash equilibrium. In sum, there seem to be two Nash equilibria in this game, namely (Opera, Opera) and (Football, Football) IE the game had three strategies for each player, there would be 9 possible strategy combinations for us to check for Nash equilibria. 6.5 Nash Equilibrium in Mixed strategies Some games do not seem to admit Nash equilibria in pure strategies. Consider the game below called “matching-pennies.” Player 2 Heads Tails Heads [1-1 | -1,1 Playert Tails{ -11_ [| 1-1 Notice that this game seems to have no Nash equilibria, at least in the sense that they have been described thus far. But, in fact it does have a Nash equilibrium in mixed strategies. The first stage in the argument is to enlarge the strategy space by constructing probability distributions over the strategy set 5).Birkbeck Economics 32 Definition 10 (Mixed Strategy) A méced strategy a; is a probability distri- bution over the set of (pure) strategies. In the matching pennies game, a pure strategy might be Heads. A mixed strategy could be Heads with probability 1/3, and Tails with probability 2/3 Notice that a pure-strategy is only a special case of a mixed strategy. A Nash equilibrium can now be defined in the usual way but using mixed strategies instead of pure strategies. Definition 11 (Nash Equilibrium) A mized-strategy profile (07, Nash equilibrium if for each player i, u(o.o2) > wlsino-i) Vai € Si. ‘The essential property of a mixed strategy Nash Equilibrium in a 2 player game is that each player's chosen probability distribution must make the other player indifferent. between the strategies he is randomizing over. In a n player game, the joint distribution implied by the choices of each player in every combination of (n—1) players must be such that the n-th player receives the same expected payoff from each of the strategies he plays with positive probability. Once we include mixed strategy equilibria in the set of Nash Equilibria, we have ‘the following theorem, ‘Theorem 5 (Existence) Bvery finite-player, finite-strategy game has at least one Nash equilibrium. Clearly, if a game has no equilibrium in pure strategies, the use of mixed- strategies is very useful. However, even in games that do have one or more pure strategy Nash equilibria, there might be yet more equilibria in mixed-strategies. For instance, we could find a mixed-strategy Nash Equilibrium in the Battle of the Sexes. 6.6 Games in Extensive Form ‘The extensive form is particularly useful when the interaction is principally dynamic. It provides a clear description of, say, the order in which the players move, what their choices are, the information that each player has at each stage and so on. The extensive form is often represented by a game tree. ‘The description involves the following elements. 1. A finite set of players J = {1,2,...,n}. In addition, there may be an additional player to capture the uncertainty, called Nature (denoted by N).Birkbeck Economics 33 2. A game tree consists of a set of nodes with a binary precedence relation- ship. Think of it as a configuration of nodes and branches. A node (more accurately, a decision node) represents a point at which a player (or ‘na- ture’) must choose an action. The choice of an action takes that player down a branch to a successor node. The idea of an initial node and termi- nal node(s) is obvious in this context. A game tree is a configuration of nodes and branches running from the initial node to the terminal nodes, with the restriction that there be no closed loops in the tree. 3. One player or (nature) is assigned to each node. This is just a way of specifying which player must choose (take an action) at that node. 4, For each node, there is a finite set A of available actions, which lead to the immediate successor nodes of that node. 5, Each players nodes are partitioned in to information sets, which measures the fineness of the information available to that player when s/he chooses an action. If two nodes lie in the same information set, the player knows that s/he is at one of those two nodes but does not know which one. 6. An assignment of payoffs, one for each player, at each terminal node. 7. A probability distribution over nature’s moves. ‘The notion of a strategy is fairly straightforward in a normal form game. How- ever, for an extensive form game, it is a little bit more complicated. To under- stand what a typical element of the strategy set S; for player i is, let h be a typical information set for player é, and A(h) the set of actions available at that information set. A (pure) strategy for player i specifies which action she must take at each of her information sets. The set of all strategies for that player is given as S; = 11,A(h). Note: It is very important to understand the distinetion between actions and strategies for an extensive form game. A strategy is a complete plan of actions.Birkbeck Beonomics 34 6.7 Actions and Strategies 6.7.1 Gamel Payot 101 Payot to 2 @) a) @® @) Game 1 In game 1, player 1 has a total of 4 actions: A, B, C, D. However, player 1 moves at three different: nodes - at each of these nodes 1 has 2 possible actions. ‘Thus the total number of strategies for player 1 is 2 x 2.x 2=8. For example, one of the strategies of player 1 is : play A initially, and then if player 2 plays L, then play C and if player 2 plays R, then play D. Such a strategy is written as ACD. The eight strategies of player 1 are: 1acc 2ACD 3. ADC 4.ADD 5. BCC 6 BOD 7. BDC 8 BDD Player 2 has 2 actions - L and R, and 4 strategies: LLLBirkbeck Economics 35 2LR BRL 4.RR 6.7.2 Game2 Payot to Payoff to ) (0) GQ @ @) Game 2 In game 2, on the other hand, player 1 moves only at two different information sets (each node is also a trivial information set). Thus 1 has only 4 strategies: AC, AD, BC, BD. Player 2 only moves at one information set - thus for 2, actions and strategies coincide. Player 2 has only 2 actions as well as 2 strategies: L and R. 6.8 Analyzing Extensive Form Games Now that we can write down the strategies for players, how do we identify the Nash equilibria of such games? We do so by converting extensive form games to normal form games. ‘To every extensive form game there is a corresponding strategic form game. But a given strategic form game can, in general, corre- spond to several different extensive form games, The normal form for game 2 is as follows:Birkbeck Economics 36 Player 2 LR AC [0,0 Puyert AD [1,1 BC {1,0 BD [10| From this, it is easy to see that there are 2 pure strategy Nash equilibria: (AC, R) and (AD, L). 6.9 Equilibrium Refinement ‘The use of Nash equilibrium as a guide to the likely outcome of a game-theoretic situation requires some justification. It is hard to convey the nuances in a brief guide but some remarks follow. First, one could argue that if there is a self evident way to play the game, it must be a Nash equilibrium. Two, we could view the Nash equilibrium as the "outcome? of preplay negotiation: if players carry out pre-play negotiation, we would like the agreed-to action to be self- enforcing in the sense that given that others will fulfil their part of the deal, it is in your interest to not deviate. Three, we can think of Nash equilibrium as the outcome of evolution, or of learning, Regarding the first one of these remarks, it suggests that being Nash is a nee- essary condition for a strategy combination to be a self-evident way to play the game but may not be sufficient. In any game where the game-form admits more than one equilibrium, we need other criteria to narrow down the equilibrium set This procedure is described as refinement. ‘The actual notions used to refine the set of equilibria are varied and do not all have universal acceptability in the discipline. ‘Weak Dominance We may choose to eliminate equilibria that involve the use of weakly dominated strategies. Subgame perfect equilibria (SPE), and the issue of credibility A sub- ‘game is a game consisting of a node which is a singleton, that node’s successors and the payoffs at the associated end-nodes. A strategy combination is a subgame perfect equilibrium (SPE) if it is a Nash ‘Equilibrium (NE) for the entire game and the implied strategies for any subgame are a NE for that subgame. We will discuss some illustrative examples in the lecture. Here is a simple example: Consider the following extensive form game.Birkbeck Economics 37 (rons a (3) (%) @) Game 3 ‘The normal form is given by Player 2 tt nt tn nn T | 1,1 | 2,0 | 1,1 | 2,0 N{0,2]0,2 [3,3 [3,3 ere are 3 pure strategy Nash Equilibria: (T,tt), (N,nn), (N,tn). However, in the subgame on the left hand side, the (trivial) Nash equilibrium (in this subgame only one player plays - so 2's optimal strategy in the subgame is trivially the Nash equilibrium for the subgame) is ‘t’. In the subgame on the right hand side, the Nash equilibrium is ‘n’. Thus the only Nash equilibrium that induces Nash equilibria in all subgames is (N,tn) - this is therefore the only subgame perfect Nash equilibrium, ‘The issue of subgame perfection is closely linked to those of credible threats and of credible promises. Very crudely speaking, consider a Nash equilibrium which is not subgame perfect. That implies there must be a subgame such that the strategy over that, remaining subgame is not a Nash equilibrium for that subgame. So if by chance we end up at that sub game, we do not expect that the players will ind it profitable to stick to that °portion’ of the strategy: if not, that portion of the strategy is not credible. ‘The issue is best discussed through some examples such as entry deterrence, credibility of government policy ete. Other Refinements There are other kinds of refinements that result in, say, sequential equilibria, trembling-hand perfection, etc. but constraints of time will prevent us from exploring these in any detail. ‘Those interested in these are advised to read more extensively in these areas.Birkbeck Economics 38 6.10 Games of Incomplete Information ‘We have implicitly assumed that (for the extensive form) the players know what ‘the game-tree looks like, they know that the other players know what the game tree looks like, and so on, This recursion is formalized as common knowledge. Definition 12 (Common Knowledge) Information is common knowledge if itis known to all players, each player knows that they all know it, each of them knows that all of them know that all of them know it, and so on. ‘The players might or might not, have full information about all aspects of a game that they are about to play. Below we consider the possibilities. Perfect information There is no uncertainty arising from moves of nature. When such uncertainties exist, the situation is one of imperfect information. Complete information All players know all the relevant information about each other, including payofis that each receives from various outcomes. When this is not the case, so that players might not know about others’ payoffs. This creates a problem in analyzing the game. We now need to know about a player's beliefs about other players preferences, his beliefs about the beliefs of others about his preferences, his beliefs about the beliefs of others about his beliefs and s0 on. This complicates the situation thoroughly. However, we do have an approach that avoids this problem. ‘This was the contribution of Harsanyi (one of the three economists who won the Nobel Prize for their work in Game Theory in 1994 - the other two were Nash and Selten).. In this approach, it is assumed that each player's preferences are determined by ‘the realization of a random variable. The actual realization is known to only the player in question, but the distribution of the random variable is common knowledge. The random variable for a player is labelled his “type.” Since the distribution of each players types are common knowledge, the game becomes exactly like a game of imperfect information where nature makes the first move and chooses realizations of types (each player observes the realization of only his own type). A game with this structure is called a Bayesian game. ‘What is an equilibrium of such a game? A Bayesian Nash equilibrium is a Nash equilibrium of this imperfect information game. It is a set of (type-contingent) strategies such that each player maximizes his expected utility, contingent on her type, and taking the other player’s type-contingent strategies as given. 6.11 Repeated Games and the Folk Theorem Suppose a particular game such as the ‘Prisoners’ Dilemma’ is played a large number of times. Can we say something about the behaviour of players in such ‘supergames’ that is not obvious in the analysis of the one-shot game?Birkbeck Economics 39 ‘To anticipate the argument, we will try and establish that if the game is repeated a large number of times, we cannot rule out some outcomes that are clearly unlikely in case the game was played just once. First, we need to have an appropriate notion of payoffs in the supergame, or more accurately, the relationship between the payoff in the supergame and in the one-shot constituent game (now called the stage game). The average payoff over the supergame is some aggregate measure of the payofis from the stage games, with later payoffs possibly discounted for the lag with which they will become available. The ‘folk theorems’ for repeated games are usually some variant of a simple idea, namely that, if the players are sufficiently patient, then any feasible, individually rational payoff combination can be supported as a Nash equilibrium (the two adjectives need some explanation). One conclusion that emerges from this is that outcomes that were ruled out in the one-shot game (eg. cooperation in the Prisoners’ Dilemma) can be ‘sustained’ in the associated repeated game. ‘The intuition for this might run as follows: if players deviate from say, an agreement to cooperate they could be punished by the other(s) in subsequent rounds which would cause loss of utility in every subsequent period. (How hard they can be punished depends on what their reservation utility is; how costly this future punishment seems to them depends on their discount factor). Fearing this reduction of utility in later rounds (which would reduce the average payoff) breeds compliance or cooperation. ‘This has considerable application in models of say, lender-debtor interaction, product quality, entry deterrence, etc. This idea is quite useful in understanding why cooperation is sustained in the real world, 6.12 Some game-theoretic models of oligopoly ‘We now apply some of the techniques developed so far to model the oligopolistic interaction between firms. Since the aim here is to see how notions such as Nash equilibrium relate to oligopoly, the details of the underlying market are kept as simple as possible. For instance, we readily assume that firms are identical, demand curves are conveniently linear, and so on. More general versions of the examples presented here can easily be generated, 6.13 Bertrand Equilibrium: Nash equilibrium in prices We first consider a case with two identical firms, each producing the same homo- geneous product. Production is costless. The firms face a standard downward- sloping demand curve, and their individual share of the total market demand depends on the combination of prices charged by them. Put simply, the firm with the lower price captures the entire market; if they charge the same price, they get half the market each. The prices are chosen ‘simultaneously’. Let us see how we can model this as a game.Birkbeck Economics 40 Label the firms as firm 1 and 2. The strategy set for each firm can be thought of as the set of non-negative prices, p. The payoff function measures the profit made by a firm for any given strategy combination. This can be written as Ti(Pisp3), for i= 1,2 and j #4. In this case, (Pi, P2) = Pig aq ifnpm 4/2 fp, = pe Firm 2's payoff function is defined symmetrically, 2(p1,p2) = page. Here a ‘Nash equilibrium is a pair of prices (pj,p3) such that 1. (Pj, P2) 2 71(P1,P3) Yer, and 2. mo(pi, ph) 2 ™(Pi,P2) Vee. How do we compute the equilibrium? hat is, how do we find the prices that satisfy conditions 1 and 2? Let us look at some candidate solutions, Case 1 pj > pi > 0. Does this satisfy 1 and 2? At this configuration of prices, 7, = 0 and m2 > 0. (Why?) If firm 1 had chosen some other price, say p}—¢, it could have captured the entire market. Hence the solution could not be of this form, Case 2 pt =p >0. ‘This, too, is not an equilibrium. Either firm, if it had chosen a lower price, could have captured a much larger share of the market and made higher profits. (Why?) Case 3 pj > ph = 0. ‘This cannot be an equilibrium because firm 2 could have charged a slightly higher price without losing its market share and made strictly greater profits. By the process of elimination, we reach the conclusion that the unique Nash equilibrium is given by PL= PR ‘The paradoxical feature of this solution is that, in equilibrium, the firms make no (super-normal) profits. This seems unrealistic. As an exercise comment on the outcome in the above game if each firm produces at positive, but constant, marginal cost c. What if the marginal cost is not the same for the two firms?Birkbeck Economics 41 6.14 Cournot equilibrium: Nash equilibrium in quan- tities Assume, as in the previous example, that there are two identical firms produc- ing costlessly but, now suppose the firms choose, not prices, but quantities to supply to the market. Assuming that the market clears, each firm obtains the market clearing price for each unit of the good supplied. To obtain the payoff- function we now need to incorporate the market demand function explicitly. For algebraic simplicity, we use a very simple demand function, 120~p. Let qi and gp be the output levels supplied by firm 1 and 2, respectively. The market clearing condition amounts to q: + q = 9. For any given strategy- combination (91,92) the market price is p = 120 — (q1 + g2). Therefore, the ayoff-function (profits) for firm 1 is given as 71(41, 42) = pen = (120-1 — a2) and likewise for firm 2. A Nash equilibrium is defined in the usual way: it is a pair of strategies (namely, quantity choices) (93, 93) such that 1. m(gj,@3) 2 ™1(q1595) Var, and 2. ma(gi,q2) > m(qj,a2) Van. Since in this case the payoff functions are differentiable in the strategy variable, ‘we can obtain the Nash equilibrium by the standard optimization techniques. First, compute the profit maximizing output choice for each firm taking the rival's output as parametric. Setting ‘we get, q1 = 60 —.5g2. This is the “reaction function” for firm 1. Similarly, we can obtain a reaction function for firm 2, g2 = 60 — 54 We then ‘solve’ the two reaction functions to obtain the Cournot-Nash solution, For this example, it turns out to be gf = qj = 40. ‘As an exercise generalize the above for the case where the are n firms. 6.15 Duopoly with sequential moves: the Stackel- berg equilibrium ‘Suppose the choice variable is output level as in the Cournot case but the choices are made sequentially. Firm 1 (the ‘leader’) chooses its output level before firmBirkbeck Economics 42 2 (the ‘follower’ does, and the follower can observe the leader's choice before she makes her output choice. It is not hard to see that the follower’s reaction function is as in the previous example but the leader now optimizes taking the follower’s reaction function (rather than output level) as given. In other words, the leader maximizes m(qu, 92) = Pa = (120 — 1 — (60 ~ 5q1))qn- ‘The Nash equilibrium of this game, also known as the Stackelberg equilibrium, is here given by gj = 60, gj = 40.