Professional Documents
Culture Documents
Lecture 5
Lecture 5
Consider the following Rock-Paper-Scissors game: Note that RPS is a zerosum game.
R P S
0,0
1,1
1,1
1,1
0,0
1,1
1,1
1,1
0,0
This game has no pure-strategy Nash equilibrium. Whatever pure strategy player 1 chooses, player 2 can beat him. A natural solution for player 1 might be to randomize amongst his strategies. Another example of a game without pure-strategy NE is matching pennies. As in RPS the opponent can exploit his knowledge of the other players action.
1,1
1,1
1,1
1,1
Fearing this what might the opponent do? One solution is to randomize and play a mixed strategy. Each player could ip a coin and play H with probability 1 and T with probability 1 . 2 2 Note that each player cannot be taken advantage of. Denition 1 Let G be a game with strategy spaces S1 ,S2 ,..,SI . A mixed strategy i for player i is a probability distribution on Si i.e. for Si nite a mixed strategy is a function i : Si + such that si Si i (si ) = 1. Several notations are commonly used for describing mixed strategies. 1. Function (measure): 1 (H) =
1 2
and 1 (T ) =
1 2
2. Vector: If the pure strategies are si1 ,..siNi write (i (si1 ) , .., i (siNi )) 1 e.g. 2 , 1 . 2 3.
1 H 2
+ 1T 2
Class Experiment 1 Three groups of two people. Play RPS with each other 30 times. Calculate frequency with which each strategy is being played. Players are indierent between strategies if opponent mixes equally between all three strategies. In games such as matching pennies, poker blung, football run/pass etc you want to make the opponent guess and you worry about being found out.
Write i (also (Si )) for the set of probability distributions on Si . Write for 1 .. I . A mixed strategy prole is an I-tuple (1 , .., I ) with i i . We write ui (i , i ) for player is expected payo when he uses mixed strategy i and all other players play as in i . ui (i , i ) =
si ,si ui (si ,si )i (si )i (si )
(1)
Remark 1 For the denition of a mixed strategy payo we have to assume that the utility function fullls the VNM axioms. Mixed strategies induce lotteries over the outcomes (strategy proles) and the expected utility of a lottery allows a consistent ranking only if the preference relation satises these axioms. Denition 2 A mixed strategy NE of G is a mixed prole such that
ui i , i ui i , i
The denition of MSNE makes it cumbersome to check that a mixed prole is a NE. The next result shows that it is sucient to check against pure strategy alternatives. Proposition 1 is a Nash equilibrium if and only if
ui i , i ui si , i
Because of symmetry is it sucient to check that player 1 would not deviate. If he plays his mixed strategy he gets expected payo 0. Playing his two pure strategies gives him payo 0 as well. Therefore, there is no incentive to deviate. Note: Mixed strategies can help us to nd MSNE when no pure strategy NE exist. 3
Denition 3 In a nite game, the support of a mixed strategy i , supp (i ), is the set of pure strategies to which i assigns positive probability supp (i ) = {si Si |i (si ) > 0}
Proposition 2 If is a mixed strategy Nash equilibrium and si ,si supp (i ) then ui si , i = ui si , i
Dene a new mixed strategy i for player i i (si ) + i (si ) 0 i (si ) = i (si )
by if si = si if si = si otherwise
=
si Si
ui si , i i (si ) si Si ui si , i si Si
ui si , i i (si )
[i (si ) i (si )]
Note that a mixed strategy NE is never strict. The proposition suggests a process of nding MSNE. 1. Look at all possible supports for mixed equilibria. 2. Solve for probabilities and check if it works. 4
1,1
0,4
0,2
2,1
It is easy to see that this game has no pure strategy Nash equilibria. For a mixed strategy Nash equilibrium to exist player 1 has to be indierent between strategies U and D and player 2 has to be indierent between L and R. Assume player 1 plays U with probability and player 2 plays L with probability .
u1 (U, 2 ) u2 (L, 1 ) + 2 (1 )
= = = =
u1 (D, 2 ) 2 (1 ) u2 (R, 1 ) 4 + (1 )
We can deduce that = 1 and = 2 . There is a unique mixed Nash 4 3 equilibrium with 1 = 1 U + 3 D and 2 = 2 L + 1 R 4 4 3 3 Remark 2 Recall the Battle of the Sexes experiments from last class. It can be shown that the game has a mixed NE where each agent plays her favorite strategy with probability 2 . This was not quite the proportion of 3 people playing it in class (but pretty close to the proportion of people choosing it in the previous year)! In many instances of this experiment one nds that men and women diered in their aggressiveness. Does that imply that they were irrational? No. In a mixed NE players are indierent between their strategies. As long as men and women are matched completely randomly (i.e. woman-woman and man-man pairs are also possible) it only matters how players mix in the aggregate! It does NOT matter if subgroups (i.e. men and women) mix at dierent states, although it would matter if they 5
would play only against players within their subgroup. Interestingly, that suggests that letting women segregate into their own communities should make them more aggressive, and men less aggressive. The term aggressive is a bit misleading because it does not result in bigger payos. However, you could come up with a story of lexicographic preferences - people care rst of all about payos, but everything else equal they want to t gender stereotypes - so playing football is good for mens ego.
Finding mixed NE in 2 by 2 games is relatively easy. It becomes harder if players have more than two strategies because we have to start worrying about supports. In many cases it is useful to exploit iterated deletion in order to narrow down possible supports. Proposition 3 Let be a NE of G and suppose that (si ) > 0 then si Si , i.e. strategy si is not removed by ISD. Proof: see problem set 2 Having introduced mixed strategies we can even dene a tighter notion of IDSDS. Consider the next game. No player has a strategy which is strictly dominated by another pure strategy. However, C for player 2 is strictly dominated by 1 L + 1 R. Thus we would think that C wont be used. 2 2
1,1
0,2
0,4
0,2
5,0
1,6
0,2
1,1
2,1
After we delete C we note that M is dominated by 2 D + 1 U . Using the above 3 3 proposition we can conclude that the only Nash equilibria are the NE of the 2 by 2 game analyzed in the previous section. Since that game had a unique mixed strategy equilibrium we can conclude that the only NE of the 3 by 3 1 game is 1 = 4 U + 3 D and 2 = 2 L + 1 R. 4 3 3 It is useful to adjust the formal denition of IDSDS and allow for mixed strategy domination: Denition 4 The set of strategy proles surviving iterated strict dominance is S = S1 S2 .. SI where
Si Si0 Sik+1
=
k=1
Sik
Remark 3 Recall the above 3 by 3 game. If we would look for possible mixed NE with supports (U,M) and (L,C) respectively, we would nd a potential NE 2 U + 1 M, 5 L + 1 C. However, this is NOT a NE because player 2 would play 3 3 6 6 R instead. Remark 4 In the RPS game we cannot reduce the set of strategies through IDSDS. Therefore we have to check all possible supports and check if it works.
Denition 5 A correspondence F : A B is a mapping which associates to every element of a A a subset F (a) B. The mixed strategy best response correspondence for player i BRi : i i is dened by BRi (i ) = arg max ui (i , i )
i i Proposition 4 is a Nash equilibrium if and only if i BRi i for all i.
2 3 2 3 2 3 1 4 1 4 1 4
Interpretation of Mixed NE
1. Sometimes players explicitly ip coins. That ts games like Poker, soccer etc., where players have to randomize credibly. 2. Large populations of players with each player playing a xed strategy and random matching. Thats very related to the social norm explanation of pure Nash equilibrium. 3. Payo uncertainty (Harsanyi, purication). Roughly their argument goes as follows in the matching penny game. There are two types of players - those who get slightly higher payo from playing heads, and those who get higher payo from getting tails (their preferences are almost the same - think of one guy getting 1 dollar and the other guy getting 1 + = 1.01 dollars from playing head). Also, we assume that there is an equal probability that my opponent is of type 1 or type 2. In this circumstances no player loses from just playing her favorite strategy (i.e. a pure strategy) because it will do best on average. To show that this is a NE we have to introduce the notion of incomplete information which we dont do just yet. Harsanyi-Selten then let the payo uncertainty go to 0 such that the game in the limit approaches the standard matching pennies. Players average strategies converge to the mixed equilibrium.