32

I am trying to understand how to compute all Nash equilibria in a 2 player game, but I fail when there are more than 2 possible options to play. Could somebody explain to me how to calculate a matrix like this (without computer) \begin{matrix} 1,1 & 10,0 & -10,1 \\ 0,10 & 1,1 & 10,1 \\ 1,-10 & 1,10 & 1,1 \end{matrix}

I tried it with the support concept, but I don't get it...

Nameless
  • 4,155
  • 2
  • 22
  • 37
user106371
  • 371

3 Answers3

42

A Nash equilibrium is a profile of strategies $(s_1,s_2)$ such that the strategies are best responses to each other, i.e., no player can do strictly better by deviating. This helps us to find the (pure strategy) Nash equilibria.

To start, we find the best response for player 1 for each of the strategies player 2 can play. I will demonstrate this by underlining the best responses: \begin{matrix} & A &B&C \\ A&\underline{1},1 & \underline{10},0 & -10,1 \\ B&0,10 & 1,1 & \underline{10},1 \\ C&\underline{1},-10 & 1,10 & 1,1 \end{matrix} Player 1 is the row player, player 2 is the column player. If 2 plays column A, then player 1's best response is to play either row $A$ or $C$, which gives him 1 rather than 0 as payoff. Similarly, the best response to column $B$ is row $A$, and to column $C$ it is row $B$.

Now we do the same for player 2 by underlining the best responses of the column player: \begin{matrix} & A &B&C \\ A&\underline{1},\underline{1} & \underline{10},0 & -10,\underline{1} \\ B&0,\underline{10} & 1,1 & \underline{10},1 \\ C&\underline{1},-10 & 1,\underline{10} & 1,1 \end{matrix} So, if player 1 plays row $A$ then player 2 best responds either with column $A$ or column $C$, giving him 1 rather than 0. We also find the best responses for row $B$ and $C$.

Now a pure strategy Nash equilibrium is a cell where both payoffs are underlined, i.e., where both strategies are best responses to each other. In the example, the unique pure strategy equilibrium is $(A,A)$. (There may also be mixed strategy equilibria.) In all other cells, at least one player has an incentive to deviate (because it gives him a higher payoff).

EDIT: How to compute mixed strategy equilibria in discrete games?

In a mixed Nash strategy equilibrium, each of the players must be indifferent between any of the pure strategies played with positive probability. If this were not the case, then there is a profitable deviation (play the pure strategy with higher payoff with higher probability).

Consider player 2. He plays column $A$ with probability $p$, $B$ with probability $q$, and $C$ with probability $1-p-q$. We need to find $p,q$ such that player 1 is indifferent between his pure strategies $A,B,C$. He is indifferent between row $A$ (left hand side) and row $B$ (right hand side) if $p,q$ are such that $$p+10q-10(1-q-p)=q+10(1-p-q).$$ He is indifferent between $B$ and $C$ if $$q+10(1-p-q)=p+q+1-q-p=1.$$ You just have to solve the first condition for $q$ as function of $p$, substitute $q$ in the second condition and you have $p$. Inserting $p$ again in the first gives you $q$.

Now we do the same with strategies for player 1 such that player 2 is indifferent. Player 1 plays $A$ with probability $x$, $B$ with probability $y$ and $C$ with probability $1-x-y$. The two conditions that follow are \begin{equation} 1x+10y-10(1-x-y)=x+10(1-x-y) \\ x+10(1-x-y)=1 \end{equation} Solve this again to find $x,y$. This is a mixed-strategy equilibrium, because neither player has a profitable deviation. Remember, we constructed the profile $(x,y;p,q)$ such that the other player is indifferent between his pure strategies. So, no matter how the other player unilaterally deviates, his expected payoff will be identical to that in equilibrium $(x,y;p,q)$. In general, depending on the game and solutions $x,y,p,q$, there may be infinitely many mixed Nash equilibria, or none. The more pure strategies there are, the more tedious it is to compute mixed strategy equilibria, since we solve for $N-1$ variables for each player ($N$ being the number of pure strategies of the other player).

Moreover, to find all equilibria, if there are more than 2 actions for a player, then every possible combination of actions has to be checked. Here, a player has 3 actions, and a mixed strategy equilibrium could entail mixing over all three or just any two of them. Since such a player would not have to be indifferent regarding the strategy played with probability 0, the equations you have to set up are different. In summary, manually checking for all possible mixed strategy equilibria if at least one player has more than two actions can require a lot of effort.

Nameless
  • 4,155
  • 2
  • 22
  • 37
  • Thank you very much for your answer, unfortunately my problem is not to find the pure strategy equilibriums, but the mixed. I am sorry that i was not clear enough in my question, I was looking for a method to compute all nash equilibriums, and i want to be sure after that i have all. Do you know any algorithm(that is calculable by hand) for it? – user106371 Nov 12 '13 at 19:06
  • 3
    I included an edit that described how to find the mixed strategy equilibria. – Nameless Nov 12 '13 at 19:34
  • Good answer! But why $(B, B)$ and $(C, C)$ are not a pure strategies Nash equilibrium ? – npisinp Apr 27 '14 at 18:16
  • Because if player 2 plays $B$, then player 1 wants to play $A$ (giving him 10) rather than $B$ (giving him 1). Similarly, if 1 plays $B$, player 2 wants to play $A$ rather than $B$. By looking at the underlined best responses, you find similar profitable deviations for $(C,C)$. – Nameless Apr 29 '14 at 19:18
  • Dear @Nameless, you wrote: ''We need to find $p,q$ such that player $1$ is indifferent between his pure strategies $A,B,C$.'' In this case, are you supposing that $A$, $B$ and $C$ are in the support of $s_1$? – rgm May 17 '17 at 21:41
  • Doesn't matter. If player $1$ is indifferent between $A,B,C$, then he can optimally play any of these strategies with any probability (even 0, because he is indifferent). The only criterion for a mixed strategy equilibrium is then for $1$ to pick probabilities of playing $A,B,C$ such that player $2$ is indifferent between his pure strategies as well (which allows him to choose $p,q$ to make player $1$ indifferent in the first place). – Nameless May 18 '17 at 01:10
  • "In general, depending on the solutions x,y,p,q, there may be infinitely many Nash equilibria, or none." Are you sure about the "none" part? Wikipedia says "Nash proved that if mixed strategies (where a player chooses probabilities of using various pure strategies) are allowed, then every game with a finite number of players in which each player can choose from finitely many pure strategies has at least one Nash equilibrium." – Bennett Jan 28 '22 at 04:20
  • Isn't this answer incomplete if you're trying to find all Nash equilibria, because you also have to check for cases where the player plays one choice with probability zero? Then the utility of that choice no longer has to be equal to the other choices; it can be less. This allows for equilibria other than the one(s) where all the probabilities are positive. (I'm not saying any such equilibria exist for the given problem, I'm just saying you have to check.) – Bennett Jan 28 '22 at 07:04
  • But Nash's result does not say a mixed strategy equilibrium always exists. You could have a pure strategy equilibrium but no mixed strategy equilibrium, as in the prisoner's dilemma. Regarding your second point, you are right, to find all equilibria we would also have to check for strategies that mix over 2 rather than 3 strategies - I amended the answer, it gets very tedious if the strategy space has more than 2 strategies. – Nameless Jan 29 '22 at 10:30
  • Notice that since the game is symmetric, the two systems of equations are identical, leading to symmetric Nash equilibria. – Dave the Sid Dec 03 '24 at 20:47
5

I think I finally understood how to get all equilibrias by the support concept.

Given payoff matrix A for Player 1 and B for Player 2. That a point $(x,y)=(x_1,x_2,x_3,y_1,y_2,y_3)$ is a Nash equilibria with supp $x =I$ and supp $ y =J$ (which means that $x_i>0 $ for $i \in I$ and $ x_i=0 $ for $ i \not\in I$) is equivalent to

$ 1) \forall i,k \in I : (Ay)_i = (Ay)_k \\ 2) \forall i \in I,\forall k \not\in I : (Ay)_k ≤ (Ay)_i \\ 3) \forall j,l \in J : (Bx)_j = (Bx)_l \\ 4) \forall j \in J,\forall l \not\in J : (Bx)_l ≤ (Bx)_j $

So to get all Nash equilibrias, I have to check for every support combination that is possible, if all equalities and inequalities hold on. If yes the equations tell for what points or intervalls there are Nash equilibrias. In the example above the equations of Nameless are the equations for the case $I=\{1,2,3\}=J$. All in all there are to check 49 equation systems, which is a lot of work...

Thank you for your answer Nameless, it helped me a lot to understand the main principle.

user106371
  • 371
0

@user106371. Your answer is exactly true. Assigning probabilities to each strategy and solving the system of equations is not a complete solution, only one case. In order to calculate number of equation systems need to be solved, you can use the following formula:

$$\sum_{k=1}^m \sum_{p=1}^n \binom{m}{k} \binom{n}{p}$$

where $m$ and $n$ are the number of strategies of the players.

alamaranka
  • 117