The document summarizes a lesson on game theory and linear programming. It discusses using linear programming to find optimal strategies in zero-sum games represented by payoff matrices. It provides examples of solving for optimal strategies in Rock-Paper-Scissors and another sample game. The key steps of formulating the column player's problem as a linear program to minimize the maximum payoff for the row player are outlined.
Navigating the Deluge_ Dubai Floods and the Resilience of Dubai International...
Lesson 35: Game Theory and Linear Programming
1. Lesson 35
Game Theory and Linear Programming
Math 20
December 14, 2007
Announcements
Pset 12 due December 17 (last day of class)
Lecture notes and K&H on website
next OH Monday 1–2 (SC 323)
2. Outline
Recap
Definitions
Examples
Fundamental Theorem
Games we can solve so far
GT problems as LP problems
From the continuous to the discrete
Standardization
Rock/Paper/Scissors again
The row player’s LP problem
3. Definition
A zero-sum game is defined by a payoff matrix A, where aij
represents the payoff to the row player if R chooses option i and C
chooses option j.
4. Definition
A zero-sum game is defined by a payoff matrix A, where aij
represents the payoff to the row player if R chooses option i and C
chooses option j.
The row player chooses from the rows of the matrix, and the
column player from the columns.
The payoff could be a negative number, representing a net
gain for the column player.
5. Definition
A strategy for a player consists of a probability vector representing
the portion of time each option is employed.
6. Definition
A strategy for a player consists of a probability vector representing
the portion of time each option is employed.
We use a row vector p for the row player’s strategy, and a
column vector q for the column player’s strategy.
A pure strategy (select the same option every time) is
represented by a standard basis vector ej or ej . For instance,
if R has three choices and C has five:
0
e2 = 1 e4 = 0 0 0 1 0
0
A non-pure strategy is called mixed.
7. Definition
The expected value of row and column strategies p and q is the
scalar
n
E (p, q) = pi aij qj = pAq
i,j=1
8. Definition
The expected value of row and column strategies p and q is the
scalar
n
E (p, q) = pi aij qj = pAq
i,j=1
Probabilistically, this is the amount the row player receives (or the
column player if it’s negative) if players employ these strategies.
9. Rock/Paper/Scissors
Example
What is the payoff matrix for Rock/Paper/Scissors?
10. Rock/Paper/Scissors
Example
What is the payoff matrix for Rock/Paper/Scissors?
Solution
The payoff matrix is
0 −1 1
0 −1 .
A= 1
−1 1 0
11. Example
Consider a new game: players R and C each choose a number 1,
2, or 3. If they choose the same thing, C pays R that amount. If
they choose differently, R pays C the amount that C has chosen.
What is the payoff matrix?
12. Example
Consider a new game: players R and C each choose a number 1,
2, or 3. If they choose the same thing, C pays R that amount. If
they choose differently, R pays C the amount that C has chosen.
What is the payoff matrix?
Solution
1 −2 −3
A = −1 2 −3
−1 −2 3
13. Theorem (Fundamental Theorem of Matrix Games)
There exist optimal strategies p∗ for R and q∗ for C such that for
all strategies p and q:
E (p∗ , q) ≥ E (p∗ , q∗ ) ≥ E (p, q∗ )
14. Theorem (Fundamental Theorem of Matrix Games)
There exist optimal strategies p∗ for R and q∗ for C such that for
all strategies p and q:
E (p∗ , q) ≥ E (p∗ , q∗ ) ≥ E (p, q∗ )
E (p∗ , q∗ ) is called the value v of the game.
15. Reflect on the inequality
E (p∗ , q) ≥ E (p∗ , q∗ ) ≥ E (p, q∗ )
In other words,
E (p∗ , q) ≥ E (p∗ , q∗ ): R can guarantee a lower bound on
his/her payoff
E (p∗ , q∗ ) ≥ E (p, q∗ ): C can guarantee an upper bound on
how much he/she loses
This value could be negative in which case C has the
advantage
16. Fundamental problem of zero-sum games
Find the p∗ and q∗ !
Last time we did these:
Strictly-determined games
2 × 2 non-strictly-determined games
The general case we’ll look at next.
17. Pure Strategies are optimal in Strictly-Determined Games
Theorem
Let A be a payoff matrix. If ars is a saddle point, then er is an
optimal strategy for R and es is an optimal strategy for C. Also
v = E (er , es ) = ars .
18. Optimal strategies in 2 × 2 non-Strictly-Determined Games
Let A be a 2 × 2 matrix with no saddle points. Then the optimal
strategies are
a − a
22 12
a22 − a21 a11 − a12 ∆
p= q = a11 − a21
∆ ∆
∆
where ∆ = a11 + a22 − a12 − a21 . Also
|A|
v=
∆
19. Outline
Recap
Definitions
Examples
Fundamental Theorem
Games we can solve so far
GT problems as LP problems
From the continuous to the discrete
Standardization
Rock/Paper/Scissors again
The row player’s LP problem
20. This could get a little weird
This derivation is not something that needs to be memorized, but
should be understood at least once.
21. Objectifying the problem
Let’s think about the problem from the column player’s
perspective. If she chooses strategy q, and R knew it, he would
choose p to maximize the payoff pAq. Thus the column player
wants to minimize that quantity. That is, C ’s objective is realized
when the payoff is
E = min max pAq.
q p
22. Objectifying the problem
Let’s think about the problem from the column player’s
perspective. If she chooses strategy q, and R knew it, he would
choose p to maximize the payoff pAq. Thus the column player
wants to minimize that quantity. That is, C ’s objective is realized
when the payoff is
E = min max pAq.
q p
This seems hard! Luckily, linearity, saves us.
23. From the continuous to the discrete
Lemma
Regardless of q, we have
max pAq = max ei Aq
p 1≤i≤m
Here ei is the probability vector represents the pure strategy of
going only with choice i.
24. From the continuous to the discrete
Lemma
Regardless of q, we have
max pAq = max ei Aq
p 1≤i≤m
Here ei is the probability vector represents the pure strategy of
going only with choice i.
The idea is that a weighted average of things is no bigger than the
largest of them. (Think about grades).
25. Proof of the lemma
Proof.
We must have
max pAq ≥ max ei Aq
p 1≤i≤m
(the maximum over a larger set must be at least as big). On the
other hand, let q be C ’s strategy. Let the quantity on the right be
maximized when i = i0 . Let p be any strategy for R. Notice that
p = i pi ei . So
m m
pi ei Aq ≤
E (p, q) = pAq = pi ei0 Aq
i=1 i=1
m
= pi ei0 Aq = ei0 Aq.
i=1
Thus
max pAq ≤ ei0 Aq.
p
26. The next step is to introduce a new variable v representing the
value of this inner maximization. Our objective is to minimize it.
Saying it’s the maximum of all payoffs from pure strategies is the
same as saying
v ≥ ei Aq
for all i. So we finally have something that looks like an LP
problem! We want to choose q and v which minimize v subject to
the constraints
v ≥ ei Aq i = 1, 2, . . . m
qj ≥ 0 j = 1, 2, . . . n
n
qj = 1
j=1
27. Trouble with this formulation
Simplex method with equalities?
Not in standard form
Resolution:
We may assume all aij ≥ 0, so v > 0
qj
Let xj = v
28. Since we know v > 0, we still have x ≥ 0. Now
n n
1 1
xj = qj = .
v v
j=1 j=1
So our problem is now to choose x ≥ 0 which maximizes xj .
j
The constraints now take the form
v ≥ ei Aq ⇐⇒ 1 ≥ ei Ax,
for all i. Another way to write this is
Ax ≤ 1,
where 1 is the vector consisting of all ones.
29. Upshot
Theorem
Consider a game with payoff matrix A, where each entry of A is
x
positive. The column player’s optimal strategy q is ,
x1 + · · · + xn
where x ≥ 0 satisfies the LP problem of maximizing x1 + · · · + xn
subject to the constraints Ax ≤ 1.
31. Rock/Paper Scissors
The payoff matrix is
0 −1 1
0 −1 .
A= 1
−1 1 0
We can add 2 to everything to make
213
˜
A = 3 2 1 .
132
32. Convert to LP
The problem is to maximize x1 + x2 + x3 subject to the constraints
2x1 + x2 + 3x3 ≤ 1
3x1 + 2x2 + x3 ≤ 1
x1 + 3x3 + 2x3 ≤ 1.
We introduce slack variables y1 , y2 , and y3 , so the constraints now
become
2x1 + x2 + 3x3 + y1 = 1
3x1 + 2x2 + x3 + y2 = 1
x1 + 3x3 + 2x3 + y3 = 1.
33. An easy initial basic solution is to let x = 0 and y = 1. The initial
tableau is therefore
x1 x2 x3 y1 y2 y3 z value
y1 2 1 3100 0 1
y2 3 2 1010 0 1
y3 1 3 2001 0 1
z −1 −1 −1 0 0 0 1 0
34. Which should be the entering variable? The coefficients in the
bottom row are all the same, so let’s just pick one, x1 . To find the
departing variable, we look at the ratios 1 , 3 , and 1 . So y2 is the
1
2 1
departing variable.
We scale row 2 by 1 :
3
x1 x2 x3 y1 y2 y3 z value
y1 2 1 31 00 0 1
y2 1 2/3 1/3 0 1/3 0 0 1/3
y3 1 3 20 01 0 1
−1 −1 −1 0
z 00 1 0
35. Then we use row operations to zero out the rest of column one:
x1 x2 x3 y1 y2 y3 z value
0 −1/3 1 −2/3 0
y1 7/3 0 1/3
x1 1 2/3 1/3 0 1/3 0 0 1/3
0 −1/3 1
y3 0 7/3 5/3 0 2/3
0 −1/3 −2/3 0
z 1/3 0 1 1/3
36. We can still improve this: x3 is the entering variable and y1 is the
departing variable. The new tableau is
x1 x2 x3 y1 y2 y3 z value
0 −1/7 1 3/7 −2/7
x3 0 0 1/7
0−
x1 1 5/7 1/7 3/7 0 0 2/7
0 18/7 0 −5/7
y3 1/7 1 0 3/7
0−
z 3/7 0 2/7 1/7 0 1 3/7
38. So the x variables have values x1 = 1/6, x2 = 1/6, x3 = 1/6.
Furthermore z = x1 + x2 + x3 = 1/2, so v = 1/z = 2. This also
means that p1 = 1/3, p2 = 1/3, and p3 = 1/3. So the optimal
strategy is to do each thing the same number of times.
39. Outline
Recap
Definitions
Examples
Fundamental Theorem
Games we can solve so far
GT problems as LP problems
From the continuous to the discrete
Standardization
Rock/Paper/Scissors again
The row player’s LP problem
40. Now let’s think about the problem from the column player’s
perspective. If he chooses strategy p, and C knew it, he would
choose p to minimize the payoff pAq. Thus the row player wants
to maximize that quantity. That is, R’s objective is realized when
the payoff is
E = max min pAq.
p q
42. The next step is to introduce a new variable v representing the
˜
value of this inner minimization. Our objective is to maximize it.
Saying it’s the minimum of all payoffs from pure strategies is the
same as saying
v ≤ pAej
˜
for all j. Again, we have something that looks like an LP problem!
We want to choose p and v which maximize v subject to the
˜ ˜
constraints
v ≤ pAej
˜ j = 1, 2, . . . n
pi ≥ 0 i = 1, 2, . . . m
m
pi = 1
i=1
43. As before, we can standardize this by renaming
1
y= p
v
˜
(this makes y a column vector). Then
m
1
yi = ,
v
˜
i=1
So maximizing v is the same as minimizing 1 y. Likewise, the
˜
equations of constraint become v ≤ (˜ y )Aej for all j, or y A ≥ 1 ,
˜ v
or (taking transposes) A y ≥ 1. If all the entries of A are positive,
we may assume that v is positive, so the constraints p ≥ 0 are
˜
satisfied if and only if y ≥ 0.
44. Upshot
Theorem
Consider a game with payoff matrix A, where each entry of A is
y
positive. The row player’s optimal strategy p is ,
y1 + · · · + yn
where y ≥ 0 satisfies the LP problem of minimizing
y1 + · · · + yn = 1 y subject to the constraints A y ≥ 1.
45. The big idea
The big observation is this:
Theorem
The row player’s LP problem is the dual of the column player’s LP
problem.
46. The final tableau in the Rock/Paper/Scissors LP problem was this:
x1 x2 x3 y1 y2 y3 z value
7/18 −5/18
x3 001 1/18 0 1/6
7/18 −5/18 0
x1 100 1/18 1/6
0 1 0 −5/18
x2 1/18 7/18 0 1/6
z 000 1/6 1/6 1/6 1 1/2
The entries in the objective row below the slack variables are the
solutions to the dual problem! In this case, we have the same
values, which means R has the same strategy as C . This reflects
the symmetry of the original game.
47. Example
Consider the game: players R and C each choose a number 1, 2,
or 3. If they choose the same thing, C pays R that amount. If
they choose differently, R pays C the amount that C has chosen.
What should each do?
48. Example
Consider the game: players R and C each choose a number 1, 2,
or 3. If they choose the same thing, C pays R that amount. If
they choose differently, R pays C the amount that C has chosen.
What should each do?
Answer.
Choice R C
1 54.5% 22.7%
2 27.3% 36.3%
3 18.2% 40.1%
The expected payoff is 2.71 to the column player.