Game Theory: Domination

Assoc. Prof. Ayman Ghoneim

Operations Research and Decision Support Department

Faculty of Computers and Artificial Intelligence, Cairo University
Contact Email: [email protected]

Decision and Game Theory (DS414/DS411) - 2024

Ayman Ghoneim (FCAI - Cairo University) Game Theory: Domination Spring 2024 1 / 30
Recall - Prisoners Dilemma
Prisoners Dilemma Game:

Figure: Prisoners Dilemma.

Choose D whatever your opponent will do.

Ayman Ghoneim (FCAI - Cairo University) Game Theory: Domination Spring 2024 2 / 30
Categorizing Strategy Games
Strategy games can be classified into different categories:
1 a two-agent (2-Players) game or a multi-agent (n-Player) game;
2 a cooperative game where communication between agents is permitted
and agents can enforce cooperation (e.g., by using contracts) or a
non-cooperative game where there is no communication allowed;
3 a symmetric game where agents can swap identities without changing the
utility of their strategies or an asymmetric game otherwise;
4 a zero (or constant) sum game where what an agent wins is equal to what
its opponents lose (i.e., the agents’ total utility sums to zero or a
constant) or a non-zero sum game otherwise; and/or
5 a simultaneous game where agents take their actions simultaneously, or a
sequential/dynamic game where agents take their actions in a
chronological order.
Note: A strategy game can belong to several categories in the same time, e.g.,
the prisoner’s dilemma game is a 2-player, non-cooperative, symmetric,
non-zero sum and simultaneous game.
Ayman Ghoneim (FCAI - Cairo University) Game Theory: Domination Spring 2024 3 / 30
Strategy

Let:
Ai be the set of all possible actions available to agent i; and
Σ be the set of all possible states of the world that may occur in a game.

Agent i will adopt a strategy si before playing the game, defined as follows.
Definition (Strategy)
A strategy si : Σ → Ai of agent i determines the action to be taken by the
agent in each state of game.

A strategy is a complete contingent plan for a player in the game.

In simple games where there is only one state, the agent’s strategy becomes
simply an action.

Ayman Ghoneim (FCAI - Cairo University) Game Theory: Domination Spring 2024 4 / 30
Strategy - Iterated PD

Iterated prisoner’s dilemma: a series of repeated rounds of the game is

played.
Future Shadow, influenced of game history.
Requires a more complicated strategy.

Ayman Ghoneim (FCAI - Cairo University) Game Theory: Domination Spring 2024 5 / 30
Strategy - Iterated PD
A strategy contains a history portion which is used to keep the players’
actions for the previous L history steps, and a lookup table of size 2nL .

Figure: A lookup table representation of a strategy with one history step.

The strategy indicates an action for each possible history (i.e., state) in the
game.
Ayman Ghoneim (FCAI - Cairo University) Game Theory: Domination Spring 2024 6 / 30
A Game Setting

Let:
I = {1, 2, ..., n} be the set of players, where n is a positive integer.
Si = {1, 2, ..., mi } be finite set of strategies for player i, where mi ≥ 1.
Also known as strategy space of agent i.
si ∈ Si be a strategy of player i.
S be the joint strategy space of the game, where S = ×i Si is the cartesian
product of the players’ pure strategy sets.
s = (sl , s2 , ..., sn ) be a strategy profile which holds a strategy si for each
player i.
s−i = (sl , s2 , ..., si−1 , si+1 , ..., sn ) be a strategy profile which holds a
strategy si for all players except for player i.
Given the definition of s−i , we can write s = (si , s−i ).

Ayman Ghoneim (FCAI - Cairo University) Game Theory: Domination Spring 2024 7 / 30
A Game Setting

Let:
πi (s) be the associated payoff to player i ∈ I given any strategy profile
s ∈ S, i.e., πi : S → R is a pure-strategy payoff function.
π(s) = (π1 (s), ..., πn (s)) be the full vector of payoffs given a
pure-strategy profile s, i.e., π : S → Rn is a joint pure-strategy payoff
function of the game.
G = (I, S, π) be a triplet describing a pure strategy game, where I is its
player set, S its pure-strategy joint space, and π its joint payoff function.

“We will assume for now that all the above are known".

Ayman Ghoneim (FCAI - Cairo University) Game Theory: Domination Spring 2024 8 / 30
Example 1 - Prisoners Dilemma

Figure: Prisoners Dilemma.

Define I, Si for each player i, S, and π for each strategy profile.

Ayman Ghoneim (FCAI - Cairo University) Game Theory: Domination Spring 2024 9 / 30
Example 1 - Prisoner’s Dilemma

Figure: Prisoners Dilemma.

I = {1, 2} , where n = 2.
S1 = S2 = {1, 2} or {Cooperate, Defect} or {C, D}.
S = {(C, C), (C, D), (D, C), (D, D)}.
π((C, C)) = (3, 3), π((C, D)) = (0, 5), π((D, C)) = (5, 0) and
π((D, D)) = (1, 1).

Ayman Ghoneim (FCAI - Cairo University) Game Theory: Domination Spring 2024 10 / 30
Example 2

Player 2
L C R
U 5,-1 11,3 0,0
Player 1 D 6,4 0,2 2,0

Define I, Si for each player i, S, π1 (U, C) and π2 (D, L).

I = {1, 2} , where n = 2.
S1 = {U, D} and S2 = {L, C, R}.
S = {(U, L), (U, C), (U, R), (D, L), (D, C), (D, R)}.
π1 (U, C) = 11 and π2 (D, L) = 4.

Ayman Ghoneim (FCAI - Cairo University) Game Theory: Domination Spring 2024 11 / 30
Strict Domination

Definition (Strictly Dominant - Informal Definition)

We say that my strategy D strictly dominates my strategy C if my payoff from
D is strictly greater than C regardless of what others do.

Never play a strictly dominated strategy, because if instead i played the

strategy that dominates it, i will do better in every case.

Ayman Ghoneim (FCAI - Cairo University) Game Theory: Domination Spring 2024 12 / 30
Strict Domination

Ok, but what about the following argument:

If we both choose C, then we will get 3, which is better than we both choose
D and get 1.
What are the problem(s) with the previous argument ?
I can’t affect other people choices in a way or another.
Even if i forced people to choose C, then i will choose D.
Rationality lead for an inefficient outcomes.

Ayman Ghoneim (FCAI - Cairo University) Game Theory: Domination Spring 2024 13 / 30
Strictly and Weakly Dominated Strategies

Definition (Strictly Dominated Strategy)

Player i strategy s0i is strictly dominated by player i strategy si if
ui (si , s−i ) > ui (s0i , s−i ) for all possible s−i .

Definition (Weakly Dominated Strategy)

Player i strategy s0i is weakly dominated by player i strategy si if
ui (si , s−i ) ≥ ui (s0i , s−i ) for all possible s−i , and ui (si , s−i ) > ui (s0i , s−i ) for at
least one s−i .

Ayman Ghoneim (FCAI - Cairo University) Game Theory: Domination Spring 2024 14 / 30
Example - Dominated Strategy

Let’s consider the following payoff matrix:

Player 2
L C R
U 7,4 11,3 2,0
Player 1 D 6,1 0,2 2,0

For player 1: Strategy U weakly dominates strategy D, since 7 > 6, 11 > 0

and 2 = 2.
For Player 2: Strategy L strictly dominates strategy R, since 4 > 0 and 1 > 0.
For Player 2: Strategy C strictly dominates strategy R, since 3 > 0 and 2 > 0.

Ayman Ghoneim (FCAI - Cairo University) Game Theory: Domination Spring 2024 15 / 30
Domination - “Put yourself in others people shoes"
Let’s consider the following payoff matrix:

Do i have any dominated strategies ?

But my pair has α dominates β, so her choice will be α.
Do i have any dominated strategies ?
Iterative elimination of dominated strategies: Put yourself in others shoes and
try to figure out what they will do.
Note: In reality, you may not know my opponent(s)’ payoff.
Ayman Ghoneim (FCAI - Cairo University) Game Theory: Domination Spring 2024 16 / 30
Example - Iterative Elimination of strictly dominated
strategies
Let’s consider the following payoff matrix:
Player 2
L C R
U 7,4 11,3 2,0
Player 1 D 6,1 0,2 2,0
Solve the game by eliminating strictly dominated strategies.
Start by Player 1: It has no any strictly dominated strategies to eliminate (U
“weakly" dominates D).
Go to Player 2: Strategy C strictly dominates strategy R. Thus, removing
strategy R result in the following payoff matrix
Player 2
L C
U 7,4 11,3
Player 1 D 6,1 0,2
Ayman Ghoneim (FCAI - Cairo University) Game Theory: Domination Spring 2024 17 / 30
Example - Iterative Elimination of strictly dominated
strategies

Back to player 1: Now strategy U strictly dominates strategy D. Thus,

removing strategy D resulting the following payoff matrix

Player 2
L C
Player 1 U 7,4 11,3

Back to Player 2: Strategy L strictly dominates strategy C. Thus we eliminate

strategy C and the game solution is (U, L).
When iterative elimination results in a single strategy profile for the game,
this game is called a “Dominance solvable game".

Ayman Ghoneim (FCAI - Cairo University) Game Theory: Domination Spring 2024 18 / 30
Example - Iterative Elimination of weakly dominated
strategies
Let’s consider the following payoff matrix:
Player 2
L C R
U 7,4 11,3 2,0
Player 1 D 6,1 0,2 2,0

For player 1: Strategy U weakly dominates strategy D, thus we eliminate

strategy D resulting in the following payoff matrix.
Player 2
L C R
Player 1 U 7,4 11,3 2,0

For Player 2: Strategy L strictly dominates strategies C and R, thus we

eliminate strategies C and R, and the game is dominance solvable with a
strategy profile (U, L).
Ayman Ghoneim (FCAI - Cairo University) Game Theory: Domination Spring 2024 19 / 30
Iterative Elimination: Pick a Number

Pick a Number Game:

Without showing your neighbor classmate what you are doing. Pick a whole
number between 1 and 100. the winner is the one who chooses the closest
number to 2/3 of the average of the number chosen by the students in the class.
The winner will get 5 L.E. minus the number in piasters between his/her
choice and 2/3 of the average number.

Ayman Ghoneim (FCAI - Cairo University) Game Theory: Domination Spring 2024 20 / 30
Iterative Elimination: Pick a Number

Example:
Assume we have 3 students who chose 25, 5, and 60.
Total: 25 + 5 + 60 = 90
Average: 90/3 = 30
2/3 × 30 = 20.
The winner is the guy who choose 25 and he/she will get 5 L.E. - 5 piasters =
4.95 L.E.

Ayman Ghoneim (FCAI - Cairo University) Game Theory: Domination Spring 2024 21 / 30
Iterative Elimination: Pick a Number

Keep the following points in mind:

Don’t choose a strictly or a weakly dominated strategy.
Put yourself in other people shoes.

Let’s consider the following argument:

If people choose randomly between 1 and 100, then the average will be 50
and 2/3 × 50 = 33. Thus, i should report a number close to 33. What is
wrong with this reasoning ?
People will not choose random.

If all people reason in the same way, then they will choose a close
number to 33, which will make the average around 33 and the 2/3 × 33
is around 22.

Ayman Ghoneim (FCAI - Cairo University) Game Theory: Domination Spring 2024 22 / 30
Iterative Elimination: Pick a Number

Let’s think in a game theoretic way, are there any strategies which we can rule
out ?
Yes, If all people choose 100, then the average will be 100 and 2/3 × 100 is
66. Thus choosing a number greater than 67 is dominated.
By eliminating all the numbers above 67, we are left with choosing a number
between 1 and 67.
By using a similar argument, if all people choose 67, then the average will be
67 and 2/3 × 67 is 45. Thus choosing a number greater than 45 is dominated.
The numbers between 45 and 67 were not dominated in the original game, but
they become dominated once we eliminated the numbers above 67.
Using similar arguments, we will remove the numbers between 30 and 45,
then the numbers between 20 and 30, and so on.

Ayman Ghoneim (FCAI - Cairo University) Game Theory: Domination Spring 2024 23 / 30
Iterative Elimination: Pick a Number

What is the consequence of rationality when playing this game ?

Simply being rational rules any dominated strategies. To eliminate based on
the already eliminated dominated strategies, i need to be rational and assume
that others are rational, and other knows that i am rational.

We need a “common knowledge of rationality".

Ayman Ghoneim (FCAI - Cairo University) Game Theory: Domination Spring 2024 24 / 30
Iterative Elimination: Politics Application
Two candidate are choosing their political position in a political spectrum for
an election. This political spectrum has 10 positions, and voters are uniformly
distributed on the political spectrum where 10% of the voters exists at each
position.

Voters in a certain position will vote for the closest candidate to them. In case
of a tie, voters will split between the two closest candidates.
The payoff of each candidate is the total percentage of voters who voted for
him.
Ayman Ghoneim (FCAI - Cairo University) Game Theory: Domination Spring 2024 25 / 30
Iterative Elimination: Politics Application

Does position 2 dominate position 1 regardless of the position of the other

candidate ?

Candidate 2 Position Comparing Candidate 1 Utility at positions 1 and 2

Position 1 u1 (1, 1) = 50% < u1 (2, 1) = 90%
Position 2 u1 (1, 2) = 10% < u1 (2, 2) = 50%
Position 3 u1 (1, 3) = 15% < u1 (2, 3) = 20%
Position 4 u1 (1, 4) = 20% < u1 (2, 4) = 25%

After the first two rows, we observe a pattern where choosing position 2 give
me a 5% better than choosing position 1.
Conclusion: Position 2 strictly dominates position 1.
By similarity, position 9 strictly dominates position 10.

Ayman Ghoneim (FCAI - Cairo University) Game Theory: Domination Spring 2024 26 / 30
Iterative Elimination: Politics Application
Any other position is dominated ? Does position 3 dominate position 2 ? No,
because candidate 2 can position himself at position 1.

Candidate 2 Position Comparing Candidate 1 Utility at positions 2 and 3

Position 1 u1 (2, 1) = 90% > u1 (3, 1) = 85%

OK, what if we removed the dominated positions 1 and 10 ? Assume that

candidate 2 is rational and apply iterative elimination, i.e., candidate 2 will
never position himself in positions 1 and 10 similar to candidate 1.

Ayman Ghoneim (FCAI - Cairo University) Game Theory: Domination Spring 2024 27 / 30
Iterative Elimination: Politics Application

Does position 3 dominate position 2 ? Yes

Candidate 2 Position Comparing Candidate 1 Utility at positions 1 and 2

Position 2 u1 (2, 2) = 50% < u1 (3, 2) = 80%
Position 3 u1 (2, 3) = 20% < u1 (3, 3) = 50%
Position 4 u1 (2, 4) = 25% < u1 (3, 4) = 30%
Position 5 u1 (2, 5) = 30% < u1 (3, 5) = 35%

Thus, positions 2 and 9 are dominated once we realized that positions 1 and
10 will not be considered by the two candidates.

Ayman Ghoneim (FCAI - Cairo University) Game Theory: Domination Spring 2024 28 / 30
Iterative Elimination: Politics Application

Using the previous analysis, iterative elimination of dominated strategies

(positions) leads to deleting all positions except positions 5 and 6.
Conclusion: Candidates will position themselves close to the center of the
political spectrum.
This is a famous model and result in political sciences, known as the median
voter theorem (Anthony Downs 1957).
It is also a famous application in economics known as product placement
(Hotelling 1929).

Ayman Ghoneim (FCAI - Cairo University) Game Theory: Domination Spring 2024 29 / 30
Iterative Elimination: Politics Application

Is this model realistic ? What is it missing ?

Voters are not evenly distributed.
Missing the issue of not turning out for voting.
Assuming a one-dimensional political spectrum. (higher dimensions do
exist)
Assuming two candidates only. (there is always an additional candidate
if people didn’t turn out for voting)
Candidates may be lying and voters will not believe them (their
positions) unless they committee to policies which reflect their claimed
positions.

Any model is an abstraction and will always miss things out. Start with a
basic model for the problem under investigation, and then keep gradually
adding more factors to enrich the basic model.

Ayman Ghoneim (FCAI - Cairo University) Game Theory: Domination Spring 2024 30 / 30