Theory of Games

"Game theory attempts to study decision-making in situations where two or more intelligent
and rational opponents are involved under conditions of conflict and cooperation. The approach of
game theory is to seek to determine a rival’s most profitable counter-strategy to one's own best
moves and to formulate the appropriate defensive measures.”
Game theory is a type of decision theory in which one’s choice of action is determined after
taking into account all possible alternatives available to an opponent playing the same game, rather
than just by the possibilities of several outcomes.

The mathematical analysis of competitive problems is fundamentally based upon the minimax
(maximin) criterion of J. Von Neumann (called the father of game theory). This criterion implies the
assumption of rationality from which it is argued that each player will act to maximize his minimum
gain or minimize his maximum loss.

Definition:
Game is defined as an activity between two or more persons on involving activities by each
person according to a set of rules, at the end of which each person receives some benefit or
satisfaction or suffers loss (negative benefit).
Characteristics of game theory
1. Chance of Strategy: If in a game activities are determined by skill, it is said to be a game of
strategy, if they are determined by chance, it is a game of chance. In general, a game may involve
game of strategy as well as a game of chance.
2. Number of persons: A game is called an n- person game if the number of persons playing is n. The
person means an individual or a group of particular object.
3. Number of activities: These may be finite or infinite.
4. Information to the players about the past activities of other players is completely available, partly
available, or not available at all.
5. Payoff: A quantitative measure of satisfaction a person gets at the end of each play is called a
payoff.

Basic Definitions
Competitive Game: A competitive game has the following four properties.
a) There is finite number (n) of competitors called players.
b) Each player has a finite number of strategies (alternatives) available to him.
c) A play of the game takes place when each player employs his strategy.
d) Every combination of activities determines an outcome which results in a gain of payments to each
player.

Zero-sum Games: Competitive games are classified according to the number of players involved, i.e.
as a two person game, three person game, etc. If the players make payments only to each other i.e.,
the loss of one is the gain of others, and nothing comes from outside, the competitive game is said to
be zero-sum. Zero-sum games are also called strictly competitive. Mathematically, suppose an n-
person game is played by n players P1, P2, … Pn whose respective pay-offs at the end of the game are
a1, a2, … an then the game will be called zero-sum if ∑𝑛𝑖=1 𝑎𝑖 = 0 at each play of the game.

Non-zero-sum Games: A game which is not zero-sum is called a non-zero-sum game. When one
player's gain does not necessarily mean another player's loss (and vice versa). These types of games
are referred to as non-zero-sum games, because the gains and the losses in the game do not always
add up to zero. Poker game is an Example.

Strategy: A strategy for a player is defined as a set of rules or alternative courses of action available to
him in advance, by which player decides the course of action that he should adopt. A strategy may be
of two types:
a) Pure Strategy: If a player knows exactly what the other player is going to do, a deterministic
situation is obtained and objective function is to maximize the gain or to minimize the loss.
Therefore, the pure strategy is to select a particular course of action (same strategy each time).
b) Mixed Strategy: If a player is guessing as to which course of action is to be selected by the other
players on any particular occasion, a probabilistic situation is obtained and objective of the player
is to maximize expected gains or to minimize expected losses. Thus, the mixed strategy is a
selection among pure strategies with fixed probabilities.
 Mathematically, a mixed strategy for a player with m ( ≥ 2) possible courses of action is denoted
by the set S of m non-negative real numbers whose sum is unity, representing probabilities with which
each course of action is chosen. If 𝑥𝑖 (𝑖 = 1,2, … , 𝑚) is the probability of choosing the course i, then
𝑆 = (𝑥1 , 𝑥2 , … , 𝑥𝑚 ) Subject to the conditions 𝑥1 + 𝑥2 + . . . +𝑥𝑚 = 1 and 𝑥1 ≥ 0, 𝑥2 ≥ 0, … , 𝑥𝑚 ≥ 0.
Note: The difference between pure and mixed strategy games is that pure strategy games processes a
saddle point where as mixed strategy games do not.

Two-Person zero-sum Game: A game of two persons (say player A and Payer B), in which the gains of
one player are equivalent to the losses of the other player is called a Two person zero-sum game, i.e.,
in a two-person zero-sum game, the algebraic sum of their net gains to both the players is zero. Two-
person zero-sum games are also known as rectangular games since they are represented by
rectangular payoff matrix.

Pay-off Matrix: A strategy is a course of action taken by one of the participant, and the payoff (gains
or losses) is the result or outcome of the strategy can be represented in the form of the matrix called
the payoff matrix.
Suppose the player A (always the gainer) has m strategies 𝐴1 , 𝐴2 , … 𝐴𝑚 and player B (always
the loser) has n strategies 𝐵1 , 𝐵2 , … 𝐵𝑚 . Then a payoff matrix can be represented by adopting the
following rules:
a) Row designations for each matrix are activities available to player A.
b) Column designations for each matrix are activities available to player B.
c) Cell entry 𝑎𝑖𝑗 is the payment to player A in A's payoff matrix when A chooses the activity i and B
chooses the activity j.
d) With a two-person zero-sum game, the cell entry in the player B's payoff matrix will be negative of
the corresponding cell entry 𝑎𝑖𝑗 in the player A's payoff matrix so that sum of payoff matrices for
player A and player B is ultimately zero.

The Payoff matrix to player B is (−𝑎𝑖𝑗 ).

Consider a coin tossing game with two persons only. Each player tosses an unbiased coin
simultaneously and gets either a head (H) or a tail (T). If (H, H) or (T, T) occurs; then player B pays Rs. 5
to player A, otherwise player B wins Rs. 2 from A. This game is a two-person zero-sum game, since the
winnings of one player taken as the losses for the other. In this game each player has two strategies
(H or T) and this yields the following 2x2 game matrix expressed in terms of the payoff to Player A.
Player B
𝐻 𝑇
𝐻 5 −2
Player A [ ]
𝑇 −2 5
Maximin and Minimax Criterion: According to this principle, the player adopts a pessimistic attitude
and plays safe i.e., his strategy is always that which results in the best out of the worst outcomes. In
other words the player A (the maximizing player) decides to play that strategy which corresponds to
the maximum of the minimum gains for his different courses of action. This is known as Maximin
Criterion. Similarly, the player B (the minimizing player) would also like to play safe and in that case he
selects that strategy which corresponds to the minimum of the maximum losses for his different
courses of action and this is known Minimax Criterion.

Optimal strategy: A course of action that puts any player in the most preferred position, irrespective
of the course of action his competitor(s) adopt, is called as optimal strategy. In other words, if the
maximin value equals the minimax value, then the game is said to have a saddle point. The payoff at
the saddle point is called Value of the game and the corresponding strategies are called optimal
strategies.
Mathematically, if a payoff matrix (𝑎𝑖𝑗 ) is such that 𝑚𝑎𝑥[min(𝑎𝑖𝑗 )] = 𝑚𝑖𝑛[max(𝑎𝑖𝑗 )] = 𝑣
then the matrix is said to have a saddle point.
Remarks:
a. A saddle point of a matrix is, sometimes called the equilibrium point of the payoff matrix.
b. A game is said to be a fair game if the value of the game is zero.
c. A game is said to be strictly determinable if the maximin and minimax values of the game are
equal and both equal to the value of the game.

Solution of Pure Strategy Games (with Saddle Point)

Rules to determine Saddle Point
1. Select the minimum (lowest) element in each row of the payoff matrix and note under row
minima. Then, select the largest element among these elements and enclose it in a rectangle,
2. Select the maximum (largest) element in each column of the payoff matrix and note under column
maxima. Then select the lowest element among these elements and enclose it in a circle,
3. Identify the element(s) that is same in the circle as the well as rectangle and mark the position of
such element(s) in the matrix. This element represents the value of the game and is called the
saddle point.

Solution of Mixed Strategy Games (without Saddle Point)

In certain cases, no saddle point exists, i.e. maximin value ≠ minimax value. In all such cases,
players must choose the mixture of strategies to find the value of game and an optimal strategy. The
value of game obtained by the use of mixed strategies represents the least payoff, which player A can
expect to win and the least which player B can expect to lose. The expected payoff to a player in a
game with payoff matrix (𝑎𝑖𝑗 ) of order m × n is defined as: 𝐸(𝑝, 𝑞) = ∑𝑚 𝑛 𝑇
𝑖=1 ∑𝑗=1 𝑝𝑖 𝑎𝑖𝑗 𝑞𝑗 = 𝑃 𝐴 𝑄
where 𝑃 = (𝑝1 , 𝑝2 , … , 𝑝𝑚 ) and 𝑄 = (𝑞1 , 𝑞2 , … , 𝑞𝑛 ) denote probabilities associated with m strategies
of player A and n strategies of player B respectively.

Maximin-Minimax Criterion: Consider an m x n game (𝑎𝑖𝑗 ) without any saddle print i.e., strategies
are mixed. Let 𝑝1 , 𝑝2 , … , 𝑝𝑚 and 𝑞1 , 𝑞2 , … , 𝑞𝑛 be the probabilities with which player A and Player B will
play their moves with 𝐴1 , 𝐴2 , … , 𝐴𝑚 and 𝐵1 , 𝐵2 , … , 𝐵𝑛 respectively. Here 𝑝𝑖 ≥ 0 (𝑖 = 1,2, … , 𝑚), 𝑞𝑗 ≥
0 (𝑗 = 1,2, … , 𝑛) and 𝑝1 + 𝑝2 + … + 𝑝𝑚 = 1, 𝑞1 + 𝑞2 + … + 𝑞𝑛 = 1.
The expected payoff function for player A. therefore, will be given by 𝐸(𝑝, 𝑞) = ∑𝑚 𝑛
𝑖=1 ∑𝑗=1 𝑝𝑖 𝑎𝑖𝑗 𝑞𝑗
Making use of maximin-minimax criterion, we have
For Player A,
𝑚

𝑣 = 𝑚𝑎𝑥. 𝑚𝑖𝑛. 𝐸(𝑝, 𝑞) = max [min (∑ 𝑝𝑖 𝑎𝑖𝑗 )]

𝑖=1
𝑚 𝑚 𝑚

max [min (∑ 𝑝𝑖 𝑎𝑖1 , ∑ 𝑝𝑖 𝑎𝑖2 , … , ∑ 𝑝𝑖 𝑎𝑖𝑛 ,)]

𝑖=1 𝑖=1 𝑖=1
Here 𝑚𝑖𝑛(∑𝑚 th
𝑖=1 𝑝𝑖 𝑎𝑖𝑗 ,) denotes the expected gain to player A, when player B uses his j strategy.
For Player B,
𝑛 𝑛 𝑛

𝑣 = min [max (∑ 𝑞𝑗 𝑎1𝑗 , ∑ 𝑞𝑗 𝑎2𝑗 , … , ∑ 𝑞𝑗 𝑎𝑚𝑗 ,)]

𝑗=1 𝑗=1 𝑗=1

Here 𝑚𝑎𝑥(∑𝑛𝑗=1 𝑞𝑗 𝑎𝑖𝑗 ,) denotes the expected loss to player A, when player B uses his ith strategy.
The relationship 𝑣 ≤ 𝑣 holds well in general and when 𝑝𝑖 and 𝑞𝑗 correspond to the optimal strategies
the relation holds inequality sense and the expected value for both the players becomes equal to the
optimum expected value of the game.

Arithmetic Method:
It is an easy method for finding optimal strategies for each player in a payoff matrix of size 2x2
without saddle point. steps for this method are

Step-1: Find the difference between the values in the first row and put it against the second row of
the matrix, neglecting sign.
Step-2: Find the difference between the values in the second row and put it against the first row of
the matrix, neglecting sign.
Step-3: Repeat steps 1 and 2 for columns also. The matrix thus obtained by swapping the difference
(oddments) represents the optimal relative frequencies of payoff for both player strategies. This may
be converted to probabilities by dividing each of them by their sum.

Note:
1. Arithmetic method is easier than the algebraic method but it cannot be applied to larger games.
2. This method should not be used to the game that has a saddle point because the method yields an
incorrect solution.
3. Some of row oddments and column oddments must be equal then only we can calculate optimal
strategies

Dominance Property
In dominance, we can reduce the size of the payoff matrix by deleting those strategies which
are dominated by the others. Thus if each element in one row, say k th of the payoff matrix (𝑎𝑖𝑗 ) is less
than or equal to the corresponding elements in some other row, say rth, then player A will never
choose kth strategy. In other words, probability Pk = P (choosing the kth strategy) is zero, if 𝑎𝑘𝑗 ≤ 𝑎𝑟𝑗
for all j = 1, 2, … , n.
The value of the game and the non-zero choice of probabilities remain unchanged even after
the deletion of kth row from the payoff matrix. In such a case the kth strategy is said to be dominated
by the rth one.
General rules for dominance are:
a. If all the elements of a row (say kth) are less than or equal to the corresponding elements of any
other row (say rth) then kth row is dominated by rth row.
b. If all the elements of a column, say kth are greater than or equal to the corresponding elements of
any other column (say rth) then kth column is dominated by the rth column.
c. Dominated rows or columns may be deleted to reduce the size of payoff matrix, as the optimal
strategies will remain unaffected.

The Modified Dominance Property: The dominance property is not always based on the superiority of
pure strategies only. A given strategy can also be said to be dominated if it is inferior to an average of
two or more other pure strategies. More generally, if some convex linear combination of some rows
dominates the ith row, then ith row will be deleted. Similar arguments follow for columns.

Remark: Rules (principles) of dominance are used when the payoff matrix is a profit matrix for the
player A and a loss matrix for player B. Otherwise the principle gets reversed.

Graphic Solution of 2xn and mx2 Games:

Graphic solution of 2xn and mx2 games enables us to reduce the 2xn and mx2 game to a simple 2x2
game. Consider the following 2xn game:

Player B
𝐵1 𝐵2 … 𝐵𝑛
𝐴1 𝑎11 𝑎12 … 𝑎1𝑛
Player A ( 𝑎22 … 𝑎2𝑛 )
𝐴2 𝑎21

It is assumed that the game does not have a saddle point. Let the optimum mixture strategies
𝐴 𝐴2
for A be given by 𝑆𝐴 = [ 1 ]where p1 + p2 = 1. The average (expected) payoff for A when he plays
𝑝1 𝑝2
𝑆𝐴 against these pure moves 𝐵1 , 𝐵2 , … 𝐵𝑛 is given by

According to the maximin criterion for mixed strategy games, player A should select the values of p1
and p2 so as to maximize his minimum expected payoffs. This may be done by plotting the expected
payoff lines: 𝐸𝑗 (𝑝1 ) = (𝑎1𝑗 − 𝑎2𝑗 )𝑝1 + 𝑎2𝑗 (𝑗 = 1,2, … , 𝑛)
The highest point on the lower envelope of these lines will give maximum of the minimum
(i.e., maximin) expected payoffs to player A as also the maximum value of pi.

The two lines* passing through the maximin point identify the two critical moves of B which,
combined with two of A, yield the 2x2 matrix that can be used to determine the optimum strategies of
the two players, for the original game, using the results of the previous section.
The (m x 2) games are also treated in the same way where the upper envelope of the straight lines
corresponding to B's expected payoffs will give the maximum expected payoff to player B and the
lowest point on this then gives the minimum expected payoff (minimax value) and the optimum value
of q1.

*If there are more than two lines passing through the maximin point, there are ties for the optimum
mixed strategies for player B. Thus, any two such lines with opposite sign slopes will define an
alternative optimum for B.

Algebraic Method for rectangular games:

The payoff for a rectangular game can always be given by mxn matrix, when player A has m possible
course of actions and player B has n possible course of actions and the payoff matrix is (𝑎𝑖𝑗 ) as shown
below

It can be shown mathematically that

1. Each rectangular game has a specific value 𝑉. This value is unique.
2. There exists a best strategy 𝑋 for player A i.e., there exists frequencies 𝑋1 , 𝑋2 … 𝑋𝑚 such that
𝑋1 + 𝑋2 + ⋯ + 𝑋𝑚 = 1 and player A can assure himself, at least an expected gain of V, the value
of the game.
3. Similarly, for player B, there exists best strategy 𝑌 = 𝑌1 , 𝑌2 … 𝑌𝑛 ; ∑ 𝑌𝑗 = 1 and Player B assures
himself at most the loss of 𝑉, it can be shown that the unknown
frequencies𝑋1 , 𝑋2 … 𝑋𝑚 , 𝑌1 , 𝑌2 … 𝑌𝑛 and 𝑉 can be found from the following relation:
a) 𝑋1 + 𝑋2 + ⋯ + 𝑋𝑚 = 1; 𝑋𝑖 ≥ 0 ----------> (1)
b) 𝑌1 + 𝑌2 + ⋯ + 𝑌𝑛 = 1; 𝑌𝑗 ≥ 0 ----------> (2)
c) 𝑋1 𝑎1𝑗 + 𝑋2 𝑎2𝑗 + ⋯ + 𝑋𝑚 𝑎𝑚𝑗 ≥ 𝑉 ∀(𝑗 = 1, 2, … , 𝑛) ---------->(3)
d) 𝑌1 𝑎𝑖1 + 𝑌2 𝑎𝑖2 + ⋯ + 𝑌𝑚 𝑎𝑖𝑚 ≤ 𝑉 ∀(𝑖 = 1, 2, … , 𝑚) ----------> (4)

The algebraic method is a direct method to solve the unknown from the equations (1), (2), (3) and
from the fundamental theorem of games. The value of the game 𝑉 is unique, Therefore, the idea is to
find such value 𝑉 which satisfies all the above four equations by using the first step i.e., to assuming
the inequations (3) & (4) as equalities.

General Solution of mxn Rectangular Games:

In a rectangular game with an mxn pay off matrix, if there does not exist any saddle point and also it is
not possible to reduce the size of the game mx2 or 2xn payoff matrix by using the concept of
dominance, the following methods are generally used to solve the game:
1. Linear Programming Method and
2. Iterative Method for approximate solution
Linear Programming Method
Let us consider an mxn payoff matrix (𝑎𝑖𝑗 ) for player A. Let
𝐴 , … , 𝐴𝑚 𝐵 , … , 𝐵𝑛
𝑆𝐴 = [ 1 … , 𝑝 ] and 𝑆𝐵 = [ 1 … , 𝑞 ] where ∑𝑚 𝑛
𝑖=1 𝑝𝑖 = ∑𝑗=1 𝑞𝑗 = 1
𝑝1, 𝑚 𝑞1, 𝑛
be the mixer strategies for two players respectively.
Then the expected gains 𝑔𝑗 (𝑗 = 1, 2, … , 𝑛) of player B against B’s pure strategies will be

𝑔1 = 𝑎11 𝑝1 + 𝑎21 𝑝2 + ⋯ + 𝑎𝑚1 𝑝𝑚

𝑔2 = 𝑎12 𝑝1 + 𝑎22 𝑝2 + ⋯ + 𝑎𝑚2 𝑝𝑚
⋮
𝑔𝑛 = 𝑎1𝑛 𝑝1 + 𝑎2𝑛 𝑝2 + ⋯ + 𝑎𝑚𝑛 𝑝𝑚

The expected losses 𝑙𝑖 (𝑖 = 1, 2, … , 𝑚) of player B against A’s pure strategies will be

𝑙1 = 𝑎11 𝑞1 + 𝑎12 𝑞2 + ⋯ + 𝑎1𝑛 𝑞𝑛

𝑙2 = 𝑎21 𝑞1 + 𝑎22 𝑞2 + ⋯ + 𝑎2𝑛 𝑞𝑛
⋮
𝑙𝑛 = 𝑎𝑚1 𝑞1 + 𝑎𝑚2 𝑞2 + ⋯ + 𝑎𝑚𝑛 𝑞𝑛

The objective of player A is to select 𝑝𝑖 (𝑖 = 1, … , 𝑚) such that he can maximize his minimum
expected gains; and the player B desires to select 𝑞𝑗 (𝑗 = 1, … , 𝑛) that will minimize his expecte losses.
Thus, if we let 𝑢 = 𝑚𝑖𝑛 ∑𝑚 𝑛
𝑖=1 𝑎𝑖𝑗 𝑝𝑖 (𝑗 = 1, 2, … , 𝑛) and 𝑣 = 𝑚𝑎𝑥 ∑𝑗=1 𝑎𝑖𝑗 𝑞𝑗 (𝑖 = 1, 2, … , 𝑚); the
problem of two players could be written as:
Player A
1 𝑝𝑖
𝑀𝑎𝑥𝑖𝑚𝑖𝑧𝑒 𝑢 = 𝑀𝑖𝑛𝑖𝑚𝑖𝑧𝑒 𝑢 = ∑𝑚 𝑖=1 𝑢 subject to the constraints:
∑𝑚𝑖=1 𝑎𝑖𝑗 𝑝𝑖 ≥ 𝑢 𝑎𝑛𝑑 ∑ 𝑝𝑖 = 1, 𝑝𝑖 ≥ 0 (𝑖 = 1, 2, … , 𝑚)
Player B
1 𝑞𝑗
𝑀𝑖𝑛𝑖𝑚𝑖𝑧𝑒 𝑣 = 𝑀𝑎𝑥𝑖𝑚𝑖𝑧𝑒 𝑣 = ∑𝑚 𝑖=1 𝑣 subject to the constraints:
∑𝑚𝑖=1 𝑎𝑖𝑗 𝑞𝑗 ≥ 𝑣 𝑎𝑛𝑑 ∑ 𝑞𝑗 = 1, 𝑞𝑗 ≥ 0 (𝑗 = 1, 2, … , 𝑛)

Assuming that 𝑢 > 0 and 𝑣 > 0, we introduce new variables defined by

𝑝𝑖 𝑞𝑗
𝑝𝑖′ = 𝑎𝑛𝑑 𝑞𝑗′ = (𝑖 = 1, 2, … , 𝑚; 𝑗 = 1,2, … 𝑛)
𝑢 𝑣
Then, the pair of linear programming problems can be re-written Player A
Player A
𝑀𝑖𝑛𝑖𝑚𝑖𝑧𝑒 𝑝0 = 𝑝1′ + 𝑝2′ + ⋯ + 𝑝𝑚 ′
subject to the constraints:
′ ′ ′ ′
𝑎1𝑗 𝑝1 + 𝑎2𝑗 𝑝2 + ⋯ + 𝑎𝑚𝑗 𝑝𝑚 ≥ 1 𝑎𝑛𝑑 𝑝𝑖 ≥ 0 (𝑖 = 1, 2, … , 𝑚; 𝑗 = 1,2, … 𝑛)
Player B
𝑀𝑎𝑥𝑖𝑚𝑖𝑧𝑒 𝑞0 = 𝑞1′ + 𝑞2′ + ⋯ + 𝑞𝑛′ subject to the constraints:
𝑎1𝑗 𝑞1′ + 𝑎2𝑗 𝑞2′ + ⋯ + 𝑎𝑚𝑗 𝑞𝑛′ ≤ 1 𝑎𝑛𝑑 𝑞𝑗′ ≥ 0 (𝑖 = 1, 2, … , 𝑚; 𝑗 = 1,2, … 𝑛)

Remarks:
1. It is easy to note that the L.P.P.'s of the two players represent a primal-dual pair. Therefore, by
fundamental theorem of duality one can read off the optimal solution of one player, just from the
optimum simplex table of the opponent. That is, we need to solve just one player's L.P.P. by
simplex method.
2. Linear programming technique requires all variables to be non-negative and, therefore, to obtain a
non-negative value of the game, the data to the problem, i.e., 𝑎𝑖𝑗 in the payoff table should all be
non-negative. If there are some negative elements in the payoff table, a constant to every element
in the payoff table must be added so as to make the smallest element zero; the solution to this
new game will give an optimal mixed strategy for the original game. The value of the original game
then equals the value of the new game minus the constant.

Iterative Method of Approximate Solution

Iterative method can be applied to solve 3×3 or higher games which cannot be solved by
Arithmetic method and are extremely tedious to be solved by method of linear programming. This
method gives an approximate solution for the value of the game and the true value can be
determined to any degree of accuracy. Optimum strategies can also be determined but not so
satisfactorily. The method assumes that each player acts under the assumption that past is the best
guide for the future and will play in such a manner so as to maximize the expected gain or to minimize
the expected loss.