Professional Documents
Culture Documents
Theory
Presented as an undergraduate
class in
Multimedia Mathematics
Paul Trafford
paul.trafford@stx.oxon.org
6 July 2011
Characteristics of Game
Theory
What is it about?
Fundamentally about the study of decision-making
Investigations are concerned more with choices and
strategies than best solutions.
It seeks to answer the questions:
i. What strategies are there?
ii. What kinds of solutions are there?
Examples:
. Chess, Go, economic markets, politics, elections, family
relationships, etc.
6
History (1)
The study of games is many centuries old.
More systematic developments in Game Theory
took place in the first half of the 20th Century.
Main Founders
John Von Neumann (mathematician)
Oskar Morgenstern (economist)
Image sources: Los Alamos National Laboratory, http://
www.lanl.gov/history/atomicbomb/images/NeumannL.GIF
and American Mathematical Society, http://www.ams.org/samplings/feature-column/fcarc-rationality
7
History (2)
Main publication: von Neumann & Morgenstern:
Theory of Games and Economic Behaviour. Princeton
University Press, 1944.
Goal:
Application of mathematical methods to broadly
analyse games
A new scientific approach to the study of economics.
Applications:
Aided by computers, theory has been broadly
applied in large-scale operations such as
international trade.
8
Game
Theory
Mathematic
s
Economic
s
Strategy
How should one play the game?
Definition: A strategy is a plan of action by which
a player has a decision rule to determine their set of
moves for every possible situation in a game.
A strategy is said to be pure if it at every stage in the
game it specifies a particular move with complete
certainty.
A strategy is said to be mixed if it applies some
randomisation to at least one of the moves.
For each game, there are typically multiple pure
strategies. Note that the randomisation is a set of fixed
probabilities, where the sum of the probabilities is 1.
11
12
A selection of games
Photo credits: Morten Johannes Ervik [Go], Jose Daniel Martinez [Chess], William Hartz
(Scrabble), David ten Have (Ludo), WikiJET (Janken), Cyron Ray Macey (Tic Tac Toe),
Dayland Shannon (Monopoly), Denise Griffin (Bridge), Steve Snodgrass (Draughts)
15
How to classify?
There are a number of [orthogonal]
criteria that may be used as the basis
for classifying games.
A common one uses two: im/perfect
information and
chance/not
chance.
Nondeterministic
(Chance
moves)
Deterministic
(No chance
Perfect
Information
Imperfect
Information
16
Classification of games:
Perfect Information
Imperfect Information
Chance
Moves
No
chance
moves
Photo credits: Morten Johannes Ervik [Go], Jose Daniel Martinez [Chess], William Hartz (Scrabble), David
ten Have (Ludo), WikiJET (Janken), Cyron Ray Macey (Tic Tac Toe), Dayland Shannon (Monopoly), Denise
Griffin (Bridge), Steve Snodgrass (Draughts)
17
18
19
Water plant
Add fertiliser
Communicate with plant
Place in sunlight
Shelter plant
It rains
It is stormy (heavy rain and
wind)
It is sunny
There is frost
Payoffs
How to Model?
Photo credit: Manjith Kainickara http://www.fotopedia.com/items/flickr-1061718736
22
Rules
Two players take turns removing objects from a
single heap or pile of objects.
On each turn, a player must remove exactly one or
two objects.
The winner is the one who takes the last object
Demonstration:
http://education.jlab.org/nim/index.html
23
24
25
Rules
Two players take turns removing objects from a single
heap or pile of objects.
On each turn, a player must remove exactly one or two
objects.
The winner is the one who takes the last object
Strategy
Leave a multiple of 3.
Demonstration: http://education.jlab.org/nim/index.html
26
28
29
30
B1
B2
A1
A2
-2
(Note the payoffs are the values that will be given to Player A)
Each row or column of payoffs is called an imputation.
Player A has two moves: A1 and A2.
Player B has two moves: B1 and B2.
The payoff for a game is given by the intersection. Thus if the
moves are respectively A1 and B2, then the payoff is zero.
32
B1
B2
A1
A2
-2
33
B1
B2
vL
4
2
A2
1
3
2*
vM
3*
A1
35
B1
B2
vL
A1
0*
A2
-2
-2
vM
0*
36
Simplification using
Dominance
For larger matrices, we may often simplify.
B1 B2 B3 B4 B5
A 4
5
6
1
5
1
A for4simplification
3
5
3is to4 compare pairs of
The main technique
columns, C and C, 2
say, and delete those columns where the
payoff in C is always greater than that in C or vice versa. In this
case we say C dominates C. (Similarly for rows).
B2
B4
3*
38
39
Mixed Strategies:
Expectation 1/2
Scenario: Game is played repeatedly. In this case choosing
the same pure strategy is not always optimal, so we can
vary these pure strategies.
To determine how we vary the strategies, we can apply
probability theory.
Key concept is Expectation := the product of the
probability of the occurrence of an event and the value
associated with the occurrence of a given event.
A player can use a mixed strategy this is more than one
pure strategy, where each pure strategy is played
randomly according to a fixed probability yielding an
expected payoff.
40
Mixed Strategies:
Expectation 1/2
We
then can determine the expected value of a
game.
Formally, as before, we denote the moves available
to Players A and B as A1,A2, ,An and B1,B2, ,Bm
respectively. Suppose the moves in As mixed
strategy are played with probabilities x = (x1, x2, ,
xn); and for B, y = (y1, y2, , ym). Suppose the
payoffs are given by .
Then the games expected value for A is E(x), where
The expectation is the same for Player B.
41
B1 B2
A 1
4
1
3
2A are A 1,A2 and to
The moves available toA Player
2 the moves in As mixed
Player B: B ,B . Suppose
1
42
B1
B2
A1
A2
Player A:
45
B2
4
2
(1)1*x + 3*y = v
(2)4*x + 2*y = v
B1
B2
4
2
48
Minimax Limitations
Whilst the Minimax theorem provides a solution, its
macro-oriented, i.e. not sensitive to individual variations.
Thus
It ensures an average payoff
Assumes repeated play and is a result that is more
reliable the more times played
In practice, it takes no account of the strategy of the
opponent even if they keep playing the same pure
strategy, the expected return is no more, no less
The optimisation reflects a collective philosophy that
markets find their natural level.
49
PART D: Non-Zero-Sum
Games
50
An Overview of Non-Zero-Sum
Games
[Recap] A game is said to be non-zero-sum if wealth may
be created or destroyed among the players (i.e. the total
wealth can increase or decrease).
In general, unlike for zero-sum games, in non-zero-sum
games, wealth can be mutually created through
cooperation.
Cooperation may be achieved whether or not there is
direct communication.
Where there is no communication, information is
necessarily imperfect.
Where there is communication, there may be bargaining.
51
Analysis of Non-Zero-Sum
Games
Methods of mathematical logical, such as use of induction, are
effective for determining strategies in Zero-sum games with
perfect information. However they are less so for games of
imperfect information, and are often not applicable to non-zero
games.
IF some assumptions are made THEN some mathematical
techniques may be effectively applied.
Prerequisites:
Understand the environment,
understand the individual and collective psychology
(Thus we are moving from the domain of pure mathematics to
embrace social sciences, particularly psychology and economics.)
52
Utility
Payoffs are given as utility the perceived worth of
something
Utility is a key concept and is determined by social
and psychological factors.
They depend upon personal preferences
The same material payoff may have different utility
(In economics, personal preference is often reckoned
in terms of ranking a selection of consumer offerings.
[Economic] agents are said to be rational if this
ranking system is complete.)
53
100% chance
20% chance
54
B1
B2
A1 (-1,-1)
(-10,0)
A2 (0,-10)
(-5,-5)
58
64
W1 W2
H1 (4,1) (0,0
)
H2 (0,0) (1,4
)
65
Hence
E1=5xy x-y+1 and E2= 5xy+-4x-4y+xy
67
69
Conclusions
Games occur in many life situations
Mathematical analysis requires
understanding of the context and
rules
Games played repeatedly yield
different responses from games
played only once.
Human psychology often yields
unexpected behaviour.
72