Game Theory: an Online Course

Jackson and Shoham

copyright 2012
do not circulate or reproduce without permission
 At a Glance
• Six weeks
– Lecture videos posted at the beginning of week; in-lecture quizzes
– Weekly review quiz
– Weekly games in Lab
– Weekly “screenside chat”

• Resources
– Lecture videos
– Slides
– Booklet
 Assignments, Grading and Certificate
• Grades
– Weekly problem sets and final exam count towards the grade
• Problem sets 70%
• Final exam 30%
– Quizzes (both in-video and weekly review) and labs don’t count towards grade
– You can answer as often as you like, but only first attempt counts

• Certificate
– Cumulative score above 70% earns you an (electronic) certificate of completion
– Cumulative score above 90% earns you an (electronic) designation “with distinction”
– No, this doesn’t count towards a Stanford degree, sorry…

• Honor code
– By all means consult with each other, but please don’t share answers to quizzes, problem sets, or final
exam. Not directly, and indirectly by posting them anywhere. Thanks.
 Community Support
• Discussion forums on class page
– Post questions, answers
– Find and fix bugs
– Translate to other languages; etc.
– We will acknowledge the most useful contributors

• Study groups
– Arrange on Study Group forum
– Conduct on Google+ Hangout, or email
 Introduction: Game Theory
• Models of strategic interactions:
– Outcomes depend on more than one player’s decision

• The decisions for one player cannot be separated from those

of others!

• Models miss (important?) aspects of reality

– That is why they are models
 Wide set of applications
• War: Why do some countries arm heavily while others do
not?
• Tragedy of the commons: Why do we need international
agreements on fishing and how should they be structured?
• Markets: What will happen if two companies are allowed to
merge?
 Wide Set of Applications
• E-commerce: How should an on-line auction be structured
to maximize revenue? How should sellers be policed?
• Legal: How should random audits of taxes be conducted?
• Sports: Should a soccer player adjust the fraction of time
that s/he kicks penalty kicks to a goalies right or left based on
the goalie?
 Game Theoretic Models
• Provide insights when interactions affect incentives

• Applications in many fields

• Methodology/tool box
– tools will vary with the applications!
 Questions About Models

• Players: who makes the decisions?

• Strategies: what are the actions available?
• Timing: who does what and when?
• Information: what do players know when choosing?
• Payoffs: what happens as a function of the actions? and,
what motivates players?
• ``Solution’’ : how do we predict what will happen?
 Course Outline

• Normal Forms

• Extensive Forms

• Repeated Games

• Incomplete Information: Bayesian Games

 Normal Form
• An example of a normal form game

• The formal model

– The main ingredients: Players, Strategies, Payoffs
– Normal Form games
• Dominance
• Equilibrium
 Predator Prey Games
• Lets examine an analysis based on a paper by Chen Shi and
Sheng Bao 2008 ``A Game Theory Based Predation Behavior
Model’’ (Center for Game Theory in Economics Archive, 2010 Conference)
http://www.gtcenter.org/Archive/2010/Conf/Chen956.pdf

• Strategies: Active or Passive

– Predators - active: dragonfly, passive: spider
– Prey – active: fly, passive: mealybugs
 ``Payoffs’’
• Predator: expected calories captured less expected calories
expended
– Active raises calories expended and captured

• Prey: penalty for expected death rate less expected calories

expended
– Active raises calories expended and lowers chances of death

• But the expected capture and expended calorie rates are interactive
 Predator/Prey

Pred\Prey Active Passive

Active 1.7, -.8 3,-1
Passive 1.6, -.7 0,0

• Large Mammal estimates

 Predator/Prey

Pred\Prey Active Passive

Active 1.7, -.8 3,-1
Passive 1.6, -.7 0,0

• Large Mammal estimates

 Predator/Prey

Pred\Prey Active Passive

Active 1.7, -.8 3,-1
Passive 1.6, -.7 0,0

• Large Mammal estimates

 Predator/Prey

Pred\Prey Active Passive

Active 1.7, -.8 3,-1
Passive 1.6, -.7 0,0

• Large Mammal estimates

 Are most large mammals active?
• predators: lions, tigers, leopards, wolves, coyotes, dogs,
cats….
• prey: deer, antelope, zebras, horses, mice, …
• But not all

• Simplified game - could enrich strategy set and payoff

structure
 Predator/Prey

Pred\Prey Active Passive

Active 2, -7 6, -8
Passive 3, -6 -1,0

• Small predator prey – insects

 Predator/Prey

Pred\Prey Active Passive

Active 2, -7 6, -8
Passive 3, -6 -1,0

• Small predator prey – insects

 Predator/Prey

Pred\Prey Active Passive

Active 2, -7 6, -8
Passive 3, -6 -1,0

• Small predator prey – insects

 Predator/Prey

Pred\Prey Active Passive

Active 2, -7 6, -8
Passive 3, -6 -1,0

• Small predator prey – insects

 Predator/Prey

Pred\Prey Active Passive

Active 2, -7 6, -8
Passive 3, -6 -1,0
 Predator/Prey

Pred\Prey Active Passive

Active 2, -7 6, -8
Passive 3, -6 -1,0

• No stable pair of strategies: prey wants to match, predator

wants to mismatch
 Predator/Prey

Pred\Prey Active Passive

Active 2, -7 6, -8
Passive 3, -6 -1,0

• No stable pair of strategies: prey wants to match, predator

wants to mismatch
• Randomization: mixed strategies
 Quiz  

 
Consider  the  following  modified  game:  

Pred\Prey   Ac<ve   Passive  

Ac<ve   2,  -­‐2   3,-­‐1  
Passive   1,  -­‐1   0,0  
 Quiz  
Use  the  logic  from  the  lecture  (for  large  mammals)  to  find  a  
stable  pair  of  pure  strategies  for  the  predator  and  prey:  
a) (Ac;ve,  Ac;ve);  
b) (Ac;ve,  Passive);  
c) (Passive,  Passive);  
d) (Passive,  Ac;ve);  
e) No  stable  pair.  
 Quiz:  Explana<on  
(b)  is  true.  
• If  Prey  is  ac;ve,  it  is  beCer  for  Predator  to  be  ac;ve  (earning  2)  than  
to  be  passive  (earning  1).  
• If  Prey  is  passive,  it  is  also  beCer  for  Predator  to  be  ac;ve  (earning  3)  
than  to  be  passive  (earning  0).  
• Thus  in  any  stable  point  the  Predator  chooses  to  be  ac;ve.  
• Given  Predator  is  ac;ve,  it  is  beCer  for  Prey  to  be  passive  (earning  
-­‐1)  than  be  ac;ve  (earning  -­‐2).  
• Thus  (Ac;ve,  Passive)  is  the  unique  stable  pair.  
 Normal Form
• An example of a normal form game

• The formal model

– The main ingredients: Players, Strategies, Payoffs
– ``Solution concepts’’
• Dominance
• Equilibrium
 Players
• Who makes the decisions?

• Do we model the US when it negotiates with another country

as a single player maximizing some payoff, or as a group of
politicians maximizing something else, or as a country full of
people interacting to make decisions?
 Strategies
• What are the possible actions for each player to take?

• Is passive/active enough to capture the strategies of a

predator/prey relationship?

• ...
 The Normal Form
• N={1,...,n} the set of players

• Ai the set of actions of player i

• A= A1 x ∙∙∙ x An profiles of actions

• ui: A → R utility function of player i

 Predator/Prey

Pred\Prey Active Passive

Active 2, -7 6, -8
Passive 3, -6 -1,0
N={Pred,Prey}
A1={Active,Passive} A2={Active,Passive}
u1(Active,Active)=2 ....
A Simple Collective-Action Game

Revolt Not
Revolt 1, 1 -1,0
Not 0,-1 0,0
 A Simple Collective-Action Game

Revolt Not
Revolt 1, 1 -1,0
Not 0,-1 0,0

N={1,2}
A1={Revolt, Not} A2={Revolt, Not}
u1(Revolt, Revolt)=1 ....
 Quiz  
Consider  the  following  normal  form:  
• N={1,  2}  
• Ai={Le<,  Right}  
• Player  1  prefers  to  choose  the  same  acBon  as  player  2:        u1
(a1=a2)=1,  u1(a1≠a2)=-­‐1  
• Player  2  prefers  to  choose  the  opposite  acBon  from  player  
1:    u2(a1=a2)=-­‐1,  u2(a1≠a2)=1  
 Quiz  

1  \  2   Le+   Right  
Le+   1,  -­‐1  
Right   ?,  ?  
What  should  be  filled  in  (?,  ?):  
 a)  1,  -­‐1  ;        
b) 0,  1  ;  
c) -­‐1,  1  ;  
d) 1,  0  .  
 Quiz:  ExplanaFon  
(c)  is  true.  
• In  (?,  ?),  a1=  Right  and  a2=  Le<,  such  that  a1  ≠  a2.  
• From  the  uBlity  funcBon,  we  know  u1(a1≠a2)=-­‐1  and  u2
(a1≠a2)=1.  
• Thus  the  payoff  is  (-­‐1,  1).  
 The utility function
• What should the payoffs be?

• What motivates players?

– own payoff?
– are they altruistic?
– do they care about relative outcomes?
– fairness…
 Two Interesting Families of Games

• Team games: Ultimate cooperation

• Zero-Sum games: Ultimate competition

Team Games

L R
L 1,1 0,0
R 0,0 1,1
Sidewalk Shuffle
 Team Games

• For any two players i and j and any action profile a, it is the
case that ui(a) = uj(a)

• A more general condition is that the players have identical

preference over strategy profiles. That is, for any two players i
and j and any two strategy profiles s1 and s2, it is the case that
ui(s1) ≥ ui(s2) iff uj(s1) ≥ uj(s2)

• AKA: “pure coordination”, “common-payoff” games

 Zero-Sum Games

R P S
H T R 0,0 -1,1 1,-1
H 1,-1 -1,1 P 1,-1 0,0 -1,1
T -1,1 1,-1 S -1,1 1,-1 0,0
Matching Pennies Rock-Paper-Scissors
(Rochambeau)
 Zero-Sum Games
• Limited to two-person games

• u1(a1,a2) + u2(a1,a2) = 0

• Any n-person game can be converted to an “equivalent” (n+1)-person zero-sum

game; hence only two-person ZS games are interesting as a category

• “Zero” is not critical; generalize to “constant-sum” games: u1(a1,a2) + u2(a1,a2) = k

• AKA: “pure conflict”, “pure competition”, “adversarial” games

• Most of the early work in game theory was on two-person zero-sum games
 Quiz  

 
Consider  the  following  zero-­‐sum  game:  

Le8   Center   Right  

Le8  
Center  
Right   -­‐5,  ?  
 Quiz  
What  should  be  filled  in  ?  :  
 
a) -­‐1;  
b) 0;  
c) -­‐5;  
d) 5.  
 Quiz:  Explana@on  
(d)  is  true.  
• In  a  zero-­‐sum  game,  u1(a1,a2)  +  u2(a1,a2)  =  0,  for  all  possible  
(a1,a2).    
• Thus  ?  =  u2(Right,  Right)=  0  –  u1(Right,  Right)  =  0  –  (-­‐5)  =  5.  
 Normal Form
• An example of a normal form game

• The formal model

– The main ingredients: Players, Strategies, Payoffs
– Reasoning about games
• From ``Optimal Strategy’’ to ``Solution concepts’’
• Dominance
• Equilibrium
 From ``Optimal Strategy’’ to
``Solution concepts’’
• Rationality means many things:

• In game theory it generally means that ``players maximize

their expected payoffs given their beliefs’’

• What makes game theory more complex than decision theory

is that beliefs of players are intertwined:
– what does player 1 believe player 2 will do?
 `Solving’ Normal Forms

• Dominance

• Equilibrium
 Dominance
• A strategy always leads to at least as high (higher) payoff than
another

• Then no matter what a player believes about other players,

one strategy leads to a higher payoff than the other.
 Dominant Strategies
• A strategy ai is a (weakly) dominant strategy for a player i

ui (ai, a-i ) ≥ ui (ai’, a-i ) for all ai’ and a-i

• A strategy ai is a strictly dominant strategy for a player i

ui (ai, a-i ) > ui (ai’, a-i ) for all ai’ and a-i
 Predator/Prey

Pred\Prey Active Passive

Active 1.7, -.8 3,-1
Passive 1.6, -.7 0,0

• Predators have a strictly dominant strategy here

 The Classic Game where both players
have Strictly Dominant Strategies
• Prisoners’ Dilemma

Confess Silent
Confess -4,-4 0,-5
Silent -5,0 -1,-1
 The Classic Game with Strictly Dominant
Strategies
• Prisoners’ Dilemma

Confess Silent
Confess -4,-4 0,-5
Silent -5,0 -1,-1
 The Classic Game with Strictly Dominant
Strategies
• Prisoners’ Dilemma

Confess Silent
Confess -4,-4 0,-5
Silent -5,0 -1,-1
 The Classic Game with Strictly Dominant
Strategies
• Prisoners’ Dilemma

Confess Silent
Confess -4,-4 0,-5
Silent -5,0 -1,-1
A Game with Weakly Dominant
Strategies

Movie Home
Movie 1, 1 0,0
Home 0,0 0,0
 Quiz  

 
Consider  the  game:  

1  \  2   Movie   Home  
Movie   2,  1   1,  1  
Home   0,  1   0,  2  
 Quiz  
“Movie”  is  a  “    ”  dominant  strategy  for  player  1,  and  “Home”  is  a  “    ”  
dominant  strategy  for  player  2.  What  should  be  filled  in  the  blanks?  
a) Strictly;  Strictly.  
b) Strictly;  weakly  (not  strictly).  
c) Weakly  (not  strictly);  strictly.  
d) Weakly  (not  strictly);  weakly  (not  strictly).  
 Quiz:  Explana?on  
(b)  is  true.  
• “Movie”  is  a  strictly  dominant  strategy  for  player  1  because  
– When  2  plays  Movie,  1  gets  2  from  Movie  and  0  from  Home;  When  2  plays  
Home,  1  gets  1  from  Movie  and  0  from  Home.  (Movie  is  always  strictly  
beHer.)  
• “Home”  is  a  weakly  (not  strictly)  dominant  strategy  for  player  2  
because  
• When  1  plays  Movie,  2  gets  1  from  either  Movie  or  Home  (so  is  indifferent);  
When  1  plays  Home,  2  gets  1  from  Movie  and  2  from  Home.  (Home  is  strictly  
beHer.)  
 Normal Form
• An example of a normal form game

• The formal model

– The main ingredients: Players, Strategies, Payoffs
– Reasoning about games
• From ``Optimal Strategy’’ to ``Solution concepts’’
• Dominance
• Equilibrium
 Nash Equilibrium
• ``Equilibrium’’

• Strategies such that if played, nobody would want to change

• Stable even if know what will happen

• Don’t regret the strategy played

 Equilibrium
• Nash Equilibrium - Named after John Nash

• Also sometimes called Cournot-Nash equilibrium – Discussed

by Cournot (1838) in the context of a game of oligopoly:
competition between firms
 Equilibrium
• One reasoning:

• Non-equilibrium points should not last

• Predicting non-equilibrium points requires a prediction that

some player is not optimizing against the others’ behaviors

• Not a justification for equilibrium, but a strike against non-

equilibrium
 Nash Equilibrium
• Best Responses

• Pure Strategy Equilibrium

• Mixed Strategies

• Existence of Equilibrium
 Nash Equilibrium
• Best Responses

• Pure Strategy Equilibrium

• Mixed Strategies

• Existence of Equilibrium
 Best responses
• Game: (N, (Ai)i , (ui)i)

• ai is a best response to a-i (strategies of others) if

ui (ai, a-i ) ≥ ui (ai’, a-i ) for all ai’

• Always have a best response in a finite game

(every finite set of numbers has a max)
 Predator/Prey

Pred\Prey Active Passive

Active 1.7, -.8 3,-1
Passive 1.6, -.7 0,0

• Large Mammal estimates

 Quiz  

 
Consider  the  collec:ve-­‐ac:on  game:  

1  \  2   Revolt   Not  
     Revolt       1,  1   -­‐1,  0  
Not   0,  -­‐1   0,  0  
 Quiz  
 
When  player  1  plays  “Not”,  for  player  2  
 
a) “Revolt”  is  a  best  response.  
b) “Revolt”  is  a  dominant  strategy.  
c) “Not”  is  a  best  response.  
d) “Not”  is  a  dominant  strategy.  
 Quiz:  Explana:on  
(c)  is  true.  
• When  player  1  plays  “Not”,  player  2  gets  -­‐1  from  “Revolt”  and  0  
from  “Not”.  Thus  “Not”  is  a  best  response.  
• No  strategy  is  a  dominant  strategy:    
– When  the  other  player  plays  “Not”,  it  is  strictly  beEer  to  play  “Not”;  
– When  the  other  player  plays  “Revolt”,  it  is  strictly  beEer  to  play  “Revolt”;  
– No  strategy  always  dominates  the  other  strategy.  
 Nash Equilibrium
• Best Responses

• Pure Strategy Equilibrium

• Mixed Strategies

• Existence of Equilibrium
 Nash Equilibrium
• A profile of strategies (a1,...,an) is a (pure strategy) Nash
equilibrium if ai is a best response to a-i for all i

• So: ui (ai, a-i ) ≥ ui (ai’, a-i ) for all i and ai’

 Nash Equilibrium versus Dominant
Strategies
• ai is a best response for i:
ui (ai, a-i ) ≥ ui (ai’, a-i ) for all ai’

• ai is a (weakly) dominant strategy for i:

ui (ai, a-i ) ≥ ui (ai’, a-i ) for all ai’ and a-i
 Quiz  

Two  players  work  on  a  project  with  the  following  symmetric  

payoffs:  

1\2   Work  (W)   Shirk  (S)  

Work   4,  4   0,  2  
Shirk   2,  0   0,  0  
 Quiz  
List  all  (weakly)  dominant  strategies  and  all  pure  strategy  Nash  
equilibria:  
a) “Work”  is  dominant;  equilibrium  is  (W,  W).  
b) “Work”  is  dominant;  equilibria  are  (W,  W)  and  (S,  S).  
c) No  strategy  is  dominant;  equilibrium  is  (W,  W).  
d) No  strategy  is  dominant;  equilibria  are  (W,  W)  and  (S,  S).  
 Quiz:  Explana:on  
(b)  is  true.  
• “Work”  is  a  dominant  strategy:  
– When  the  other  player  plays  “Work”,  it  is  strictly  beEer  to  work  (4  vs  2  
from  shirk).  
– When  the  other  player  plays  “Shirk”,  a  player  is  indifferent  between  work  
and  shirk  (0  vs  0).  
• (W,  W)  and  (S,  S)  are  both  pure  strategy  Nash  equilibria:    
– Neither  player  would  be  strictly  beEer  by  deviaNng  from  either  of  these  
prescribed  pairs  of  acNons,  presuming  the  other  plays  the  prescribed  
ac4on:    if  the  other  Shirks,  then  a  player  is  indifferent  and  also  willing  to  
Shirk.  
 ``Battle of the Sexes’’ Game
• No strategy is (weakly) dominated

L P
L 3,1 0,0
P 0,0 1,3
 ``Battle of the Sexes’’ Game
• There are two pure strategy Nash equilibria:

L P
L 3,1 0,0
P 0,0 1,3
Nash Equilibrium in technology
adoption:

No dominant strategies
Two equilibria: Old, Old and New, New

Old New
Old 1,1 0,0
New 0,0 2,2
Nash Equilibrium in technology
adoption:

No dominant strategies
Two equilibria: Old, Old and New, New

Old New
Old 1,1 0,0
New 0,0 2,2
 A Simple Collective-Action Game

Revolt Not
Revolt 1, 1 -2,0
Not 0,-2 0,0

Two Nash equilibria (in pure strategies):

both revolt, both not revolt

Again there are no dominated strategies

 A Simple Collective-Action Game

Revolt Not
Revolt 1, 1 -2,0
Not 0,-2 0,0

See the importance of communication!

Not revolting is the ``safer strategy’’

 But can involve weakly dominated
strategies:
Movie Home
Movie 1, 1 0,0
Home 0,0 0,0

Home, Home is a Nash equilibrium...

 Quiz  

 
Consider  a  seller-­‐buyer  game:  

S\B   Buy   Not  

Good   1,  1   0,  0  
Bad   2,  -­‐1   0,  0  
 Quiz  
Find  all  pure  strategy  Nash  equilibrium:  
a) Only  (Good,  Buy).  
b) Only  (Bad,  Not).  
c) Only  (Good,  Not).  
d) Only  (Bad,  Buy).  
e) Both  (a)  and  (b).  
 Quiz:  ExplanaBon  
(b)  is  true.  
• (Bad,  Not)  is  a  pure  strategy  Nash  equilibrium:  
– When  the  buyer  chooses  Not,  seller  gets  0  from  either  Good  or  Bad,  so  has  
no  incenBve  to  deviate  from  Bad.  
– When  the  seller  chooses  Bad,  the  buyer  gets  0  from  Not  and  -­‐1  from  Buy,  so  
has  no  incenBve  to  deviate  from  Not.  
• (Good,  Buy)  is  not  a  pure  strategy  Nash  equilibrium:    
– When  buyer  chooses  Buy,  the  seller  gets  1  from  Good  and  2  from  Bad.  So  the  
seller  would  deviate  by  deviaBng  to  Bad.  
• Similarly,  you  can  check  that  from  each  of  the  other  combinaBons  of  pure  
strategies  some  player  would  strictly  benefit  from  deviaBng.  
 Justifications for Equilibrium
• Self-fulfilling

• No regrets

• Any non-equilibrium point is not stable in the sense that

somebody could do better by deviating

• Is equilibrium justified by some process? How do we get

there?