Game Theory (5en254)
Game Theory (5en254)
Game Theory (5en254)
ZS 2020/2021
Klára Žemlová
1
Lecture 1
Example 1.1
• Same position
• Do not know each other, will never meet again
• Enough time to think
• Held in separate cells
• Want to minimize the years in Gulag
• Not important whether guilty or not
• Get explained consequences (punisments)
Game
• Game = strategic situation
o Payoffs mutually depend
o Cooperative/non-cooperative
• Types:
o Static games
▪ Simulatanous/order does not matter
o Repeated static games
▪ Same situation again and again
o Dynamic games
▪ Structure/order of decisions matter
• Elements of games
o Players
▪ Decision makers
▪ Ex. People, groups, firms, governments, genes…
o Actions
▪ What players can do
▪ Action set = collection of all actions of a player
▪ Action profiles
• Situations that may happen
• List with one action/decision per each player
▪ Outcomes = anything of value for players
• Ex. Money, prestige, food, corporate profit
o Prefecences & payoff-functions:
▪ Over outcomes associated with action profiles
▪ Only ordinal meaning for now
• Definition: Normal-form Game
o I players, for each i: Actions, preferences/utility function
o Normal-form = table (matrix), typically:
▪ Row-player
▪ Column-player
▪ Actions
▪ Cells with utility for players
o Conductor, Tchaykovskij
• Actions:
o A=(Confess – C, Not confess – NC)
o Action profiles: {(C,C),(NC,NC), (C,NC),(NC,C)}
o Outcomes: years in prison
• Preference relations:
o (Over outcomes of action profiles)
o Conductor: {C,NC} > {C,C} > {NC, NC} > {NC,C}
o Tchaykovskij: {NC, C} > {NC, NC} > {C,C} > {C, CN}
Additionally…
• Information structure
o Who knows what
o For now, everybody knows everything
• Assumptions on behavior
o Make life easier
o Proceed from simple to complicated
• No cheating
• No cooperation
Assumptions
• Players care only about own utility/payoff
o Rational selfishness
o Not harming others
• Rationality of players
o Irrelevance of unimportant information
o No emotins
o Cognitive cooperation, memory, iteration…
• Common knowledge of rationality
o P1 is rational.
o P2 is rational.
o P1 knows that P2 is rational.
o P2 knows that P1 is rational.
o P1 knows that P2 knows that P1 is rational.
o …
3
Models
• Model = simplification of reality that represents certain features of reality
o Map vs globe vs the Earth
• Assumptions
o Simplity reality, while serving a purpose
o Capture the essence, not irrelevant details => good enough approximation
o Shorter than verbal description
o Ass. Of basic microeconomics : homo-economicus (often unrealistic, non-existent
later on)
Economic models
• Why models?
o Intuition unreliable
▪ Analyze complex situations in a tractable way
▪ Help to undestand the realitionships, recognize patterns
▪ Provide non-trivial insights and common language
o Help with predictions
▪ Practical implications
o Give a bechmark toward real behavior
• Real-life: homo sapiens sapiens
o (sub)conscious psychological drives
o Cognitive limitations
o Not well-defined goals
o Lack of ability to achieve them, not maximizing utility..
o When is this important?
Lecture 2
How will the game likely end?
Method 2: leave out “bad” decisions
• Rational players would not play them
1. Compare actions – find out which are bad
2. Cross out “bad” ones
3. Repeat in a reduced game
Bad decisions
• “Bad” action x = if another action y is “always” better
o Y brings higher utility than x no matter what other players do
o Action x is strictly dominated by y
▪ Dominated = bad
▪ Dominating = good
Assumptions
• We assume rational players will not play dominated strategies
o Rationally
• Other players know this
o Assume: common knowledge of rationality
• Rephrase: There is no belief that a player can hold such that it would be optimal to play
dominated action
o Cannot rationalize the action
▪ Rationalize = defend that based od my beliefs, this was the best I could do
• => we can leave them out (the bad decisions)
Party game
Angelina and Brad can come to a party Early or late
• If both come early, both enjouy each other’s company = 10 h
• If both come late, both enjoy each other’s company = 5 h
• If one comes early and has none to speak to, gets drunk and does not have fun (0 h) while
the other later makes fun of him/her, but this is less fun than other’s company (3 h)
Any dominated strategies?
Dominant equilibrium
Definition:
• Action profile a* if for every player her action is strictly dominating all other actions, given
other players also play stricly dominating strategies
Must be unique if strict domination is used.
Example 3
Any strickly bad (dominated) actions? For P1? P2?
Equilibrium?
Lecture 3
Practice homework
Solve the matrix using IEDS
Dominant equilibrium: {D,M}
o A “steady” state
o A plausible final situation of game with rational agents
o Holding decisions of others constant, can one single player improve? NO
equilibrium
o No regret principle
▪ If having chance to come back and repeat my decision as the onli one from
all players, a player does not want to
o Self-enforcing contract
▪ If you know others will make their equilibrium action, you want to also make
your eq. Action
o Players have “coordinated” beliefs
• Solution of a game: a theory/way how to find an equilibrium
Level-K thinking
Level O (L0): non-strategic thinking
• Think of number between 0 and 100
• Expected average? 50
• Eliminate numbers above 50
Level 1 (L1): think all others are L0 and responds
• Think the average guess is 50
• Writes down 33
Level 2 (L2): thinks all others are L1 and responds to
• Writes down 22..
Full rationality?
Perfect rationality vs common knowledge of rationality
Weak dominance
If action z almost always better, but sometimes yields the same payoff as x
• = action x i weakly dominated by z
• Z and x could be the same
Formally (definition):
• Player i’s action a weakly dominates action b if ui(a, a-i) >/= u-i (b, a-i) for every list a-i of other
players’ actions
• And for some actions, ui(a, a-i) > u-i (b, a-i)
Lecture 4
Best response
Core idea: if one player knows what the others are going to do, her decision is easy
Example 3.6
Method 3: Best response intersection
BR1(L) = T, M
BR1(C) = M
BR1(R) = T
BR2(T) = L
BR2(M) = L
Example 3.7
Can you solve the matrix using IEDS?
Write down the BR of both playes and
indicate them in matrix.
BR1(A) = Y BR2(X) = B, C
BR1(B) = Y BR2(Y) = B
BR1(C) = Z BR2(Z) = A
Do the BR cross? Yes, crossing by playing
actions {Y, B}.
Is X dominated? No.
• Never a best response =/=> strictly
dominated
• Strictly dominated => never a best response
10
Example 4.1
1. Find NE
2. Find DE using weak elimination
• If we eliminate T which is weakly
dominated by B, and then eliminate
L which is dominated by R, we lose
nonstrict NE {TL}
Example 3.8
Underline the BR of both players
Do the BR cross? Yes, the cross by playing
action {DM}
Example 3.9
Underline the BR of both players
Do the BR cross? Yes, they cross by playing actions {(XA),(ZC)}
Can you solve the matrix using IEDS – strict or weak? No
Lecture 5
Example 5.1 – hand feeding squirrel
Situation: matching pennies
What is the bes decision here?
No Nash Equilibria in deterministic decisions
• No pair of actions is complatible with a steady
state
1) If squirrel goes always left:
13
Probability distribution
Giving probabilities to outcomes
Example: tossing a fair coin
• Outcomes: {HT}
• Probabilities: (1/2, ½)
Example: playing fair die
• Outcomes: {1, 2, 3, 4, 5, 6}
• Probabilities (1/6, 1/6, 1/6, 1/6, 1/6, 1/6)
Example: playing two fair dice, focus on sum
• Outcomes: {2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12}
• Probabilities: (1/36, 1/18, 1/12, 1/9, 5/36, 1/6, 5/36, 1/9, 1/12, 1/18, 1/36)
Expected payoffs
EP = sum (probabilityi * payoffi)
Example: tossing a fair coin
• Payoffs: H => +1, T => -1
• Probabilities: (1/2, ½)
• EP = ½*1 + ½*(-1) = 0
Example: playing fair die
• Payoffs: {1, 2, 3, 4, 5, 6}
• Probabilities: 1/6, 1/6, 1/6, 1/6, 1/6, 1/6
• EP = 1/6*1 + 1/6*2 + … + 1/6*6 = 3,5
Example: playing two fair dice, focus on sum
• Payoffs: {2, 3, 4, 4, 6, 7, 8, 9, 10, 11, 12}
• Probabilities: (1/36, 1/18, 1/12, 1/9, 5/36, 1/6, 5/36, 1/9, 1/12, 1/18, 1/36)
• EP = 1/36*2 + 1/18*3 + … + 1/36*12 = 7
• E.g. si = ( ½, ¼, ¼) is mixed strategy where player i plays “left” with probability ½ and “middle”
and “right” with probability ¼
• Probabilities have to sum up to 1!
• Randomization over actions
Mixed strategy may assign probability 1 to a single action – pure strategy
• E.g. si = (0, 1) is pure strategy where player i always plays “R”
Support of mixed strategy – set of actions played with positive probability.
Graph
Graph – continue
15
Payoffs
Formally:
• Expected utility of the mixed strategy profile for player y:
o Sum (over all cells in the matrix) of utility of the cell weighted by the probability of
reaching the cell undet given mixed strategy profile
o Probability of reaching the cell is given by the product of probabilities of each player
playing his action that leads him to this cell
Graph – solution of my BR
Computing MSNE
Example 5.3 – Battle of sexes again
• Every Thursday after work they meet
• Even though they may have decided upon where to go, Brad often forgets the decision
• So Angelina is uncertain about what Brad will do: will he assume they go shopping or that
they go to a boxing game?
o (Or, Brad is fed-up with always going boxing and would like to have a fair
relationship)
• Brad can not communicate with Angelina
o This would show her that he forgot
• What should they do?
o They should somehow best-respond to that uncertainty
Best responses:
• We found MSNE point {(2/3,1/3), (1/3,2/3)}
o Or {1/3, 2/3} in this plane
• What if Angelina knows brad goes to Box less often than p?
19
Best response
• Intersection of BR
o MSNE
• Check it is really NE
o By deviating, can anybody strictly
gain?
▪ Any pure strategies better?
• 3 intersection = 3 equilibria
o Two pure strategy
o One mixed strategy
Computing MSNE
1. Guess the support
o Support – the actions played with positive probabilities
2. Assign “some” probabilities to actions
o Assume players are mixing and reacting to others
3. Calculate the expected payoffs based on probabilities
o Assume indifference
4. Construct the best-response functions
20
Revision 1
Venus vs Serena
• Serena is at the net
• Venus has the ball
• Where will Venus send the ball?
o To Serena’s left or right?
o Serena has to guess
• Pure-strategy NE? What are BR?
she do better?
o If I play P, what is my EP?
o If I play S, what is my EP?
o If I play (1/3, 1/3, 1/3), what is my EP?
• MSNE = {(1/3, 1/3, 1/3); (1/3, 1/3, 1/3)}
Conclusion
To be able to compare expected values we need different type of preferences – cardinal
• Von Neumann – Morgenstern preferences
Re-definition of Best-Response, NE, game
• Mixed strategies
Hint: dominated strategy never used in mixing
Interpretation of mixing:
• Reaction to uncertainty
• Uncoordinated repeated play
• Confusing your opponent
Sequential games
Famous example: prisoner’s dilemma
• What if player 2 gets info about move of player 1?
• Time dimension: sequential game
• Representation
o Extensive form = game “tree”
o Capture the order of decisions
o P1 moves, P2 observes decision and moves…
22
Game tree
Pure strategies:
S1 = {(A,G), (A,H), (B,G), (B,H)}
S2 = {(C,E), (C,F), (D,E), (D,F)}
Lecture 7
Subgame perfection
Subgame
• Remaining portion of game three after some moves
• Restriction of the original game
• Can be a game on its own
• Full game is its own subgame
“Subgame-perfect” Nash Equiibrium
• In no subgame can any player i do better by choosing a strategy different from s*i, given that
every other player j plays s*j
25
Subgame
Backward induction
26
Example 7.4
Find SPNE. => SPNE = {(L,R,R), (U,D); (L, R, R),(D,D)}
27
Example 7.7
Two players
• Incumbent monopolist
• A potentional entrant
Entrant chooses between entering the market or not.
• If the entrant enters, the monopolist profit will be lower.
• The monopolist may try to deter enter by threatening to fight
o Engage in an expensive advertising war, predatory pricing, preventing the entrant
from making positive profit.
29
• Such fighting will, however, be costly for the monopolist as well as the entrant.
Two NE:
{Enter (not fight, not fight)} – also SPNE
{not enter (fight, not fight)}
Conclusion
• Extensive representation of games
o Extension of definition of games
o New definition of strategy
• Problem of NE: incredible threat
• Definition – subgame
• Subgame-perfect Nash Equilibrium
Lecture 8
Recap from previous topic
• Sequential games
o Game tree
• New definition of strategy (!)
• Transformation of game from seq to normal form
• Subgame
• Subgame-perfect Nash Equilibrium
• (In-)credible threat
• Backward induction
• Interpretation of NE (B,(B,S)):
o Given Brad’s beliefs and actions of both types of Angelina, Brad is playing the best
response
o Given Brad’s action, both types types of Angelina are playing best response
• Interpretation of NE if Angelina is the 1st type:
o Angelina wants to meet Brad and chooses B
o Brad chooses B and believes that 1st type Angelina chooses B and the 2nd type
Angelina chooses S
• Interpretation of NE if Angelina is the 2nd type:
o Angelina wants to avoid brad and chooses S
o Brad chooses B and believes that 1st type Angelina chooses B and the 2nd type
Angelina chooses S
Definition of a Bayesian game – Battle of sexes
• Players
o Brad and Angelina
• States
o Meet, avoid
• Actions
o B and S
• Signals and signal function
o B: one signal both states
o A: two signals
• Beliefs
o Probabilities of states after observing signals
o B: ½ to each state after receiving signal
o A: 1 to either state after signal
o Associated expected action of other player
• Payoffs
o Given in the table
Solution:
• One way – create a table, three players, p = 0.5…
• Shortcut:
o if you are good suspect, strictly dominant strategy NOT shoot
o if you are bad, dominant strategy is to shoot
▪ you can leave the dominated strategies out
• Expected utility for sheriff:
o E(shoot) = -1*(1-p) + 0*p = 0 – 2*p => p = 1/3
▪ If p > 1/3, sheriff shoots
▪ If less, does not shoot
▪ If equal, then mixing
Example 2 – too much information may hurt
• Single-person decision problem
o Player cannot be worse off if she has more information: if she wishes, she can ignore
the information
• Strategic game
o If a player has more information and the other players know that she has it, she may
33
be worse off
• Following game has two possible states, both players believe that states S1 and S2 are
equally likely
34
Conclusion
Bayesian games – information is incomplete
• Same players, actions, different payoffs
How to find NE in static Bayesian games:
• Consider each type as a single player
• Given beliefs, compute expected payoffs
• Find NE in this game
Lecture 9
Revision
Incomplete information games
• Min one player does not know full game
• Uncertain about part of game, e.g. Payoff/type of other player
• Know probability of events happening
o Chance of meeting good/bad Angelina
Solution procedure
• Treat two types of one player as two players
• Create all combinations of beliefs about their actions
• Create a “big” matrix by calculating expected payoff of uncertain player and the two types
Equilibrium
• Actions optimal given beliefs and actions of other players; beliefs are correct
35
Prisoner’s dilemma:
Game tree – general
• For P2, action C different from E
• Can choose differently in each node
• Strategy set = {CE, CF, DE, DF}
Intuition 2
How to analyze this game?
ANS: no domination for P2 this time.
But! Domination for P1
• (P1: if P2 Y -> A)
P2 “knows” this -> X
• Believes
Could be transformed to a normal form game
• NE: (A,X)
More complicated…
Example 1
How to analyze this game?
• Without R easy:
o T dominates B
o Thus P1 plays L
o Holds also with R
Subgames?
Finding Nash
• Let’s transform it to normal-form game
• Find NE
o Two NE: (L,T) and (R,B)
• Are these subgame perfect as well?
o Only one subgame, whole game (subgames can not cut
through information sets) => yes, they are SPNE as well
Main point 1
• What about (R,B)? Is it plausible?
o (R,B) is empty/incredible threat
▪ P2 will never play B
o We want to eliminate this, keep only (L,T)
• How to eliminate empty threat?
o Dynamic games with perfect information
▪ SPNE eliminates non-credible threats
o Dynamic games with imperfect information
▪ SPNE is not strong enough!
37
Practice example
Consider following 3-player game
• How many proper subgames it has? 3
• Find wPBNE.
• What about (A, L, L’, p=0)
38
Example 3
Pepsi is considering a new marketing war against Coca Cola. Pepsi can decide to stay out of the
market (O). If Pepsi decides to enter the market Coca Cola knows that there is a new increased
competition but does not observe whether Pepsi is ready (R) or unready (U). Coca Cola decides to
accept (A) the new competitor or to fight (F).
1. Write down normal form of the game.
2. Find Nash equilibria (optimal actions)
3. Check NE one by one to see if consistent beliefs can be found (if yes, we have wPBNE; if no,
no wPBNE)
This works because wPBNE is a subset of NE. There are two NE: (O,F) and (U,A)
1. (U,A) is NE. Is it wPBE?
o Option 1: information set is reached
o Beliefs have to be consistent with actions => p=0
and 1-p=1
o Yes, wPBE 1 is (U,A) and p=0, 1-p=1
▪ P1 behaves optimally given action of P2
▪ P2 behaves optimally given his beliefs
and P1’s action
▪ P2’s beliefs are consistent with P1’s
actions
2. (O,F) is NE. Is it wPBNE?
o Option 2: information set is not reached
o Beliefs have to be consistent with actions =>
beliefs have nothing to be consistent with set p
such that actions of all players are still optimal
o => F has to be optimal for Coca-Cola EP(F) >/=
EP(A)
o WPBE 2 is (O,F) and any p > ½ (F is better than A)
▪ P1 behaves optimally given action of P2
▪ P2 behaves optimally given his beliefs and
P1’s action
▪ P2’s beliefs are consistent with P1’s action
39
Summary
Static games
• Complete information: NE
• Incoplete information: Bayesian NE
Dynamic games
• Perfect information: SPNE
• Imperfect information: weak perfect Bayesian equilibrium
Weak perfect Bayesian equilibrium
• Sequential rationality
• Consistency of beliefs
• Elimenates on-credible threats
3. P1 moves
o Obtain education or no
Lecture 10
Revision: weak perfect Bayesian Nash Equilibrium
Definition: A wPBE equilibrium consists of equilibrium strategies and beliefs systems satisfying:
1. Sequential rationality
o Each players’ strategy is optimal whenever she has to move, given her belief and the
other players’ strategies
▪ Similar idea to subgame perfection
▪ Rules out incredible threat in an imperfect-information game
2. Consistency of beliefs
o With strategies – each player’s belief is consistent with strategy profile (strategies of
all players)
▪ With players believe others are doing is true
▪ Required if information set with uncertainty is reached
Signaling games
• Extensively used in Economics
o Also in biology – credibility enhancing displays
o Peakock’s tail
• Michael Spence – Nobel prize 2001
• Job-market signaling
• Sender of a signal/message
o Worker can signal productivity – get MBA
• Receiverer of a signal
o Employer observes signal and decides to hire or not
• Spence shows it may make sense to have MBA even though it may be useless
o Costs less high-ability employees than low-ones
o Signal can separete them
• Rephrase MBA and abilities to beer and quiche
41
Signaling games
Beer or quiche
• 0. Nature chooses the type of P1:
o Weak, strong with certain probabilities
• 1. PA chooses his breakfast: Beer of Quiche
• 2. P2 (a bully) does not know the type of P1 but
observes breakfast.
o Then he decedes whether to fight P1.
• P1 gets 1 point if he had his favorite meal, and gets
additional 2 points if there was no fight.
• P2 gets 1 point if he fought the Weak type did not
fight the Strong
Equilibria:
• Separating equilibrium
o Each type (P1) chooses a different action
o Receiver (P2) knows the sender’s type
o E.g.: strong -> Beer; weak -> quiche
• Pooling equilibrium
o All types (P1) choose the same action
o Receiver (P2) has no clue about sender’s type
o E.g.: strong -> beer; weak -> beer
• Partially separating/pooling equilibrium
o Some types of sender send one message; some types of sender mix the messages
o E.g.: strong -> beer; weak -> mix
44
Pooling NE 1 Pooling NE 2
Separated NE 1 Separated NE 2
WBNE: P1S, P1W -> B, P2(Q) -> F, P2(B) -> NF, P > 1/2, q = 0.1
WBNE: P1S, P1W - > Q, P2(Q) -> NF, P2(B) -> F, q > 1/2 , p=0.1
Lecture 11
Revision
Static games
• Perfect information: nash equilibrium
o Conductor and Tchaykovskij
• Mixed strategy NE
o Squirrel and me
• Imperfect information: (Bayesian) Ne
o Angry Angelina or a nice Angelina?
46
Dynamic games
• Perfect infromation: SPNE
o Billa, Tesco and Albert entering market
• Imperfect information: weak perfect Bayesian equilibrium
o Bully and Kung-fu master / weak guy
Beer of quiche
Analyze all possible equilibria:
• Separating equilibrium 1: weak-quiche, strong-beer
• Separating equilibrium 2: weak-beer, strong-quiche
• Pooling equilibrium 1: weak-beer, strong-beer
• Pooling equilibrium 2: weak-quiche, strong-quiche
In each case:
1. Start with player 1 actions
2. Determinate player 2’s beliefs
3. Find player 2’s optimal response
4. Check if player 1’s action is optimal
5. Pooling equilibrium: find beliefs s.t. P1’s action is optimal
Pooling NE 1 Pooling NE 2
Separated NE 1 Separated NE 2
Pooling NE 1 - possible
Pooling NE 2 – not OK
Pooling equilibrium:
P1W,S -> Q
P2(Q) - F
P2(B) – F
Q=0.3, p <13/20