Games &Computers.ppt 2564KB Jun 23 2011 12:12:22 PM
Games Computers
(and Computer Scientists)
Play
Avi Wigderson
Computer
Science
Games
Game
Theory
=
Information Processing by
Computers
Agents
•
•
•
•
•
•
Competing
Cooperating
Faulty
Colluding
Secretive
Adversarial
Computationally Bounded
Communicating Digitally
Plan
• Complexity of Games
• Implementation of Games
• Design of Games
• Games against Clairvoyance
Complexity of Games
Theorem [Zermelo] : In every finite
win/lose perfect information 2player game, White or Black can
force a win.
Extensive Form
Question: Can a winning strategy be
efficiently computed?
Rectangle Game
m=4
n=5
1
m
1
5
2
4
3
n
.Theorem: White has a winning strategy
. Proof: Assume Black has a winning strategy
!Then White can mimic it and win. Contradiction
Question: What is the winning strategy?
Zero-Sum Games
Matching Pennies
(simultaneous play)
H
T
H1
-11 1-
T1
1-1 -1
Strategic Form
“Best” strategy for each player is to flip a fair coin. Game value is 0.
m
1
2
j
: ]Theorem [von Neumann ‘28
1
Every 0-sum game has a
2
.Min-Max) value(
v -v
Question: Can the value, i
strategies be computed? n
: ]Theorem [Khachian ‘80
.Yes – Efficient linear programming algorithm
ij
ij
Nash Equilibrium
Chicken [Aumann]
C
C
D
11 02
Strategic Form
Probabilistic strategies (Sw, Sb).
D 2 0 -3 -3
Nash Equilibrium: No player has an incentive to
.change its strategy given the opponent’s strategy
¼[here Sw=Sb = [C with prob ¾, D with prob
.Theorem [Nash]: Every (matrix) game has an equilibrium
Question: Can the players compute (any) equilibrium?
).Best known algorithm: exponential time (infeasible
Implementing Games
The Millionaires’ Problem
A
Alice
B
Bob
Both want to know who is richer
Neither gets any other information
Question: Is that possible?
Joint random decisions
3/4 1/4
Nash eq. With Independent Strategies
C
Expected value = 3/4
Prob[CC[ = 9/16
Prob[CD[ = 3/16
Prob[DC[ = 3/16
Prob[DD[ = 1/16
3/4 C
D
11 02
1/4 D 2 0 -3 -3
Nash eq. With Correlated Strategies [Aumann]
Prob[CD[ = 1/2
Prob[DC[ = 1/2
Prob[CC[ = 0
Prob[DD[ = 0
Expected value = 1
Question: How to flip a coin jointly?
Simultaneity
1/2 1/2
H
T
Expected value = 0
)if they play simultaneously( 1/2
H1
-11 1-
1/2
T1
1-1 -1
Question: How do we guarantee simultaneity
A computational representation:
outcome
Parity Function
xW xB Parity(xW, xB )
P
xW
xB
0
1
0
1
0
1
1
0
0
0
1
1
Privacy vs. Resilience
• Voting
Majority Function
M
x1 x2 x3
x1
x2
x3 Majority(x1, x2, x3 )
0
0
0
1
0
1
1
1
0
0
1
0
1
0
1
1
0
1
0
0
1
1
0
1
0
0
0
0
1
1
1
1
Q1: How to guarantee x15?
Q2: How to guarantee x1 remains private?
• Millionaire’s Problem
• Poker
• Any game
Completeness Theorem
Theorem [Yao, Goldreich –Micali –Wigderson[:
1. More than 1/2 of the players are honest
2. Players computationally bounded
3. Trap-door functions exist (e.g. factoring integers is hard)
Every game,
with any secrecy requirements,
can be digitally implemented
s.t. no collusion of the bad players can affect:
* correctness (rules, outcome)
* privacy (no information leaks)
Hard problems can be useful!
Correct & Private digital implementation
Trusted party
Ideal implementation
Secrets
Preferences
Strategies
s1
s2
sn
1
2
n
Internet
Digital implementation
Internet
How to ensure Privacy
Oblivious Computation [Yao[
1
f(inputs)
P
1
1
0
M
P
1
1
0
P
M
P
1 0 0
1 0
1 0
How to ensure Correctness
Definition [Goldwasser-Micali-Rackof[:
zero-knowledge proofs:
• Convincing
• Reveal no information
Theorem [Goldreich-Micali-Wigderson[:
Every provable mathematical statement has a
zero-knowledge proof.
Corollary: Players can be forced to act legally,
without fear of compromising secrets.
Where is Waldo? [Naor[
Designing Games
How to minimze players’ influence
Public Information Model [Ben-Or—Linial] :
Function
Joint random coin flipping
Parity
Every good player flips, then combine Majority
majorit
parit
y y
M
M
M
P
M
Influenc
1
1/7
Iterated
Majority
1/8
M
Theorem [Kahn—Kalai—Linial] : For every function, some
player has non-proportional influence.
Theorem [Alon—Naor] : There are “multi-round” function
for which no player has non-proportional influence.
How to achieve cooperation, efficiency, truthfulness
Players (agents) are selfish
• Auction
Question: How to get players to bid their true
values?
Theorem [Clarke—Groves—Vickery[:
2nd price auction achieves truthfulness.
• Internet Games
Question: How to get players to cooperate?
[Nisan[: Distributed algorithmic mechanism design.
[Papadimitriou[: Algorithms, Games & the Internet
New CS Issues: Pricing, incentives
New GT Issues: Complexity, Algorithms
Coping with Uncertainty
Competing against
Clairvoyance
On-line Problems
Investor’s Problem (One-way trading)
price
day
1
Muggle’s
action
Wizard’s
action
2
3
4
5
6
7
8
9
Profit/loss
On-line problems are everywhere:
• Computer operating systems
• Taxi dispatchers
• Investors’ decisions
• Battle decisions
•
•
•
Competitive Analysis [Tarjan—Slator[:
For every sequence of events,
Bound the competitive ratio:
muggle-cost(sequence)
wizard-cost(sequence)
Can be achieved in many settings.
Huge, successful theory.
“Online Computation and Competitive Analysis”
[Borodin—El-Yaniv[
Every Game? Any secrecy requirements?
Incomplete information
Game in Extensive form
Nature
Alice
Bob
...
...
...
...
Nature
Alice
...
Information Sets
• Player’s action depends
only on its information set
Completeness Theorems
Theorem [Yao, Goldreich –Micali –Wigderson[:
1. More than 1/2 are honest
2. Players computationally bounded
3. Trap-door functions exist (e.g. factoring integers is hard)
Every game, with any secrecy requirements, can
be
digitally implemented s.t. no collusion of the bad
players can affect:
* correctness (rules, outcome)
* privacy (no information leaks)
Theorem [Ben-Or –Goldwasser –Wigderson[:
1’.
2’. At least 3 players, more than 2/3 are honest
3’. Private pairwise communication
(and Computer Scientists)
Play
Avi Wigderson
Computer
Science
Games
Game
Theory
=
Information Processing by
Computers
Agents
•
•
•
•
•
•
Competing
Cooperating
Faulty
Colluding
Secretive
Adversarial
Computationally Bounded
Communicating Digitally
Plan
• Complexity of Games
• Implementation of Games
• Design of Games
• Games against Clairvoyance
Complexity of Games
Theorem [Zermelo] : In every finite
win/lose perfect information 2player game, White or Black can
force a win.
Extensive Form
Question: Can a winning strategy be
efficiently computed?
Rectangle Game
m=4
n=5
1
m
1
5
2
4
3
n
.Theorem: White has a winning strategy
. Proof: Assume Black has a winning strategy
!Then White can mimic it and win. Contradiction
Question: What is the winning strategy?
Zero-Sum Games
Matching Pennies
(simultaneous play)
H
T
H1
-11 1-
T1
1-1 -1
Strategic Form
“Best” strategy for each player is to flip a fair coin. Game value is 0.
m
1
2
j
: ]Theorem [von Neumann ‘28
1
Every 0-sum game has a
2
.Min-Max) value(
v -v
Question: Can the value, i
strategies be computed? n
: ]Theorem [Khachian ‘80
.Yes – Efficient linear programming algorithm
ij
ij
Nash Equilibrium
Chicken [Aumann]
C
C
D
11 02
Strategic Form
Probabilistic strategies (Sw, Sb).
D 2 0 -3 -3
Nash Equilibrium: No player has an incentive to
.change its strategy given the opponent’s strategy
¼[here Sw=Sb = [C with prob ¾, D with prob
.Theorem [Nash]: Every (matrix) game has an equilibrium
Question: Can the players compute (any) equilibrium?
).Best known algorithm: exponential time (infeasible
Implementing Games
The Millionaires’ Problem
A
Alice
B
Bob
Both want to know who is richer
Neither gets any other information
Question: Is that possible?
Joint random decisions
3/4 1/4
Nash eq. With Independent Strategies
C
Expected value = 3/4
Prob[CC[ = 9/16
Prob[CD[ = 3/16
Prob[DC[ = 3/16
Prob[DD[ = 1/16
3/4 C
D
11 02
1/4 D 2 0 -3 -3
Nash eq. With Correlated Strategies [Aumann]
Prob[CD[ = 1/2
Prob[DC[ = 1/2
Prob[CC[ = 0
Prob[DD[ = 0
Expected value = 1
Question: How to flip a coin jointly?
Simultaneity
1/2 1/2
H
T
Expected value = 0
)if they play simultaneously( 1/2
H1
-11 1-
1/2
T1
1-1 -1
Question: How do we guarantee simultaneity
A computational representation:
outcome
Parity Function
xW xB Parity(xW, xB )
P
xW
xB
0
1
0
1
0
1
1
0
0
0
1
1
Privacy vs. Resilience
• Voting
Majority Function
M
x1 x2 x3
x1
x2
x3 Majority(x1, x2, x3 )
0
0
0
1
0
1
1
1
0
0
1
0
1
0
1
1
0
1
0
0
1
1
0
1
0
0
0
0
1
1
1
1
Q1: How to guarantee x15?
Q2: How to guarantee x1 remains private?
• Millionaire’s Problem
• Poker
• Any game
Completeness Theorem
Theorem [Yao, Goldreich –Micali –Wigderson[:
1. More than 1/2 of the players are honest
2. Players computationally bounded
3. Trap-door functions exist (e.g. factoring integers is hard)
Every game,
with any secrecy requirements,
can be digitally implemented
s.t. no collusion of the bad players can affect:
* correctness (rules, outcome)
* privacy (no information leaks)
Hard problems can be useful!
Correct & Private digital implementation
Trusted party
Ideal implementation
Secrets
Preferences
Strategies
s1
s2
sn
1
2
n
Internet
Digital implementation
Internet
How to ensure Privacy
Oblivious Computation [Yao[
1
f(inputs)
P
1
1
0
M
P
1
1
0
P
M
P
1 0 0
1 0
1 0
How to ensure Correctness
Definition [Goldwasser-Micali-Rackof[:
zero-knowledge proofs:
• Convincing
• Reveal no information
Theorem [Goldreich-Micali-Wigderson[:
Every provable mathematical statement has a
zero-knowledge proof.
Corollary: Players can be forced to act legally,
without fear of compromising secrets.
Where is Waldo? [Naor[
Designing Games
How to minimze players’ influence
Public Information Model [Ben-Or—Linial] :
Function
Joint random coin flipping
Parity
Every good player flips, then combine Majority
majorit
parit
y y
M
M
M
P
M
Influenc
1
1/7
Iterated
Majority
1/8
M
Theorem [Kahn—Kalai—Linial] : For every function, some
player has non-proportional influence.
Theorem [Alon—Naor] : There are “multi-round” function
for which no player has non-proportional influence.
How to achieve cooperation, efficiency, truthfulness
Players (agents) are selfish
• Auction
Question: How to get players to bid their true
values?
Theorem [Clarke—Groves—Vickery[:
2nd price auction achieves truthfulness.
• Internet Games
Question: How to get players to cooperate?
[Nisan[: Distributed algorithmic mechanism design.
[Papadimitriou[: Algorithms, Games & the Internet
New CS Issues: Pricing, incentives
New GT Issues: Complexity, Algorithms
Coping with Uncertainty
Competing against
Clairvoyance
On-line Problems
Investor’s Problem (One-way trading)
price
day
1
Muggle’s
action
Wizard’s
action
2
3
4
5
6
7
8
9
Profit/loss
On-line problems are everywhere:
• Computer operating systems
• Taxi dispatchers
• Investors’ decisions
• Battle decisions
•
•
•
Competitive Analysis [Tarjan—Slator[:
For every sequence of events,
Bound the competitive ratio:
muggle-cost(sequence)
wizard-cost(sequence)
Can be achieved in many settings.
Huge, successful theory.
“Online Computation and Competitive Analysis”
[Borodin—El-Yaniv[
Every Game? Any secrecy requirements?
Incomplete information
Game in Extensive form
Nature
Alice
Bob
...
...
...
...
Nature
Alice
...
Information Sets
• Player’s action depends
only on its information set
Completeness Theorems
Theorem [Yao, Goldreich –Micali –Wigderson[:
1. More than 1/2 are honest
2. Players computationally bounded
3. Trap-door functions exist (e.g. factoring integers is hard)
Every game, with any secrecy requirements, can
be
digitally implemented s.t. no collusion of the bad
players can affect:
* correctness (rules, outcome)
* privacy (no information leaks)
Theorem [Ben-Or –Goldwasser –Wigderson[:
1’.
2’. At least 3 players, more than 2/3 are honest
3’. Private pairwise communication