CSIT5900

Game Theory and Auction - A Short Introduction

Fangzhen Lin

Department of Computer Science and Engineering

Hong Kong University of Science and Technology

Fangzhen Lin (HKUST) Game Theory 1 / 34

Multi-Agent Systems (MASs)

An MAS is one where more than one agent co-exist and interact.
In addition to issues that need to be dealt with in single agent case
(such as checking if a newly observed/told fact is consistent with his
current belief set), new issues in MASs include the following:
I How do agents communicate with each other.
I How do they cooperate with each other.
I How do they act in face of adversity.

Fangzhen Lin (HKUST) Game Theory 2 / 34

Examples

Two person adversary games such as chess, go, and tic-tac-toe:

I What’s my opponent like?
I How much time do I have for each move?
I Why am I playing the game?
Two teams compete against each other including bridge and most
sport games.
Resource allocations: a number of agents need to share certain
resources, what’s a fair way of dividing the resources?
Market mechanisms in a society.

Fangzhen Lin (HKUST) Game Theory 3 / 34

Game Theory

Game theory studies how self-interested agents interact.

It has been applied in economics, political science, social science,
biology, computer science, and many other areas.
Games in normal form are the simplest games studied in game theory
and also most fundamental ones.

Fangzhen Lin (HKUST) Game Theory 4 / 34

Game Theory (Definitions)
Wikipedia:
Game theory is a branch of applied mathematics that is used in the social sciences,
most notably in economics, as well as in biology, engineering, political science, in-
ternational relations, computer science, and philosophy. Game theory attempts to
mathematically capture behavior in strategic situations, in which an individual’s suc-
cess in making choices depends on the choices of others. While initially developed to
analyze competitions in which one individual does better at another’s expense (zero
sum games), it has been expanded to treat a wide class of interactions, which are
classified according to several criteria. Today, “game theory is a sort of umbrella
or ’unified field’ theory for the rational side of social science, where ’social’ is inter-
preted broadly, to include human as well as non-human players (computers, animals,
plants)” (Aumann 1987).
Stanford Encyclopedia of Philosophy:
Game theory is the study of the ways in which strategic interactions among rational
players produce outcomes with respect to the preferences (or utilities) of those players,
none of which might have been intended by any of them. The meaning of this
statement will not be clear to the non-expert until each of the italicized words and
phrases has been explained and featured in some examples.

Fangzhen Lin (HKUST) Game Theory 5 / 34

An Example

You and your friend share a network and both of you want to download a
movie:
If both of you do that, the network is jammed, and none of you is
happy: say both of you value it 2.
If only one of you do that, the network works perfectly: the one who
did that is very happy (score of 5), the other is very unhappy (0).
If none of you do that, then none of you is very happy, but then you
can do something together. So let’s assign it a score of 3.
What should you do? Does it depend on what you think your friend will
do? Does it depend on if you two can talk about it? Will most rational
people choose to do the same thing? What if the same situations need to
be repeated a number of times?
These are some of the problems studied in game theory.

Fangzhen Lin (HKUST) Game Theory 6 / 34

Modeling Agents’ Interests

Preferences: prefer A over B.

Utility functions: A has value 100 and B 60 to me.
A utility function is a function from outcomes (situations) to real numbers between [0, 1].

Theorem (von Neumann and Morgenstern, 1944) If a preference relation  satisfies the
axioms completeness, transitivity, substitutability, decomposability, monotonicity, and continuity,
then there exists a function u: O → [0, 1] with the properties:
1 u(o1 ) ≥ u(o2 ) iff o1  o2 , and
u([p1 : o1 , ...pk : ok ]) = ki=1 pi u(oi ), where [p1 : o1 , ..., pk : ok ] is a lottery (assigning
P
2
probability pi to outcome oi ) on O.

Fangzhen Lin (HKUST) Game Theory 7 / 34

Games in Normal Form

A (finite, n-person) game is a tuple (N, A, u), where

N is a set of players, indexed by i;
A = (A1 , ..., An ), and each Ai is a set of actions (pure strategies) for
agent i;
u = (u1 , ..., un ), and each ui is a utility function for agent i, which is
a function from A to numbers.

Fangzhen Lin (HKUST) Game Theory 8 / 34

Prisoner’s Dilemma

Two prisoners are interrogated separately.

If both denies (cooperate), then each gets 1 year;
If both admit (defect), then each serves 2 years;
If one admits and the other denies, then the one who denies gets 3
years and the one who admits is free.

C D
C 3, 3 0,5
D 5, 0 1, 1

Fangzhen Lin (HKUST) Game Theory 9 / 34

Coordination Games

We can choose to drive either on the left or the right side of the road.
Safe if we choose the same.
Collision if we choose differently.

L R
L 1, 1 -1,-1
R -1, -1 1, 1

Fangzhen Lin (HKUST) Game Theory 10 / 34

Battle of Sexes

A man and a woman are sharing a TV.

There are two channels, sport and movie.
The man prefers to watch the sport channel.
The woman prefers to watch the movie channel.

S M
S 1, 0 -1,-1
M -1, -1 0, 1

Fangzhen Lin (HKUST) Game Theory 11 / 34

The Matching Pennies Game

We each flip a coin.

I win if they match, and lose if they don’t.

H T
H 1, -1 -1,1
T -1, 1 1, -1

Fangzhen Lin (HKUST) Game Theory 12 / 34

Solution Concepts: Nash Equilibria

Given a profile s = (a1 , ..., an ), and a player i, we write

si = (a1 , ..., ai−1 , ai+1 , ..., an ), and s = (si , ai ).
A strategy a∗ ∈ Ai is player i’s best response to si if for any a ∈ Ai ,
ui (si , a) ≤ ui (si , a∗ ).
A profile s = (a1 , ..., an ) is a Nash equilibrium if for each player i, ai is
player i’s best response to si .
Given a game, there may be one, more than one, or no Nash
equilibria.

Fangzhen Lin (HKUST) Game Theory 13 / 34

Zero-Sum games (Strictly Competitive Games)

A two-person game (A, B, u1 , u2 ) is a zero-sum (or strictly competitive

game) if for all a ∈ A, b ∈ B, u1 (a, b) + u2 (a, b) = 0.
In a zero-sum game, what is good for one player is bad for the other. A
zero-sum game has a unique Nash equilibria payoff: if (a1 , b1 ) and (a2 , b2 )
are both Nash equilibria, then ui (a1 , b1 ) = ui (a2 , b2 ), i = 1, 2.
In fact, if (a, b) is a Nash equilibrium, then

miny ∈B u1 (a, y ) ≥ miny ∈B u1 (x, y )

for any x ∈ A, and similarly for any y ∈ B,

minx∈A u2 (x, b) ≥ minx∈A u2 (x, y )

Fangzhen Lin (HKUST) Game Theory 14 / 34

Mixed Strategies

P strategy of an agent is a lottery of her actions π : A → [0, 1]

A mixed
and a π(a) = 1.
The utility function can be extended to mixed strategies:

X n
Y
ui (π1 , ..., πn ) = ui (a1 , ..., an ) πk (ak )
(a1 ,...,an ) k=1

A mixed strategy profile is a (mixed) NE if no one can deviate for

better by another mixed strategy.
Nash Theorem Every game has a mixed NE.

Fangzhen Lin (HKUST) Game Theory 15 / 34

Auctions

The auction setting is important for two reasons.

Auctions are widely used in consumer, corporate, and computer
science settings: Trade goods; sell important public resources; run
financial markets; allocate broadband bandwidth and processing
powers.
The second and more fundamental reason to care about auctions is
that they provide a general theoretical framework for understanding
resource allocation problems among self-interested agents: an auction
is any protocol that allows agents to indicate their interest in one or
more resources, and that uses these indications of interest to
determine both an allocation of resources and a set of payments by
the agents.

Fangzhen Lin (HKUST) Game Theory 16 / 34

Single-Item Auctions

Examples:
1 English auctions: the auctioneer sets a starting price, and agents bid
successively.
2 Japanese auctions: the auctioneer sets a starting price, and increases
the price successively; agents decide whether to be “in” at each price
point.
3 Dutch auctions: the auctioneer sets a high starting price, and
decreases the price successively; the first agent signals a buy signal
wins the item at that price.
4 Sealed-bid auctions: agents send in bids in secret and the auctioneer
uses a protocal to decide the winner. The protocal must be known to
the agents before they send in their bids.

Fangzhen Lin (HKUST) Game Theory 17 / 34

Auctions as Structured Negotiations

An auction is a negotiation with the following rules:

Bidding rules: How are offers made (by whom, when, what can their
content be)?
Clearing rules: When do trades occur, or what are those trades (who
gets which goods, and what money changes hands) as a function of
the bidding?
Information rules: Who knows what when about the state of
negotiation?

Fangzhen Lin (HKUST) Game Theory 18 / 34

Sealed Bid Auctions

First Price Auction:

Bidders send in their bids sealed.
The highest bidder wins the item at the price of her bid.
When there is a tie, break it randomly.
Second Price Auction:
Bidders send in their bids sealed.
The highest bidder wins the item at the price of the next highest bid.
When there is a tie, break it randomly.
How should agents bid?

Fangzhen Lin (HKUST) Game Theory 19 / 34

Auctions as Games

First price auction with common knowledge of agents’ values:

a set of agents N = {1, ..., n};
the same set of actions for each agent {x | 0 ≤ x ≤ 1} (bid at x);
utility functions ui (x1 , ..., xn ) = vi − xi if agent i wins the auction, and
0 otherwise, where vi is the value that the item is worth to agent i.
Questions: Does it matter how ties are broken? What are the Nash
equilibria?

Fangzhen Lin (HKUST) Game Theory 20 / 34

Auctions as Games

Second price auction with common knowledge of agents’ values:

a set of agents N = {1, ..., n};
the same set of actions for each agent {x | 0 ≤ x ≤ 1} (bid at x);
utility functions ui (x1 , ..., xn ) = vi − xj if agent i wins the auction,
and 0 otherwise, where vi is the value that the item is worth for agent
i, and xj is the second highest value in {x1 , ..., xn }.
Questions: Does it matter how ties are broken? What are the Nash
equilibria?

Fangzhen Lin (HKUST) Game Theory 21 / 34

First Price Auction with Common Same value

Assumptions:
one single item up for auction using the first price mechanism;
ties are broken randomly;
N bidders, each can bid using prices from a fixed set P;
each has value 1 for the item.

Fangzhen Lin (HKUST) Game Theory 22 / 34

Case 1

Consider N = 2 (two bidders), and P = {0, 1} (each can bid either 0 or

1). This yields the following game:

0 1
0 0.5, 0.5 0, 0
1 0, 0 0, 0

When both bid 0, they all get expected payoff of 0.5 because the tie is
broken randomly and the winner gets payoff 1. So two Nash equilibria:
(0, 0) and (1, 1).

Fangzhen Lin (HKUST) Game Theory 23 / 34

Case 1 - A Generalization
Suppose in general there are two bids 0 ≤ a < b ≤ 1:
a b
a (1-a)/2, (1-a)/2 0, 1-b
b 1-b, 0 (1-b)/2, (1-b)/2
(b, b) will always be a Nash equilibrium.
(a, a) is a Nash equilibrium iff (1 − a)/2 ≥ 1 − b.
(a, b) cannot be a Nash equilibrium: it is a Nash equilibrium iff
0 ≥ (1 − b)/2 and 1 − b ≥ (1 − a)/2, i.e. a = b = 1, which is
impossible.
(b, b) will be the only Nash equilibrium if (1 − a)/2 < 1 − b, i.e.
a > 2b − 1.
An example of last case: a=0.3, b=0.6. In this case, the game
becomes the Prisoner’s dilemma:
0.3 0.6
0.3 0.35, 0.35 0, 0.4
0.6 0.4, 0 0.2, 0.2
Fangzhen Lin (HKUST) Game Theory 24 / 34
General Case
In general, suppose there are n ≥ 2 players, and each can bid
0 ≤ a1 < · · · < am ≤ 1, m ≥ 2.
(am , ..., am ) is always a Nash equilibrium;
(ai , ..., ai ), 1 ≤ i < m, is a Nash equilibrium iff (1 − ai )/n ≥ 1 − ai+1 ;
suppose x = (x1 , ..., xn ) has two different elements, and the largest
element is ai which occurs k times.
I if k = 1, then x is not a Nash equilibrium: if ai = 1, then the player
who bids ai can increase her expected payoff by lowering her bid to
ai−1 ; if ai < 1, then players whose bids are lower than ai can increase
their expected payoff by raising their bids to ai .
I If k > 1, then x is a Nash equilibrium iff ai = am = 1.
If each player can bid arbitrarily in a continuous close interval [l, m], then
the unique Nash equilibrium is (m, ..., m). If the upper bound is open:
[l, m), then there is no Nash equilibrium.

Fangzhen Lin (HKUST) Game Theory 25 / 34

Auctions as Games

But it is not reasonable to expect agents will know each other’s

valuations of the item.
What the item is worth to an agent is her own private information.
How to model uncertainties: probabilities.

Fangzhen Lin (HKUST) Game Theory 26 / 34

Auctions as Bayesian Games
In general, a sealed single item auction consists of
a set of agents N = {1, ..., n};
for each agent, a set of possible private values, and a set of possible
bids - we assume they are both [0, 1] here;
a common joint prior on value profiles: p : [0, 1]n → [0, 1];
a payment function σ from bid profiles and agents to [0.1]:
σ : [0, 1]n × N → [0, 1],
meaning that σ(b, i) is what i needs to pay when the bid profile is b;
a winner selection function τ from bid profiles to agents.
A strategy is now a function from agent’s private values to bids. Assuming
agents are risk-neutral, the utility function for agent i is then
X
ui (f1 , ..., fn ) = p(v )ui (f1 (v1 ), ..., fn (vn ), v1 , ..., vn ),
v

where ui (b, v ) = vi − σ(b, i) if τ (b) = i, and −σ(b, i) otherwise.

Fangzhen Lin (HKUST) Game Theory 27 / 34
Nash Equilibria for Auctions

For second-price auction, the Nash quilibrium is to bid your valuation:

(v1 , ..., vn ).
For first-price auction, if each player’s valuation is drawn
independently and uniformly at random, then
(v1 (n − 1)/n, ..., vn (n − 1)/n) is a Nash equilibria.

Fangzhen Lin (HKUST) Game Theory 28 / 34

Revenue Equivalence

A landmark result of auction theory is Vickrey’s celebrated revenue

equivalence theorem: any auction in which
the bidder with the highest valuation always wins
the bidder with the lowest possible valuation expects zero payoff
all bidders are risk neutral, and
bidders’ valuations are drawn from a strictly increasing and atomless
distribution
will lead to the same expected revenue for the seller (auctioneer).

Fangzhen Lin (HKUST) Game Theory 29 / 34

Iterated Prisoner’s Dilemma

The general form of the prisoner’s dilemma:

Cooperate Defect
Cooperate R=3, R=3 S=0, T=5
Defect T=5, S=0 P=1, P=1

Here R is rewards for mutual cooperation, P penalty for mutual defection,

T the temptation to defect, and S sucker’s payoff.
In general, T > R > P > S, and for iterated prisoner’s dilemma:
2R > T + S so that cooperate twice is better than taking turns to
exploit each other.
The game has a dominant Nash equilibrium (D, D), and this is the
strategy to play if the players are to play the games for a known
number of times.

Fangzhen Lin (HKUST) Game Theory 30 / 34

Computer Prisoner’s Dilemma Tournament (Axelrod 1984)
Rules:
Each program is going to play another player an unknown number of
times, determined by a probability (the probability of ending the
match is set to be 0.00346).
Each program has available to it the history of interactions so far.
Round one:
14 entries + RANDOM, round robin, each pair of entries was
matched in 5 games of 200 moves each. An entry’s score is the sum
of its scores in all games it played.
The winner was Tit For Tat (TFT) submitted by Anatol Rapoport, a
psychologist from University of Toronto.
TFT cooperates on the first move and there after plays whatever
move the other player played last time.
Round two: conducted after the results of round one were announced:
62 entries + RANDOM, from 6 countries.
TFT were again the winner.
Fangzhen Lin (HKUST) Game Theory 31 / 34
References

Axelrod, Robert (1984), The Evolution of Cooperation, Basic Books, ISBN

0-465-02122-0
A List of Strategies:
http://www.iterated-prisoners-dilemma.net/
prisoners-dilemma-strategies.shtml

Fangzhen Lin (HKUST) Game Theory 32 / 34

Some Formal Properties (Axelrod)

Discount factor It is typical to discount future. In iterated prisoner’s

dilemma, future encounters can be similarly discounted, say by a weight w ,
0 ≤ w ≤ 1, so scores from the nth encounter contributes only w (n−1) s,
where s is the payoff received at the nth game. In the iterated game
setting, the weight is the same as the probability of playing your oponent
again the next time.
Proposition 1 If the discount weight w is sufficient large, then there is no
strategy that is best against all possible strategies used by the other player.
Think about which strategy is best against ALL D or Permanent
Retaliation (C until the oponent plays D, then ALL D).

Fangzhen Lin (HKUST) Game Theory 33 / 34

Properties (Cont’d)

TFT is also robust: performs well in a wide variety of environments. This

is an informal statement supported by the tournaments and further
experiments.
A strategy is colelctively stable if no strategy can invade it. A strategy S
can invade strategy G if given a large group of individuals all using G , a
new individual using S can get a better score than the individuals using G .
Proposition 2 TFT is collectively stable if and only if, w is large enough.
This value of w is a function of the four payoff parameters T , R, P, S.
Proposition 3 ALL D is always collectively stable. However, ALL D can
be invaded by a cluster of individuals, and this cluster needs only be a
small proposition of the original population.
Proposition 4 A strategy is nice if it is not the one to defect the first. If a
nice strategy cannot be invaded by an individual, then it cannot be
invaded by any cluster of the original population.

Fangzhen Lin (HKUST) Game Theory 34 / 34