GOTha, 12/05/06, LIP 6

Algorithmic Game Theory and

Scheduling

Eric Angel, Evripidis Bampis, Fanny Pascual

IBISC, University of Evry, France

 Outline
 Scheduling vs. Game Theory

 Stability, Nash Equilibrium

 Price of Anarchy

 Coordination Mechanisms

 Truthfulness
 Scheduling
(A set of tasks) + (a set of machines)
(an objective function)

Aim: Find a feasible schedule optimizing the

objective function.
 Game Theory
(A set of agents) + (a set of strategies)
(an individual obj. function for every agent)

Aim: Stability, i.e. a situation where no agent has

incentive to unilaterally change strategy.

Central notion: Nash Equilibrium (pure or mixed)

 Game Theory (2)

Nash: For any finite game, there is always a

(mixed) Nash Equilibrium.

Open problem: Is it possible to compute a Nash

Equilibrium in polynomial time, even for the case
of games with only two agents ?
 Scheduling & Game Theory
The KP model:
(Agents: tasks) + (Ind. Obj. F. of agent i:
the completion time of the machine on
which task i is executed)
The CKN model:
(Agents: tasks) + (Ind. Obj. F. of agent i:
the completion time of task i)
 Scheduling & Game Theory (2)

The AT model:
(Agents: uniform machines) + (Ind. Obj. F.
of agent i: the profit defined as Pi-wi/si)

Pi: payment given to i

Wi : load of machine i
: the speed of machine i
Si
 The Price of Anarchy (PA)
Aim: Evaluate the quality of a Nash Equilibrium.
[Koutsoupias, Papadimitriou: STACS’99]

Need of a Global Objective Function (GOF)

PA=(The value of the GOF in the worst NE)/(OPT)

It measures the impact of the absence of

coordination
[In what follows, GOF: makespan]
 An example: KP model
[Koutsoupias, Papadimitriou: STACS’99]

3 tasks
1 2 2 machines
1 3

0 1 2 3 time

A (pure) Nash Equilibrium

Question:
How bad can be a Nash Equilibrium ?
 An example: KP model
pij : the probability of task i to go on machine j

The expected cost of agent i, if it decides to go

on machine j with pij =1:

Ci j = li + S pj k lk
Ki
In a NE, agent i assigns non zero probabilities
only to the machines that minimize Ci j
 An example
Instance: 2 tasks of length 1, 2 machines.

A NE: pij = 1/2 for i=1,2 and j=1,2

C11= 1 + 1/2*1 = 3/2

C12= C21= C22=3/2

Expected makespan
1/4*2+1/4*2+1/4*1+1/4*1 =3/2

OPT = 1
 The PA for the KP model
Thm [KP99]: The PA is (at least and at most)
3/2 for the KP model with two machines.

Thm [CV02]: The PA is Q(log m/(log log log m))

for the KP model with m uniform machines.
 Pure NE for the KP model
Thm [FKKMS02]: There is always a pure NE for
the KP model.

Thm [V02]: The PA (pure Nash eq.), is 2-2/(m+1)

for the KP model with m identical machines

Thm [CV02]: The PA is Q(log m/(log log log m))

for the KP model with m identical machines.

Thm [FKKMS02]: It is NP-hard to find the best

and worst equilibria.
 Pure NE for KP and local
search
Nash eq. => local optimum (with Jump)

The converse is not true.

4 4 1
2 4 2 4
5 1 5

A local optimum
Not a Nash eq.
 How can we improve the PA ?
Coordination mechanisms
Aim: to decrease the PA
What kind of mechanisms ?
-Local scheduling policies in which the schedule on each
machine depends only on the loads of the machine.
-each machine can give priorities to the tasks and introduce
delays.
 The LPT-SPT c.m. for the CKN model

M1 1 1 SPT

M2 2 2 LPT
0 4

M1 1 2
M2 2 1

0 3

Thm [CKN03]: The LPT-SPT c.m. has a price of

anarchy of 4/3 for m=2.

[The LPT c.m. has a PA of 4/3-1/3m]

 The Price of Stability (PS)
The framework: A protocol wishes to propose a collective
solution to the users that are free to accept it or not.

Aim: Find the best (or a near optimal) NE

PS = (value of the GOF in the best NE)/OPT

Example:
- PS=1 for the KP model

- PS=4/3-1/3m for the CKN model (with LPT l.p.)

Nashification for the KP model
Thm [E-DKM03++]: There is a polynomial time
algorithm which starting from an arbitrary
schedule computes a NE for which the value of the
GOF is not greater than the one of the original
schedule.

Thus: There is a PTAS for computing a NE of

minimum social cost for the KP model.
 Approximate Stability
Aim: Relax the notion of stable schedule in order to improve
the price of stability.

a-approx. NE: a situation in which no agent has

sufficient incentive to unilaterally change
strategy, i.e. its profit does not increase more
than a times
its current profit.
Example: a 2-approx. NE
M1 LPT
3 3
M2 2 2 2 LPT
 The algorithm LPTswap
Thm[ABP05]: LPTswap returns a 3-approx. NE and
has an approximation ratio of 8/7. Therefore the
price of 3-approximate stability is less than 8/7 (LPT l.p).

-construct an LPT schedule

Exchange: (x1,y1), or (x1,y2),
-1st case: x1 x2 x3 or (x2,y2)
y1 y2 Return the best or LPT

-2nd case: x1 x2 x3 x4
y1 y2
Exchange: (x3+x4,y2)
Compare with LPT and return
the best
-3rd case: Return LPT
 Truthful algorithms

The framework:

Even the most efficient algorithm may lead

to unreasonable solutions if it is not
designed to cope with the selfish behavior
of the agents.
 CKN model: Truthful algorithms
• The approach:
– Task i has a secret real length li.
– Each task bids a value bi ≥ li.
– Each task knows the values bidded by the other tasks, and
the algorithm.

• Each task wish to reduce its completion time.

• Social cost = maximum completion time (makespan)

• Aim : An algorithm truthful and which minimizes the

makespan.
[Christodoulou, Koutsoupias, Nanavati: ICALP’04]
 Two models
• Each task wish to reduce its completion time
(and may lie if necessarily).
• 2 models:
– Model 1: If i bids bi, its length is li
– Model 2: If i bids bi, its length is bi

• Example: We have 3 tasks: , ,

Task 1 bids 2.5 instead of 1: 1 2 3

.
3 Model 1: C1 = 1
1 2 Model 2: C1 = 2.5

0 1 2 3 4 5 time
 SPT: a truthful algorithm
• SPT: Schedules greedily the tasks from the
smallest one to the largest one.
– Example:
1 3

– Approx. Ratio = 2 – 1/m [Graham]

• Are there better truthful algorithms ?

 LPT
• LPT: Schedules greedily the tasks from the
largest one to the smallest one.
– Approx. Ratio = 4/3 – 1/(3m) [Graham]

• We have 3 tasks: , ,
1 2 3
Task 1 bids 1: Task 1 bids 2.5:

3 C1 = 3 3
2 1 2 C1 = 1
1

0 1 2 3 4 5 time 0 1 2 3 4 5 time

Task 1 has incentive to bid 2.5, and LPT is not truthful.

 Randomized Algorithm
• Idea: to combine:
– A truthful algorithm
– An algorithm not truthful but with a good approx.
ratio.

• Task: wants to minimizes its expected

completion time.

• Our Goal: A truthful randomized algorithm

with a good approx. ratio.
 Outline

 Truthful algorithm
SPT-LPT is not truthful
Algorithm: SPT
A truthful algorithm: SPT-LPT
 SPT-LPT is not truthful
• Algorithm SPT-LPT:
– The tasks bid their values
– With a proba. p, returns an SPT schedule.
With a proba. (1-p), returns an LPT schedule.
• We have 3 tasks : , ,
– Task 1 bids its true 1value : 12 3

SPT : 1 3 3 C1 = p + 3(1-p)
– Task 1 bids2 a false value : LPT
2.5 : 2 1 = 3 - 2p

SPT : 1 1 3 C1 = 1
2 3 LPT :
1 1 2
 Algorithm SPT
• SPT:
Schedules tasks 1,2,…,n s.t. l1 < l2 < … < ln
Task (i+1) starts when 1/m of task i has been
executed.
• Example: (m=3)

1 4 7
2 5 8
3 6 9

0 1 2 3 4 5 6 7 8 9 10 11 12
 Algorithm SPT
• Thm: SPT is (2-1/m)-approximate.
• Idea of the proof: (m=3)
1 4 7
2 5 8
3 6 9

12
• 0 Idle
1 2 3 4
times : 5 6 7 8 9 10 11

idle_beginning(i) = ∑ (1/3 lj)

idle_middle(i) = 1/3 ( li-3 + li-2 + li-1 ) – li-3
idle_end(i) = li+1 – 2/3
j<i
li + idle_end(i+1)
 Algorithm SPT
• Thm: SPT is (2-1/m)-approximate.
• Idea of the proof: (m=3)
Cmax
1 4 7
2 5 8
3 6 9

0 1 2 3 4 5 6 7 8 9 10 11 12

Cmax = (∑(idle times) + ∑(li)) / m

∑(idle times) ≤ (m-1) ln and ln ≤ OPT
 Cmax ≤ ( 2 – 1/m ) OPT
A truthful algorithm: SPT-LPT
• Algorithm SPT-LPT:
– With a proba. m/(m+1), returns SPT.
– With a proba. 1/(m+1), returns LPT.

• The expected approx. ratio of SPT - LPT is

smaller than the one of SPT: e.g. for m=2,
ratio(SPT-LPT) < 1.39, ratio(SPT)=1.5

• Thm: SPT-LPT is truthful.

 A truthful algorithm: SPT-LPT
• Thm: SPT-LPT is truthful.
Idea of the proof:
• Suppose that task i bids b>li. It is now larger than
tasks 1,…, x, smaller than task x+1.

l1 < … < li < li+1 < … < lx < lx+1 < … < ln
b<
• LPT: decrease of Ci(lpt) ≤ (li+1 + … + lx)
• SPT: increase of Ci(spt) = 1/m (li+1 + … + lx)
• SPT-LPT:
change = - m/(m+1) Ci(spt) + 1/(m+1) Ci(spt) ≥ 0
AT model: Truthful algorithms
Monotonicity: Increasing the speed of
exactly one machine does not make the
algorithm decrease the work assigned to
that machine.

Thm [AT01]: A mechanism M=(A,P) is

truthful iff A is monotone.
 An example
The greedy algorithm is not monotone.

Instance: 1, e, 1, 2-3 e, for 0<e<1/3

Speeds (s1,s2) M1 M2
(1,1) e, 1 1, 2-3 e
(1,2) e, 2-3 e 1,1
 Results for the AT model
3-approx randomized mechanism [AT01]

(2+e)-approx mechanism for divisible speeds and

integer and bounded speeds [ADPP04]

(4+e)-approx mechanism for fixed number of

machines [ADPP04]
12-approx mechanism for any number of machines
[AS05]
 Conclusion
• Future work:
-Links between LS and game theory
-Many variants of scheduling problems
…