+ All Categories
Home > Documents > Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a...

Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a...

Date post: 21-Dec-2015
Category:
View: 230 times
Download: 2 times
Share this document with a friend
Popular Tags:
39
Distributed Combinatorial Optimization
Transcript
Page 1: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.

Distributed Combinatorial Optimization

Page 2: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.

Abstract

• Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards reconstructing an integer solution from the fractional one is a standard technique in a non distributed scenario.

• However, this method has not often been applied for distributed algorithms (there is only one example: Constant time distributed dominating set approximation by F.Kuhn and R. Wattenhofer)

• There are some problems that can be defined as distributed problems. For example:– To determine appropriate transmission power level for every node in sensor

networks. – MDS problem: can be very useful for routing in mobile ad-hoc networks

Page 3: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.

Abstract

• The most efficient solution for the distributed problems is distributed algorithm.

• The main contribution of the presented paper is fast LP distributed approximation algorithm, using which, combined with randomized rounding techniques, efficient distributed approximation algorithms can be obtained.

Page 4: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.

Introduction

• The challenge: achieve a global goal based on local information.

• Trade off: time complexity v.s. approximation ratio.

• The structure of ad-hoc networks is changing rapidly due to the node mobility. Along with scarceness of the resources (like energy and bandwidth) this leads to the need for low message and time complexities algorithms.

Page 5: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.

Definitions

• Why to use approximation algorithms?

For NP-hard problems, like vertex cover, TSP or knapsack there is no polynomial time solution (unless P=NP).

Approximation algorithms have to be efficient, i.e. run in polynomial time.

Page 6: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.

Definitions

• For a minimization problem, a polynomial time algorithm A is said to be an approximation algorithm with approximation ratio

if and only if for every instance of the problem, A gives a solution which is at most times the optimal value for that instance, .

• For a maximization problem, a polynomial time algorithm A is said to be an approximation algorithm with approximation ratio

if and only if for every instance of the problem, A gives a solution which is at least times the optimal value for that instance, .

• A is said to be - approximation algorithm.• is an approximation ratio of A.

1

1

Page 7: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.

Definitions

• Primal Linear Program, minimization problem (fractional covering problem) example:

• Optimal solution, if exists:

0,,

,0

11

1

ija

ib

jc

jx

,...,n, ii

bj

xm

jij

a

jx

m

jj

c

m1,...,j

to subject

minimize

),...,(***

1 nxxx

Page 8: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.

Definitions

• Dual Linear Program,

maximization problem (fractional packing problem)

example:

Optimal solution, if exists:

0

10

11

1

ji,a

i,b

j c

,...,n, ii

y

, ...,m, jj

ci

yn

iji

a

yin

ibi

to subject

maximize

)*

,...,*1(

*myyy

Page 9: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.

Definitions

• LP relaxation: In general, relaxation refers to the action of relaxing the integer requirement of linear IP (integer program) to turn it into an LP. For example: for the Vertex Cover problem linear IP program is:

• The LP corresponding to the above IP is (after relaxation action):

Vi},,{i x

Eij, jxix

nx...xx

10

1 subject to

21 min

Vi,ix

Eij, jxix

nx...xx

10

1 subject to

21 min

Page 10: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.

Definitions

• Any assignment of variables that satisfies the constraints is called a feasible solution.

• Obviously, if the LP version is infeasible, then the IP version is also infeasible.

• Rounding: Solve LP and convert fractional solution into integer solution. Rounding is polynomial, but can be time consuming e.g., if LP relaxation has exponentially many constraints.

Page 11: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.

Definitions

• Randomized Rounding: Solve LP, then randomly round fractional values to integer values.

• Round of communication: • Generating the message [ + processing the message from

previous round]• Sending the message to the outgoing neighbor

Page 12: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.

Introduction - cont

The main parts(contributions) of the paper:1. A novel deterministic distributed algorithm which achieves a

approximation for fractional covering and packing problems in only

rounds where and are the maximum

number of times a variable occurs in the inequalities of primal and dual LP, respectively, and where denotes the ratio between the largest and the smallest non-zero coefficient of the LP (here we use messages of logarithmic size).

2. If a message size isn’t important the approximations can be obtained even faster.

1

4loglog )/εΔ(ρ)Δ(O pd dp

Page 13: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.

Introduction - cont

3. Combined with randomized rounding techniques, the above algorithms can be used to efficiently approximate a number of combinatorial problems.

Page 14: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.

Algorithm – notations

• The number of primal and dual variables is m and n, respectively.

• The linear program is bound to a network graph G=(V,E).

minmax

0minmin

maxmax

/aaρ

}}\{i,ci,bij{ai,ja

}i,ci,bij{ai,ja

Page 15: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.

Algorithm – notations

• Each primal variable xi and each dual variable yj is associated with a node respectively.

• There are communication links between primal and dual nodes wherever the respective variables occur in the corresponding inequality. Thus, iff xi occurs in the j-th inequality of LP. Formally, this means that are connected iff aji > 0.

,)(

,)(

Vd

jvV

p

iv

E)(d)j,v(p)

i(v

)(,)( dj

vpi

v

Page 16: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.

Algorithm – notations LP

• The degree of is called .

• The primal degree: .

• The dual degree:

• The set of dual neighbors of is denoted by

• The set of primal neighbors of is denoted by

)( piv )( p

i

)(max piip

)(max djjd

)( pi

N)( piv

)( div )(d

iN

Page 17: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.

Algorithm – assumptions LP

• Purely synchronous communication model.• All nodes know .

m

j ijai

d Γ,n

j jiaic

c

ipΓ

1max

1max

max

hfcc dpi and ,,},max{max

p/k

p

p

/kp

p

ΓΓ

k h

Γ

kf

pp ln1

1

111

Page 18: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.

LP Algorithm

• First step: convert the coefficients to the following form:• This can be achieved by dividing every aij by bi, and

replacing bi with 1. After that, the ci and aij are divided by =minj{aji}\{0}.

• The optimal objective values stay the same.• A feasible solution for the transformed LP can be easily

converted to a feasible solution for original LP, by dividing all x-values by the corresponding , and by dividing the y –values by the corresponding bi.

1or 0,1:, ij

aij

ai

bji

i

i

Page 19: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.

LP Algorithm

• Second step: to measure the efficiency per cost ratio of a primal node , we define as follows:

– when is a variable that belongs to each dual node and which is decreased every time the corresponding primal constraint is achieved ( ).

Besides that, every dual node has a variable which counts how many times the primal constraint has been fulfilled.

)( piv

jrj

jiaic

ci max

)( div

i

1jr

if

jr

Page 20: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.

LP Algorithm

• The algorithm has two parameters: , and which determine the trade-off between time complexity and the approximation quality.

• The bigger kp and kd, the better the approximation ratio of the algorithm. On the other hand, smaller kp and kd lead to a faster algorithm.

1pk1dk

Page 21: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.
Page 22: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.
Page 23: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.

Analysis

• Lemma 1: For each primal node , at all times during Algorithm holds:

• Proof: we have to look how fast decreases comparing to the outer loop, particularly, how fast it decreases comparing to .

)(piv

pp kepi

/)2(

ipp ke

p/

Page 24: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.

Analysis

• Lemma 2: After line 12 of Algorithm, for each dual node , either or

• Proof: if after inner loop iteration then the same primal variables will be increased in the next iteration, and then of the next iteration will be greater by than the previous, and if both conditions are false we will reach a contradiction.

)( div

0~ ir d/kdjj iji Γxaw 1

0~ ir iw

d/kdΓ 1

Page 25: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.

Analysis

• Lemma 3: Each time a dual node enters increase_duals() in Algorithm, holds: (1)

– and (2)

• Proof: The condition can be violated only in increase_duals() procedure. • Inside increase_dual() procedure the interesting case we should consider is:

• The second inequality follows from the first and lemma 2.

pp

d

pp

kep

kd

ii

kep

iii

ry

wry

/

/1

/

1

12 iw

Page 26: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.

Analysis

• Lemma 4: Let be a primal node and let be the weighted sum of the y-values of its dual neighbors. Further, let be the increase of and be the decrease of during an execution of increase_duals(). We have:

• Proof: The inequality holds for every dual neighbor of .

)(piv j

jjii yaY

iY iY

i

i

i

ik

pi

kd

kp

kp

i c

cY

p

dpp

max

/1

/1/1/3

1

},max{

)(djv

)( p

iv

Page 27: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.

Analysis

.definition and 1, lemma of because

follows lemma the1

max that show weIf

. of decrease thebe and of increase be Let

1

111

ii

-j/k

p/ke

p

/kd

/kp

/kp

j

jjjj

Y

rΓΓ

},Γ{ΓΓβ

rry

ppp

dpp

Page 28: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.

Analysis

• Lemma 5: Let be a primal node and let be the weighted sum of the y-values of its dual neighbors. At line (20) of the algorithm holds:

• Proof: Look at the minimal value and the maximal value of during the iterations, an the sum of the (using Lemma 4) in every case.

i

)( piv j

jjii yaY

}d/kd,Γp/k

p{Γp/k

pΓfpkc

iciY

11max

31

max

iY

Page 29: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.

Analysis

• Lemma 6: At the end of the algorithm we have• Proof: counts the number of times the i-th constraint of (LP) is satisfied. So the claim can

be proved by looking at -th increment in the line 13 of the Algorithm.

• Lemma 7: After the main part of the algorithm (in line 21) we have:

• Proof: Let be a primal node. Look at the sum of increases (which by lemma 6 are 0 in the end) over its all dual neighbors.

ffandri ii 0:

ifiw

n

jjyc

ix

m

iic

1max

1)( p

iv

jy

Page 30: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.

Analysis

• Theorem: For arbitrary , Algorithm approximates (LP) and (DLP) by a factor

and the time complexity of Algorithm is:

1, pd kk

d/kdΓp/k

pΓp/k

pΓ1

,1

max4

)O( ),(log

log1

1

11 /1/1

dppp

pk

p

p

kp

dp

kktosimplifiesthisOkfor

kkkO

pp

Page 31: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.

Analysis

• Proof: • Each primal constraint is satisfied at least f times (lemma 6), so in line 21 of the

Algorithm all primal variables are divided by at least f.

• The sum of the y values of each primal nodes’ dual neighbors is at most according to lemma 5.

• Therefore, dividing all dual variables by

makes the dual solution feasible.max

113 max1

c

},Γ{ΓΓ)f(k

dpp /kd

/kp

/kp

p

max

11max

3

1c

}d/kd,Γp/k

p{Γp/k

pΓ)fp(kic

Page 32: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.

Analysis

• By lemma 7, the ratio between the objective functions of the primal and the dual solution becomes:

}d/kd

,Γp/k

p{Γp/k

Qp/k

pΓc

Q

p/k

pk

p/k

pk

pk

cn

jjy

m

iixic

11max

4

1

max

11

1

1

11

1

max

1

1

max

113 max

c

},Γ{ΓΓQ

dpp /kd

/kp

/kp

Page 33: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.

Analysis

• According to duality theorem for the linear programming this ratio is an upper bound on the approximation ratio for (LP) and (DLP).

• Time complexity can be obtained by the fact that number of rounds is proportional to the number of iterations of the inner most loop.Each iteration of the inner most loop takes two rounds. Hence, the algorithm has time complexity: and if we will substitute the actual values for f and h we will get the desired expression.

• -

))(( hfkkO pd

constant. a is )1/1

p( because

),O( tosimplifies this),(logFor

pk

dkpkpOpk

Page 34: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.

Analysis

• Collorary: For sufficiently small ε , Algorithm computes a (1+ ε ) -approximation for (LP) and (DLP) in rounds. In particular, a constant factor approximation can be achieved in time .

• Proof: We chose kp and kd such that

• then, we get .

'constant for )(log and )(log dOdkpOpk

)/log(log 4dpO

)log(log dpO

'1/1/1 dp kd

kp

Page 35: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.

Fast LP Algorithm

• If message size doesn’t have to be bounded, LP and DLP approximation can be obtained even faster.

• The above can be achieved by using N.Lineal and M.Sacks randomized distributed algorithm to decompose a graph into sub-graphs of limited diameter(“Low Diameter Graph Decompositions”, by N.Lineal and M.Sacks ).

Page 36: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.

Fast LP Algorithm

• Graph decompositions are used in distributed computing as a tool for fast distributed algorithms. • The goals:

• decentralize computations (symmetry-breaking)• exploit locality

• Graph decomposition is partition of its vertex set into a small number of disjoint sets (blocks), when each one of them has small diameter. Good trade off between these qualities can be used to improve performance.

Page 37: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.

Fast LP Algorithm

• For every n-vertex graph, Lineal and Sacks randomized distributed algorithm, yields a decomposition of diameter O(log n) into O(log n) blocks. The time complexity is O(log^2 n).

• In this paper, Lineal and Sacks randomized distributed algorithm, is used to decompose the linear program into sub programs which can be solved locally.

Page 38: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.

Randomized rounding

• The LP approximation algorithms can be used with randomized rounding techniques to obtain distributed approximation algorithms for a number of combinatorial problems.

• There are another two distributed algorithms presented in the paper that compute logarithmic approximation integer solutions from the solutions to relaxed LP programs by using randomized rounding technique in constant number of rounds.

Page 39: Distributed Combinatorial Optimization. Abstract Approximating integer linear programs by solving a relaxation to a linear program (LP) and afterwards.

Conclusion

• The presented algorithms alone or with randomized rounding technique can obtain efficient approximation for number of combinatorial problems, and they match or ever outperform the best known problem specific algorithms in terms of running time and approximation quality.


Recommended