randomized competitive analysis for two server problems
DESCRIPTION
Randomized Competitive Analysis for Two Server Problems. Wolfgang Bein Kazuo Iwama Jun Kawahara. k-server problem. Goal: Minimize the total distance. k-server problem. k-server problem. k-server problem. k-server problem. k-server problem. k-server problem. k-server problem. ………. - PowerPoint PPT PresentationTRANSCRIPT
Randomized Competitive Analysis for Two Server Problems
Wolfgang Bein
Kazuo Iwama
Jun Kawahara
k-server problem
Goal: Minimize the total distance
k-server problem
k-server problem
k-server problem
k-server problem
k-server problem
k-server problem
k-server problem
………
Greedy does not work,
2-server 3-point problem
2-server 3-point problem
2-server 3-point problem
2-server 3-point problem
2-server 3-point problem
a
b c
Adversary (always malicious): cababacb……Opt: cababacb……
one move per two inputs 2.0CR
Algorithm exists whose CR = 2.0
k-server: Known Facts
• Introduction of the problem [Manasse, McGeoch, Sleator 90]
• Lower bound: k [MMS90]• General upper bound: 2k-1 [Koutsoupias,
Papadimitriou 94]• k-server conjecture
– true for 2-servers, line, trees, fixed k+1 or K+2 points, ……
– still open for 3 server 7 points
Randomized k-server
• Very little is known for general cases• Even for 2-servers (CR=2 for det. case):
– On the line [Bartal, Chrobak, Lamore 98]
– Cross polytope space [Bein et al. 08]– Specific 3 points: Can use LP to derive an
optimal algorithm (but nothing was given about the CR) [Lund, Reingold 94]
– Almost nothing is known about its CR for a general metric space
1551.987
78CR
19
12CR
Randomized 2-server 3-point
a
b c
Adversary is malicious: c……
: 2.0 1.5CR
Select a server (a or b) at random
Adversary’s attempt fails with prob. 0.5
Our Contribution
• For (general) 2-server 3-point problem, we prove that CR < 1.5897.
• Well below 2.0 (=the lower bound for the deterministic case): Superiority of randomization for the server problem
• Our approach is very brute force
The Idea
• We can assume a triangle in the plain wolg.
• For a specific triangle, its algorithm can be given as a (finite) state diagram, which can be derived by LP [LR94]
• Calculation of its CR is not hard.
• Just try many (different shaped) triangles, then…..
111541511154151323
1
C
L
C
L R
11R L
R
R
1 2
The Idea
• We can assume a triangle in the plain wolg.
• For a specific triangle, its algorithm can be given as a (finite) state diagram, which can be derived by LP [LR94]
• Calculation of its CR is not hard.
• Just try many (different shaped) triangles, then…..
Testing Many Triangles
……
CR=1.5 1.53 1.489 1.533
1.536
almost the same but CR=1.89
1.537
1.0
Approximation Lemma
Line Lemma
Approximation Lemma
1.0 1a
1b
1.02a
2b
r r 1 1
2 2
max ,a b
a b
1OPT 2OPT
1 2
1a a
1
1b
11.0
1OPT
12
OPT OPT
2 2 1 1
2 1 1 1
ALG ALG ALG ALG
OPT OPT OPT OPT
Proof
algorithm A
Line Lemma1 n
1n
algorithm s.
1 11
1
1t.
1
1n
n
nn n
CR C
n
1.58191
e
e
Using Approximation Lemma
a1
b
2
( , )a bd
d
a
b
a
b
algorithm :A CR r
:a
A CR ra d
1
1
Using Line Lemma
a1
b
2
( , 1)n n
ab
: nA CR C1 n
1 11
1
11 1
n
n
nn n
C
n
decreasing
:2 n
nA CR C
n
2
a1
b
2
0Proving CR r
0
00 0
0
( )2 n
nf S C r
n
0( ) ii i
i i
af S r r
a d
( , );i i ia b CR r
id
0S
iS
finite set ofsquares(triangles)
Our algorithm = algs for squares + alg for the line
a1
b
2
Proving Better CR
jS ( ) : Maximumjf S
Computer Program
a1
b
2
a1
b
2
Computer Program
Computer Program
a1
b
2+ some heuristics
Some Data
• Conjecture: 1.5819
• Current bound: 1.5897– 13,285 squares, d=1/256~1
• Small area [5/4, 7/4, 1/16]: 1.5784– 69 squares, d=1/64~1/128
• Small area [7/4, 9/4, 1/4]: 1.5825– 555 squares, d=1/2048~1/64
(5/4, 7/4)
1/16
Proof for Line Lemma1 n
1n
algorithm s.
1 11
1
1t.
1
1n
n
nn n
CR C
n
1.58191
e
e
Final Remarks
• Strong conjecture that the real CR is e/(e-1). Analytical proof?
• Extension of our approach to, say, the 4-point case.
• Many very interesting open problems in the online server problem.
Thanks!