THE QUADRATIC ASSIGNMENT PROBLEM - Åbo Akademiweb.abo.fi/fak/tkf/at/ose/doc/Pres_29112012/The...
Transcript of THE QUADRATIC ASSIGNMENT PROBLEM - Åbo Akademiweb.abo.fi/fak/tkf/at/ose/doc/Pres_29112012/The...
OSE SEMINAR 2012
THE QUADRATIC ASSIGNMENT PROBLEM
Axel Nyberg
CENTER OF EXCELLENCE INOPTIMIZATION AND SYSTEMS ENGINEERINGÅBO AKADEMI UNIVERSITY
ÅBO, NOVEMBER 29, 2012
2 | 16
Introduction
I Introduced by Koopmans and Beckmann in 1957
I Cited by ≈ 1500
I Among the hardest combinatorial problems
I Real and test instances easily accessible (QAPLIB - A Quadraticassignment problem Library)
I Instances with N=30 are still unsolved
THE QAP
Center of Excellence in Optimization and Systems Engineering at Åbo Akademi University
Applications 3 | 16
Location theory
I Supply Chains
I Logistics
I Production
THE QAP
Center of Excellence in Optimization and Systems Engineering at Åbo Akademi University
Applications 3 | 16
Location theory
I Objective: Assign N plants between N given locations in order tominimize total flows.
1
2
3
4
5
36
4
2
2
3
3
3
4
1
A=
0 3 6 4 23 0 2 3 36 2 0 3 44 3 3 0 12 3 4 1 0
1
2
3 4
5
10
15
7
56
4
25
B=
0 10 15 0 7
10 0 5 6 015 5 0 4 20 6 4 0 57 0 2 5 0
THE QAP
Center of Excellence in Optimization and Systems Engineering at Åbo Akademi University
Applications 3 | 16
Location theory
12
24
35
43
51
3*6
4*5
2*10
2*5
3*4 3*2
4*7
1*15
I Optimal solution = 258I Optimal Permutation=[2 4 5 3 1]
A=
0 3 6 4 23 0 2 3 36 2 0 3 44 3 3 0 12 3 4 1 0
B24531 =
0 6 0 5 106 0 5 4 00 5 0 2 75 4 2 0 15
10 0 7 15 0
THE QAP
Center of Excellence in Optimization and Systems Engineering at Åbo Akademi University
Applications 4 | 16
Facility Layout-Real World Examples
I Hospital Layout - Germanuniversity hospital, KlinikumRegensburg 1972.(Optimality proved in the year2000)
I Ship Design
I Airport gate Assignment
I Nature Park Layout
THE QAP
Center of Excellence in Optimization and Systems Engineering at Åbo Akademi University
Applications 5 | 16
DNA MicroArray Layout
I Microarrays can have up to 1.3 million probes
I Small subregions can be solved as QAPs
I Objective: To reduce the risk of unintended illumination ofprobes
AGTCGCACGCGTAGA
ATTTGGAGCCGTCGA
TGTTGATGCGGAGGC
CGTCGCCTCCGAGGT
ACTCGAGGAAGATGT
Sub-region (size N = 25)
THE QAP
Center of Excellence in Optimization and Systems Engineering at Åbo Akademi University
Applications 6 | 16
Other applications for QAPs
I Backboard wiring
I Control panel and keyboard layout
I VLSI design
I Computer manufacturing
I Archeology
I Bandwith minimization of a graph
I Economics
I Image processing
THE QAP
Center of Excellence in Optimization and Systems Engineering at Åbo Akademi University
Formulations 7 | 16
Three Objective Functions
I Koopmanns-Beckmann
minA ·XBXT (1)
I SDPmintr(AXBXT ) (2)
I DLRminXA ·BX (3)
THE QAP
Center of Excellence in Optimization and Systems Engineering at Åbo Akademi University
Formulations 8 | 16
Koopmans Beckmann form
N¼i=1
N¼j=1
N¼k=1
N¼l=1
aijbkl · xikxjl
N¼i=1
xij = 1, j = 1, ...,N ;
N¼j=1
xij = 1, i = 1, ...,N ;
xij ∈ {0,1}, i , j = 1, ...,N ;
I This formulation has N2(N −1)2 bilinear terms.
THE QAP
Center of Excellence in Optimization and Systems Engineering at Åbo Akademi University
Formulations 8 | 16
Koopmans Beckmann form
N¼i=1
N¼j=1
N¼k=1
N¼l=1
aijbkl · xikxjl
N¼i=1
xij = 1, j = 1, ...,N ;
N¼j=1
xij = 1, i = 1, ...,N ;
xij ∈ {0,1}, i , j = 1, ...,N ;
I This formulation has N2(N −1)2 bilinear terms.
THE QAP
Center of Excellence in Optimization and Systems Engineering at Åbo Akademi University
Formulations 9 | 16tr(AXBXT ) = tr((A⊗B)yyT )
Q= A⊗B=
a11B · · · a1NB
... · · ·...
aN1B · · · aNNB
and y= vec(X) =
x11...
xNN
Semi Definite Programming Relaxation
minY,y
tr(Q′Y)
s.t. diag(Y) = y[1 yT
y Y
]� 0
I ⊗ is the Kronecker product
I The number of continuous variables in Y is N4 and the number ofbinary variables in y is N2
THE QAP
Center of Excellence in Optimization and Systems Engineering at Åbo Akademi University
Formulations 10 | 16
minXA ·BX
minN¼i=1
N¼j=1
a ′ijb′ij
a ′ij =n¼
k=1
akjxik ∀i , j
b ′ij =n¼
k=1
bik xkj ∀i , j
A=
0 3 5 9 63 0 2 6 95 2 0 8 109 6 8 0 26 9 10 2 0
B=
0 4 3 7 74 0 4 10 4
3 4 0 2 37 10 2 0 47 4 3 4 0
THE QAP
Center of Excellence in Optimization and Systems Engineering at Åbo Akademi University
Formulations 10 | 16
minXA ·BX
minN¼i=1
N¼j=1
a ′ijb′ij
a ′ij =n¼
k=1
akjxik ∀i , j
b ′ij =n¼
k=1
bik xkj ∀i , j
A=
0 3 5 9 63 0 2 6 95 2 0 8 109 6 8 0 26 9 10 2 0
B=
0 4 3 7 74 0 4 10 4
3 4 0 2 37 10 2 0 47 4 3 4 0
a ′23 = 5x21 +2x22 +0x23 +8x24 +10x25
b ′23 = 4x13 +0x23 +4x33 +10x43 +4x53THE QAP
Center of Excellence in Optimization and Systems Engineering at Åbo Akademi University
Formulations 11 | 16
Discrete Linear Reformulation (DLR)
minn¼
i=1
n¼j=1
Mi¼m=1
Bmi zmij
zmij ≤ Aj¼
k∈Kmi
xkj m = 1, ...,Mi
Mi¼m=1
zmij = a′ij
∀i , j
Example for one bilinear term a′23b′23
a′23 = 5x21 +2x22 +0x23 +8x24 +10x25
b ′23 = 4x13 +0x23 +4x33 +10x43 +4x53
x13 + x23 + x33 + x43 + x53 = 1
x21 + x22 + x23 + x24 + x25 = 1
THE QAP
Center of Excellence in Optimization and Systems Engineering at Åbo Akademi University
Formulations 11 | 16
Discrete Linear Reformulation (DLR)
minn¼
i=1
n¼j=1
Mi¼m=1
Bmi zmij
zmij ≤ Aj¼
k∈Kmi
xkj m = 1, ...,Mi
Mi¼m=1
zmij = a′ij
∀i , j
Example for one bilinear term a′23b′23
a′23 = 5x21 +2x22 +0x23 +8x24 +10x25
b ′23 = 4x13 +0x23 +4x33 +10x43 +4x53
x13 + x23 + x33 + x43 + x53 = 1
x21 + x22 + x23 + x24 + x25 = 1
4z123 +10z2
23
z123 ≤ 10(x13 + x33 + x53)
z123 + z2
23 = a ′23
THE QAP
Center of Excellence in Optimization and Systems Engineering at Åbo Akademi University
Formulations 12 | 16
02
46
810
02
46
810
0
20
40
60
80
100
02
46
810
02
46
810
0
20
40
60
80
100
Figure 1: Bilinear term a ′23b′23 discretized in b ′23 (to the left) and in a ′23 (to the right)
I The size of the MILP problem is dependent on the number ofunique elements per row.
I Tightness of the MILP problem is dependent on the differencesbetween the elements in each row.
THE QAP
Center of Excellence in Optimization and Systems Engineering at Åbo Akademi University
Formulations 13 | 16
I The size of the DLR is dependent on the number of uniqueelements per row.
I Tightness of the DLR problem is dependent on the differencesbetween the elements in each row.
I A can be modified to any matrix A, where aij + aji = aij +aji .
A=
0 1 2 2 3 4 4 51 0 1 1 2 3 3 42 1 0 2 1 2 2 32 1 2 0 1 2 2 33 2 1 1 0 1 1 24 3 2 2 1 0 2 34 3 2 2 1 2 0 15 4 3 3 2 3 1 0
A=
0 2 2 2 6 6 6 60 0 0 0 4 4 4 42 2 0 2 2 2 2 22 2 2 0 2 2 2 20 0 0 0 0 0 0 02 2 2 2 2 0 2 22 2 2 2 2 2 0 24 4 4 4 4 4 0 0
THE QAP
Center of Excellence in Optimization and Systems Engineering at Åbo Akademi University
Formulations 13 | 16
I The size of the DLR is dependent on the number of uniqueelements per row.
I Tightness of the DLR problem is dependent on the differencesbetween the elements in each row.
I A can be modified to any matrix A, where aij + aji = aij +aji .
A=
0 1 2 2 3 4 4 51 0 1 1 2 3 3 42 1 0 2 1 2 2 32 1 2 0 1 2 2 33 2 1 1 0 1 1 24 3 2 2 1 0 2 34 3 2 2 1 2 0 15 4 3 3 2 3 1 0
A=
0 2 2 2 6 6 6 60 0 0 0 4 4 4 42 2 0 2 2 2 2 22 2 2 0 2 2 2 20 0 0 0 0 0 0 02 2 2 2 2 0 2 22 2 2 2 2 2 0 24 4 4 4 4 4 0 0
THE QAP
Center of Excellence in Optimization and Systems Engineering at Åbo Akademi University
Formulations 14 | 16
Instance Size BKS old LB DLR Time(minutes)esc32a 32 130 103 130 1964esc32b 32 168 132 168 3500esc32c 32 642 616 642 254esc32d 32 200 191 200 10esc64a 64 116 98 116 48
Table 1: Solution times when solving the instances esc32a, esc32b, esc32c, esc32d
and esc64a from the QAPLIB to global optimality
I Previously unsolved instances presented in 1990.
I Nug30 (N = 30) solved in 2001 (in 7 days) using 1000computers in parallel.
THE QAP
Center of Excellence in Optimization and Systems Engineering at Åbo Akademi University
15 | 16
A few references
S.A. de Carvalho Jr. and S. Rahmann.Microarray layout as a quadratic assignment problem.Proceedings of the German Conference on Bioinformatics (GCB), P-83 of Lecture Notes inInformatic:11–20, 2006.
B. Eschermann and H.-J. Wunderlich.Optimized synthesis of self-testable finite state machines.In Fault-Tolerant Computing, 1990. FTCS-20. Digest of Papers., 20th InternationalSymposium, pages 390 –397, jun 1990.
Peter Hahn and Miguel Anjos.Qaplib - a quadratic assignment problem library online.University of Pennsylvania, School of Engineering and Applied Science, 2002.
T.C. Koopmans and M.J. Beckmann.Assignment problems and location of economic activities.Econometrica, 25:53–76, 1957.
Axel Nyberg and Tapio Westerlund.A new exact discrete linear reformulation of the quadratic assignment problem.European Journal of Operational Research, 220(2):314 – 319, 2012.
THE QAP
Center of Excellence in Optimization and Systems Engineering at Åbo Akademi University