Tabu
Search and Simulated Annealing for Solving
Large Quadratic Assignment
Problem
by Mohamed
Saifullah Hussin and Thomas
Stützle
July 2010
Paper Abstract
Performance comparisons between
algorithms have a long tradition in metaheuristic research since the
very initial
research efforts. An early example, are comparisons between Tabu
Search (TS) and Simulated Annealing (SA)
algorithms for tackling the Quadratic Assignment Problem (QAP). The
result of these existing comparison is, even
for only two types of algorithms, to a certain extent inconclusive.
While earlier comparisons of SA and TS algorithms
focused on rather small-sized instances, here we examine specifically
the dependence of the relative performance
between SA and TS in dependence of instance size. In particular, our
experimental results show that the assertion
whether one algorithm is better than the other can depend strongly on
the QAP instance size.
Keywords:
Tabu Search, Simulated Annealing, Quadratic Assignment Problem
Benchmark Instances
Benchmark instances that were generated for this article can be
downloaded below. The format of these instances:
n
b
A
B
where n is the
instance size, b is the best known solution (we
put a dummy value 1000). A is
the distance matrix,
and
B is the flow matrix. Instances are grouped as "*.tar.gz"
files,
where each file contains 30 instances. The files
are named using this format:
in-sp
where i is the instance
class, n is the instance
size, and sp is the sparsity.
Euclidean Structured Instances
es20-0.00
es20-0.25
es20-0.67
es20-0.90
es20-0.97
es30-0.00
es30-0.25
es30-0.67
es30-0.90
es30-0.97
es40-0.00
es40-0.25
es40-0.67
es40-0.90
es40-0.97
es50-0.00
es50-0.25
es50-0.67
es50-0.90
es50-0.97
es60-0.00
es60-0.25
es60-0.67
es60-0.90
es60-0.97
es80-0.00
es80-0.25
es80-0.67
es80-0.90
es80-0.97
es100-0.00
es100-0.25
es100-0.67
es100-0.90
es100-0.97
es200-0.00
es200-0.25
es200-0.67
es200-0.90
es200-0.97
es300-0.00
es300-0.25
es300-0.67
es300-0.90
es300-0.97
es500-0.00
es500-0.25
es500-0.67
es500-0.90
es500-0.97
Random Random Instances
rr20-0.00
rr30-0.00
rr40-0.00
rr50-0.00
rr60-0.00
rr80-0.00
rr100-0.00
rr200-0.00
rr300-0.00
rr500-0.00
Grid Random Instances
gr20-0.00
gr30-0.00
gr40-0.00
gr50-0.00
gr60-0.00
gr80-0.00
gr100-0.00
gr200-0.00
gr300-0.00
gr500-0.00
Instance Characteristics
Instance size
dd
fd
sp
ES.20.0.00 20
45.39
168.18 0.07
ES.20.0.25 20
45.89
195.29 0.27
ES.20.0.67 20
45.13
287.90 0.63
ES.20.0.90 20
45.08
556.38 0.90
ES.20.0.97 20
45.16
849.93 0.97
ES.30.0.00 30
45.78
169.25 0.06
ES.30.0.25 30
45.28
195.10 0.24
ES.30.0.67 30
46.02
293.58 0.64
ES.30.0.90 30
45.79
547.48 0.89
ES.30.0.97 30
45.95 1084.19
0.97
ES.40.0.00 40
46.05
171.98 0.05
ES.40.0.25 40
45.82
195.67 0.24
ES.40.0.67 40
45.86
296.39 0.62
ES.40.0.90 40
46.45
563.55 0.90
ES.40.0.97 40
45.96 1061.96
0.97
ES.50.0.00 50
46.31
169.39 0.04
ES.50.0.25 50
46.43
198.36 0.25
ES.50.0.67 50
46.46
304.57 0.63
ES.50.0.90 50
46.49
583.01 0.89
ES.50.0.97 50
46.73 1073.21
0.97
ES.60.0.00 60
46.46
169.68 0.04
ES.60.0.25 60
46.55
198.41 0.24
ES.60.0.67 60
46.63
299.98 0.62
ES.60.0.90 60
46.48
564.96 0.89
ES.60.0.97 60
46.43 1068.70
0.97
ES.80.0.00 80
46.61
171.25 0.04
ES.80.0.25 80
46.78
199.99 0.24
ES.80.0.67 80
46.90
299.31 0.62
ES.80.0.90 80
46.42
573.57 0.89
ES.80.0.97 80
46.60 1070.56
0.97
ES.100.0.00 100 46.63
171.96
0.03
ES.100.0.25 100 46.77
198.57
0.23
ES.100.0.67 100 46.82
300.68
0.62
ES.100.0.90 100 46.67
583.77
0.89
ES.100.0.97 100 46.86
1097.32 0.97
ES.200.0.00 200 46.88
172.01
0.03
ES.200.0.25 200 46.84
199.97
0.23
ES.200.0.67 200 47.00
302.97
0.62
ES.200.0.90 200 46.87
584.80
0.89
ES.200.0.97 200 47.05
1093.56 0.97
ES.300.0.00 300 47.02
171.86
0.03
ES.300.0.25 300 46.99
200.06
0.23
ES.300.0.67 300 46.97
303.16
0.62
ES.300.0.90 300 46.94
585.41
0.89
ES.300.0.97 300 47.02
1100.61 0.97
ES.500.0.00 500 47.01
172.09
0.03
ES.500.0.25 500 47.09
200.95
0.23
ES.500.0.67 500 47.06
303.43
0.62
ES.500.0.90 500 47.11
584.92
0.89
ES.500.0.97 500 47.04
1101.81 0.97
Table 1: Basic statistics on all ES instances from the benchmark set.
Given are the distance
dominance (dd), the flow dominance (fd), and the sparsity of the flow
matrix (sp).
Instance size
dd
fd
sp
RR20 20
56.61
56.05
0.06
RR30 30
57.27
56.92
0.04
RR40 40
57.47
57.66
0.04
RR50 50
57.83
58.12
0.03
RR60 60
57.57
57.76
0.03
RR80 80
57.80
57.82
0.02
RR100 100
57.81 57.91
0.02
RR200 200
58.05 58.07
0.01
RR300 300
58.03 58.14
0.01
RR500 500
58.16 58.16
0.01
Table 2: Basic statistics on all RR instances from the benchmark set.
Given are the distance
dominance (dd), the flow dominance (fd), and the sparsity of the flow
matrix (sp).
Instance size
dd
fd
sp
GR20 20
46.61
56.05
0.06
GR30 30
47.73
56.92
0.04
GR40 40
48.13
57.66
0.04
GR50 50
48.94
58.12
0.03
GR60 60
49.06
57.76
0.03
GR80 80
50.64
57.82
0.02
GR100 100
49.25 57.91
0.02
GR200 200
52.34 58.07
0.01
GR300 300
55.66 58.14
0.01
GR500 500
50.16 58.16
0.01
Table 3: Basic statistics on all GR instances from the benchmark set.
Given are the distance
dominance (dd), the flow dominance (fd), and the sparsity of the flow
matrix (sp).
Additional Results
Instance Sparsity
RoTS RTS
SAC SAR
20
0.00
0.00
0.00
0.04 0.46
20
0.25
0.00
0.00
0.15
0.79
20
0.67
0.00
0.00
0.09
3.01
20
0.90
0.00
0.07
4.30 17.58
20
0.97
0.02
0.00
2.99 38.62
30
0.00
0.00
0.03
0.12
0.06
30
0.25
0.00
0.05 0.36
0.10
30
0.67
0.00
0.10
0.60
1.15
30
0.90
0.00
0.64
3.93
8.46
30
0.97
0.34
0.00
18.33 37.62
40
0.00
0.00
0.12
0.38
0.08
40
0.25
0.00
0.16
0.22
0.05
40
0.67
0.00
0.13
0.77
0.20
40
0.90
0.00
1.00
3.33
2.43
40
0.97
0.72
0.00
14.99 2.09
50
0.00
0.00
0.19 0.21
0.04
50
0.25
0.00
0.17
0.25
0.03
50
0.67
0.08
0.42
0.52
0.00
50
0.90
1.81
3.84
3.78
0.00
50
0.97
0.78
0.00
8.12
6.84
60
0.00
0.01
0.15
0.15
0.00
60
0.25
0.00
0.19 0.28
0.01
60
0.67
0.26
0.36
0.54
0.00
60
0.90
1.09
4.69
0.85 0.00
60
0.97
2.73
1.42
5.66 0.00
80
0.00
0.00
0.09
0.10
0.04
80
0.25
0.00
0.15
0.09
0.00
80
0.67
0.19
0.95
0.59
0.00
80
0.90
1.14
5.07
1.59
0.00
80
0.97
7.48
8.94
7.48
0.00
100
0.00
0.00
0.07
0.04
0.01
100
0.25
0.00
0.17
0.13
0.01
100
0.67
0.19
0.95
0.59
0.00
100
0.90
1.13
4.39
1.61
0.00
100
0.97
8.64
10.05
2.36
0.00
200
0.00
0.00
0.13 0.04
0.02
200
0.25
0.00
0.13 0.03
0.00
200
0.67
0.12
0.63
0.60
0.00
200
0.90
0.48
7.30
2.16
0.00
200
0.97
10.52
15.00
0.00
3.13
300
0.00
1.52
0.00
1.58
1.53
300
0.25
0.14
0.00
0.28
0.14
300
0.67
0.16
1.19
0.84
0.00
300
0.90
0.65
13.92
4.12
0.00
300
0.97
3.00
3.44
2.45 0.00
500
0.00
4.43
0.00
4.47
4.45
500
0.25
2.24
0.00
2.27
2.23
500
0.67
0.53
0.00
1.25
0.41
500
0.90
1.18
9.63
6.47
0.00
500
0.97
9.49
14.64
6.52
0.00
Table 4: Comparison of RoTS, RTS, SAC, and SAR on instance size 20, 30,
40, 50, 60, 80, 100, 200, 300, and 500.
Entries are distance from best found across 30 instances at 10000n.
The lowest average values are printed in bold.
Instance RoTS
RTS SAC
SAR
RR20 0.00
0.00
0.08
0.33
RR30 0.04
0.00
1.21
1.00
RR40 0.14
0.00
1.31
1.27
RR50 0.17
0.00
1.46
1.36
RR60 0.59
0.00
1.59
1.69
RR80 0.32
0.00
1.06
1.39
RR100 0.45
0.00
0.82
1.47
RR200 0.49
0.00
0.48
0.42
RR300 0.15
0.00
0.23
1.01
RR500 0.06
0.00
0.15
0.82
Table 5: Comparison of RoTS, RTS, SAC, and SAR on RR instances size 20,
30, 40, 50, 60, 80, 100, 200, 300, and 500.
Entries are distance from best found across 30 instances at 10000n.
Statistically significant worse performance is indicated
in italics
font, while the best solutions are in bold font.
Instance
RoTS RTS
SAC SAR
GR20
0.00
0.00
0.00
0.00
GR30
0.00
0.00
0.02
0.01
GR40
0.02
0.00
0.09
0.05
GR50
0.01
0.00
0.06
0.03
GR60
0.01
0.00
0.16
0.03
GR80
0.00
0.00
0.02
0.02
GR100
0.00
0.00
0.01
0.03
GR200
0.00
0.03
0.01
0.02
GR300
0.55
0.00
0.56
0.56
GR500
1.41
0.00
1.40
1.41
Table 6: Comparison of RoTS, RTS, SAC, and SAR on GR instances size 20,
30, 40, 50, 60, 80, 100, 200, 300, and 500.
Entries are distance from best found across 30 instances at 10000n.
Statistically significant worse performance is indicated
in italics
font, while the best solutions are in bold font.