, 2012 Vol. 41, No.

1, 57

Application of Multistart Tabu Search to the

MAXCUT Problem
Vardenis Pavardenis
Kaunas University of Technology, Department of Multimedia Engineering
Studentu st. 50-400a, LT-3031 Kaunas, Lithuania
E-mail: itc@ktu.lt

Vardene Pavardene
Vytautas Magnus University, Department of Applied Informatics
Vileikos str. 8, LT-44404 Kaunas,Lithuania

Abstract. In this paper, we investigate two multistart tabu search implementations for the MAX - CUT problem: an
algorithm based on application of a steepeset ascent heuristic to specially constructed subproblems and the classical random restart method. Computational results on three seta of standard test problems indicate that the first of these techniques
outperforms the second one and is very competitive when compared to other heuristic algorithms.
Key words: combinatorial optimization; heuristics; tabu search; MAXCUT.

1. Introduction

3. Theorems, Definitions, Proof

Given an undirected graph G = (V, E) with vertex set V = 1, . . . , n, edge set E V V and
weights wij associated with the edges (i, j) E, the
MAX - CUT problem asks for a subset of vertices V1
such that the weight of
Pthe cut (V1 , V2 = V \ V1 )
given by w(V1 , V2 ) = iV1 ,jV2 wi j is maximized.
Introducing binary variables i, j V , the problem can be stated as follows

It is very easy to set theorems, lemmas, definitions, examples and proofs. Really, it is very easy
to set theorems, lemmas, definitions, examples and
proofs.

maximize F (x) =

wi j(xi xj )2

xi {0, 1}, i V.

Proof. It is very easy to prove.

Lemma 1. It is very easy to set lemmas.
Definition 1. It is very easy to set definitions.

(1)

i,jE

subject to

Theorem 1. It is very easy to set theorems.

(2)

The MAX - CUT problem is of considerable

practical significance. It has a large number of applications, the most known of which are found in design
automation [?, ?] and statistical physics [?].
2. Algorithms
In this section we briefly describe two multistart
tabu search algorithms adapted for solving the MAX
- CUT problem.
The first algorithm generates new staring points
by fixing values of some variables at 0 and then applying a steepeset ascent procedure to the projection of
the problem constructed by removing the fixed variables.

Remark 1. It is very easy to set remarks.It is very

easy to set theorems, lemmas, definitions, examples
and proofs. Really, it is very easy to set theorems,
lemmas, definitions, examples and proofs.
Corollary 1. It is very easy to set corollaries.
Proposition 1. It is very easy to set theorems, lemmas, definitions, examples and proofs. Really, it is
very easy to set theorems, lemmas, definitions, examples and proofs.
The last experiment was conducted on a set
of test problems taken from the DIMACS library
of semidefinite - quadratic - linear programs. This
set contains four instances of MAX - CUT, called
the torns problems, which originated from the Ising
model of spin glasses in physics. We run MST and
RRT on each instance 5 times for 1800 seconds in the
case of pm3-8-50 and g3-8 and for 7200 seconds in
the case of pm3-15-50 and g3-15. The results are reported in Table ??.

V.Pavardenis, V.Pavardene

Fig. 1. Graphical interpretation of the quadratic assignment problem

Table 1.

Solutions found by different techniques for the torns problems

Problem
pm3-8-50
pm3-15-50
g3-8
g3-15

n
512
3375
512
3375

circuit
454
2964
41684814
281029888

Fig. 2. A curve

SA
458
900
892
898

RRT
458
2930
40043061
251918092

MST
458
3000
41684814
283206561

torial Optimization, Wiley, Chichester,1997, pp. 121

136.

4. Experimental results
The main purpose of experimentation was to investigate the capabilities of tabu search in solving instances of the MAX - CUT problem and to compare
the obtained results with those reported in the literature.
5. Conclusions
In this paper we presented two multistart tabu
search implementations for the MAX - CUT problem. The algorithm can quickly find solutions that are
competitive with those found by most successful algorithms described in the literature. For 6 benchmark
graphs the solutions of weight larger than the best
known value were produced.
References
[1] F. Barahona, M. Grtschel, M. Jnger, G. Reinelt.
An application of combinatorial optimization to statistical physics and circuit lauout design. Operation Research, 1988, 36, 493 - 513.
[2] P. Festa, M. G. C. Resende. CIRCUIT + PR: rank-2
heuristic with path - relinking for MAX - CUT. Extended abstracts of the First Metaheuristics International Conference MIC2003, Kyoto, Japan, 2003.
[3] G. Pataki, S. Schmieta. The DIMACS library of
semidefinite - quadratic - linear programs. Technical
report, Computational Optimization Research Center,
Columbia University, 2002.
[4] G.C.Armour and E.S.Buffa. A heuristic algorithm
and simulation approach to relative location of facilities. Management Science,1963,9,294304.
[5] Z. Michalewicz, D.B. Fogel. How to Solve IT: Modern
Heuristics. Springer, BerlinHeidelberg, 2000.
[6] A. Hertz, E. Taillard, D. de Werra. Tabu search. In: E.
Arts and J.K. Lenstra (eds.), Local Search in Combina-

Received May 2011

Application of Multistart Tabu Search to the MAXCUT Problem

Patobulintas hibridinis optimizavimo algoritmas

kvadratinio paskirstymo udaviniui
V. Pavardenis, V. Pavardene
iame straipsnyje pasiulytas patobulintas hibridinis
euristinis optimizavimo algoritmas gerai inomam, sudetingam kombinatorinio optimizavimo udaviniui, butent, kvadratinio paskirstymo (KP) udaviniui. Tai pagerinta autoriaus ankstesnio hibridinio algoritmo versija. Naujasis algoritmas pasiymi tuo, jog c ia ipletota efektyviu euristiku
atkaitinimo modeliavimo (AM) (angl. simulated annealing) ir tabu paiekos (TP) (angl. tabu search) hibridizacijos ideja. Hibridizacija remiasi vadinamaja iteracine schema: TP algoritmas panaudojamas kaip postanalizes procedura AM algoritmo gautajam sprendiniui,
savo ruotu, AM algoritmas taikomas sprendiniu sekai,
gautai sprendiniu diversifikavimo/generavimo keliu. Svarbi
pasiulyto algoritmo savybe yra ir ta, kad jame realizuotas vadinamasis altojo restarto principas (mechanizmas),
kurio paskirtis padeti ivengti galimu paiekos stagnacijos situaciju. Naujasis algoritmas ibandytas su KP
udavinio testiniais pavyzdiais (duomenimis) i testiniu
pavyzdiu bibliotekos QAPLIB. Gauti eksperimentu
rezultatai liudija, jog nagrinetiems KP udavinio testiniams pavyzdiams siulomas algoritmas yra pranaesnis u
ankstesnius atkaitinimo modeliavimo ir tabu.