Computers & Operations Research ∎ (∎∎∎∎) ∎∎∎–∎∎∎

Contents lists available at ScienceDirect

Computers & Operations Research

journal homepage: www.elsevier.com/locate/caor

Variable Neighborhood Search for extremal

vertices : The AutoGraphiX-III system
Gilles Caporossi
GERAD & HEC Montréal 3000, chemin de la côte-ste-Catherine H3T 2A7, Canada

ar t ic l e i nf o a b s t r a c t

More than fifteen years after the beginning of the development of AutoGraphiX (AGX), a third version of
Keywords: the software is made available. Since the program was rewritten from scratch, it was the opportunity to
Variable neighbourhood search look forward and consider new avenues. From the user's point of view, the interface is completely
Graph theory changed, which allows the display of multiple information which was not possible in the previous
Extremal graphs versions. However, one of the main improvements is that it is designed to help researchers in the field of
AutoGraphiX complex networks. In these days when increasing research is applied to complex networks (such as
social networks), the use of quantities related to vertices, indicating the centrality (the importance of an
actor in the network measured as a topological indicator) naturally leads researchers toward the
mathematical study of these quantities. This new paradigm implies a complete change in the optimi-
zation algorithm that now natively handles multi objective optimization problems involving vertex-
related measures.
& 2015 Elsevier Ltd. All rights reserved.

1. Introduction As a difference, with previous versions, it is now possible to use

From the very beginning of the development of AGX, the main
principle was to find extremal graphs, which implies the optimi-
zation of quantities that may be computed from the graph
regardless the way vertices are labelled (graph invariants). If AGX-
2 was able to compute vectors and matrices, their usage was
restricted because the optimizer was not able to handle more than
a single value at a time.
The conception of AGX-III is oriented toward the study of tools
for complex networks analysis. This implies to consider quantities
that are associated to each vertex (influence or centrality), not
associated to the whole graph. Note that the software is not build
to study complex existing networks, but to mathematically study
measures that may be used to analyze complex networks, i.e.,
centrality indicators (vertex related measures). For example, it is
not designed to compute measures, such as the betweenness
centrality of vertices for a given graph, but to help mathematically
study betweenness centrality by finding graphs for which it is
extremal, and eventually find bounds related to other measures
(such as transmission) [11,15,18].
E-mail address: gilles.caporossi@gerad.ca
URL: https://www.gerad.ca/Gilles.Caporossi
optimization on one or more vectors at a time instead of just a
single value. One of the windows of AGX-III is the so called prop-
erties explorer which has a tab dedicated to graphs and a tab
dedicated to vertices (see Figs. 2 and 3). In the graph tab, vertex
related values are considered as different variables (one variable
per vertex). On the vertex tab, each observation is a vertex. An
engine based upon the geometric approach to find conjectures
described in [9] is implemented to find tight inequalities valid for
all observations (graphs or vertices depending on the choice of the
researcher). Of course, graph values may also be displayed on the
vertex tab, the same value being assigned to every vertex of the
graph. It is also possible to force a vector not to be considered as
vertex value, but as a graph value, in which case all the values will
be considered as different variables even in vertex mode (which is
useful for spectral graph theory). In such a case, even on the vertex
tab, all the different values will be considered as different
variables.
From a practical point of view, the problem definition is now
achieved through the white board (represented in Fig. 1) which
provides a graphical environment where operators are repre-
sented by nodes and argument relations as edges. Double-clicking
on a node opens a separate window by which it is possible to
change some attributes of the operator or display its content (a
value, a vector or a matrix, depending on the context).

http://dx.doi.org/10.1016/j.cor.2015.12.009
0305-0548/& 2015 Elsevier Ltd. All rights reserved.

Please cite this article as: Caporossi G. Variable Neighborhood Search for extremal vertices : The AutoGraphiX-III system. Computers
and Operations Research (2016), http://dx.doi.org/10.1016/j.cor.2015.12.009i
2 G. Caporossi / Computers & Operations Research ∎ (∎∎∎∎) ∎∎∎–∎∎∎

Fig. 1. The white board, the problem definition window in AGX-III.

Let us illustrate with an example. The energy of a graph is
defined as
X
n extremal graphs by the mean of the Variable Neighborhood Search
EðGÞ ¼ j λi ðGÞj ;
i¼1
where λi ðGÞ is the ith eigenvalue of the adjacency matrix of G. The
problem represented in Fig. 1 consists in maximizing and mini-
mizing the energy of the graph and its complement. Calculus
nodes are computed sequentially from the top to the bottom. Each
calculus node is connected to its argument by the top. Rectangular
nodes are computed directly from the graph and do not require
argument. Ellipses represent operators that may either need
arguments or be arguments. Finally, diamonds, arrows or double
arrows correspond to final state calculus nodes that may either
link to the display of the graph (highlights or values on the nodes
or edges) or to the optimizer (indicating a function to be mini-
mized, maximized or both sequentially). The white circle on the
top right rectangle (named ADJACENCY) indicates that this matrix
is computed from the complement of the graph under study. On
that example, the left sequence of calculus nodes reads as follows
(from the top to the bottom): ADJACENCY – EIGENVALUES – ABS –
SUM – MAXMIN. In the algebraic notation, this would mean
MaxMinðsumðabsðeigenvaluesðadjacencyÞÞÞÞ. The sequence on the
right is the same except that the adjacency matrix is computed for
the complement of the graph.
The energy of the graph and its complement are both max-
imized and minimized, which means that AGX-III will run simul-
taneously or sequentially (depending on the number of threads)
4 optimization problems, each of them involving a combination of
2 objectives.
This paper focuses on the optimization module and the cap-
abilities it provides. The reader could refer to the AGX-III webpage
[12] for more information on the system and its use. The following
section recalls background information about the search for
(VNS) metaheuristic [14,17]. Then, the problem of comparing
invariants is discussed before the extension of multi objective
optimization to vertex related values (and the difficulties it
involves). Some algorithms are proposed, and then compared
before the conclusion.
2. Variable neighborhood search for extremal graphs
Let G ¼ ðV; EÞ be a connected undirected graph without loops or
multiple edges. Let Gn be the set of all graphs on j V j ¼ n vertices,
and let i : Gn -R be a graph invariant, a function whose value does
not rely on the labelling of the vertices. Some of the most common
graph invariants are the number of edges, diameter, chromatic
number, independence number, etc. Graph invariants provide a
powerful analytical tool for investigation of structures of graphs.
They, combined in convenient algebraic relations, contain global
and general information about a graph.
A way to study some particular graph invariant from the
mathematical point of view is to describe it or bound it by a
function of other invariants, which could be achieved through the
search for extremal graphs and their description.
Without lose of generality, the problem of finding extremal
graphs could be described as the following combinatorial opti-
mization problem:
max iðGÞ ð1Þ
subject to

Please cite this article as: Caporossi G. Variable Neighborhood Search for extremal vertices : The AutoGraphiX-III system. Computers
and Operations Research (2016), http://dx.doi.org/10.1016/j.cor.2015.12.009i
 G. Caporossi / Computers & Operations Research ∎ (∎∎∎∎) ∎∎∎–∎∎∎ 3

ik ðGÞ r 0 ð2Þ The Variable Neighborhood Search implementation used in

il ðGÞ ¼ 0
G A Gn :
Where ik(G) and il(G) are sets of invariants corresponding to sets of
possible additional constraints. Note Gnr the set of all the realizable
solutions of the problem (1)–(4), i.e. graphs G A Gn that respect
these constraints. An alternative definition of the problem is:
maxiðGÞ
subject to
GA Gnr :
The search for extremal graphs may be used to various pur-
pose such as finding conjectures, refuting or strengthening con-
jectures, finding graphs respecting constraints, among others.
Based upon the Variable Neighborhood Search (VNS) [14,17],
AutoGraphiX [1,8] was developped since 1997 and was used to
AGX is described by Algorithm 3, where the stopping criterion
ð3Þ could either be the maximum number of evaluations of the
objective function or the maximum time. In those algorithms, the
ð4Þ
transformation to use t and the objective function i are implicit as
they are chosen prior to the application of the algorithm and
remains unchanged during the whole optimization process, as
opposed to the current graph G and the perturbation magnitude k
which are therefore presented as parameters of the local search
and perturbation algorithms.
Algorithm 3. Variable Neighborhood Search in AGX.
1: Let kmax, the maximum perturbation magnitude.
2: Let G be an initial graph
n
▹
3: Set G ’G the best known solution
Usually a random graph

find extremal graphs, and to help the researcher in graph theory

by providing counter-examples to conjectures or by finding new 4: Set k’1
conjectures (relations that hold for all graphs or a family of 5: repeat
them), etc. 6: G0 ’PERTURBðG; kÞ
0
The basic principle of the variable neighborhood search is to 7: G0 ’LSðG0 Þ
successively apply local searches and perturbations of various 8: if iðG″ Þ 4iðGn Þ then
magnitude. In AGX, both the local search and the perturbations 9: Set Gn ’G″
involve transformations of the graph. 10: Set k’1
Let the neighborhood t(G) of G be the set of graphs that could 11: else
be obtained from G by applying the transformation t. Examples of 12: Set k’ðk þ1Þmodðkmax Þ
transformations could be to add an edge, to remove an edge, to 13: until the stopping criterion is reached
apply a 2-opt, etc. The transformations used in AGX are mainly
based upon the replacement of a subgraph on 4 vertices by
another one. Often, t(G) is too large to be completely explored.
Therefore, some heuristics are used to consider the most promis- 3. Comparing invariants by multi objective optimization
ing graphs G0 A tðGÞ. The choice of the graphs to be considered in t
(G) is achieved by a learning algorithm which uses probabilities A convenient and classical way to study an invariant is to
that are updated during the optimization process. The basics of compare it with other invariants and find relations with them.
this learning algorithm are provided in [1], but a much improved Automatic search for those relations may be achieved by para-
version is described in [10]. metric optimization in which some invariants are bounded by
The VNS algorithm involves a local search described by Algo- constraints while one of them is optimized. A typical way to use
rithm 1, and a perturbation described in Algorithm 2.
AGX was thus to search for extremal graphs with regard to a given
Algorithm 1. Local search in AGX. invariant (or a function of invariants, which is also an invariant),
when varying the number of vertices (order) and the number of
1: Let G be an initial graph. edges (size). This principle was extensively applied by Aouchiche
2: function LSG

▹
and Hansen [2,3,5], who proposed to study arbitrary functions of
3: Note i(G) the value associated to G. To be two invariants by applying an operator between them (for exam-
maximized ple the summation, the substraction, the product or the ratio). This
4: repeat approach implicitly assigns a coefficient 1 to each invariant, and is
5: Let G0 A tðGÞ \ Gnr so that iðG0 Þ 4 iðGÞ therefore sensitive to the scaling of the invariants [4]. A more
6: if G0 exists then, general approach was proposed by Mélot [16] with the system
7: G’G0 . GraPHedron in which all possible graphs of limited order are
8: until G0 doesn't exist.
to t
▹ G is a local optimum according considered and studied as well as the values for the invariants
under study. Based on the values of pairs of invariants for all
graphs, GraPHedron thus provides a better description of the
9: return G
relations between the considered invariants than the one pro-
posed by Aouchiche and Hansen.
It is natural to combine both approaches and replace the heavy
Algorithm 2. Perturbation in AGX.
task of the enumeration by multi objective optimization, which
1: Let G be an initial graph. leads to a description of the relations among invariants similar to
2: Let k be a positive integer. that of GraPHedron, but based upon graphs of larger order that are
3: function PERTURB(G, k) heuristically obtained. AutoGraphiX-III implements this approach
4: for step ¼ 1-k (and the capability to study vertex-related measures).
5: Randomly choose G0 A tðGÞ. In the case of multi objective optimization, the objective
6: G’G0 function is IðGÞ A Rq , a vector whose components are graph invar-
7: return G iants ij(G), j ¼ 1; …; q.

Please cite this article as: Caporossi G. Variable Neighborhood Search for extremal vertices : The AutoGraphiX-III system. Computers
and Operations Research (2016), http://dx.doi.org/10.1016/j.cor.2015.12.009i
4 G. Caporossi / Computers & Operations Research ∎ (∎∎∎∎) ∎∎∎–∎∎∎

Let us now consider, without lost of generality, the multi 3.2. Example
objective optimization problem that consists in maximizing all the
components of I(G). This problem could be described as: The transmission tv of the vertex v is defined as the sum of
max IðGÞ ð5Þ geodesic (the shortest) distances from v to all other vertices of G.
The betweenness buv of the edge (u,v) is the number of shortest
subject to paths between pairs of vertices of G that involve the edge (u,v)
GA Gnr : ð6Þ [13]. From the betweenness can be computed the adjusted
betweenness centrality cv [11] closely related to Freeman's
Consider two graphs G1 ; G2 A Gnr and the vector I ¼ IðG1 Þ  IðG2 Þ. betweenness centrality as follows:
X
 If I A Rq þ and I a 0 (I has some non 0 components), then G1 cv ¼ buv ð10Þ
ðu;vÞ A E
dominates G2, and we write IðG1 Þ 4 IðG2 Þ.
 If I has both strictly positive and strictly negative components, Studying the relation between tv and cv, in terms of invariants
or I ¼0 (all components of I are 0), then G1 and G2, are not could be achieved by finding relations between maxðt v Þ and
comparable, and we write IðG1 Þ  IðG2 Þ. maxðcv Þ. Using AGX with multi objective optimization (with order
n¼ 10) provides a tentative pareto front as described on Fig. 2,
The set Gn  Gnr of Pareto optimal solutions for the problem (5)– where 17 r maxðt v Þ r 45 and 49 rmaxðcv Þ r 81.
(6) is such that: On n¼ 10 vertices, the automated conjecture search engine of
AGX provides the relation:
1. 8 G2 A Gnr ⧹Gn (G1 A Gn so that IðG1 Þ 4 IðG2 Þ and 2 maxðt v Þ þ maxðcv Þ r 139;
2. 8 G1 ; G2 A Gn , IðG1 Þ  IðG2 Þ.
and indicates it is respected as a equality for 9 graphs. When n ¼11,
3.1. Algorithmic issues the relation becomes:
2maxðt v Þ þ maxðcv Þ r170:
In the case of multi objective optimization, the problem is not
In both cases, the relation is satisfied as an equality by the path, for
to find an optimal solution, but a set of Pareto optimal solutions Gn .
which:
Various approaches may be considered for this problem. The first
one, which does not require any special implementation as long as nðn 1Þ
max t v ¼
the software may handle constraints, consists in optimizing an jnk2 lnm jnk  lnm 
objective function, while the others are expressed as constraints in max cv ¼  þ þ1  1 ;
which they are bounded by parameters. This approach was used 2 2 2 2
with earlier versions of AGX (for example [7]). which leads to Conjecture 1.
Using this technique, if the objective function to be optimized is
Conjecture 1. Let G ¼ ðV; EÞ be a connected simple graph. Let tv and
ip(G), with 1 rp rq, the problem (5)–(6) becomes:
cv respectively denote the transmission and adjusted betweenness
max ip ðGÞ ð7Þ centrality of the vertex v, then the inequality
 2 
subject to 3n
2 maxðt v Þ þ maxðcv Þ r  n 1 ð11Þ
is Z as 8 s ¼ 1…q; s a p ð8Þ 2
holds. Furthermore, the bound is tight if and only if G is a spider (a
GA Gnr ; ð9Þ tree with only one vertex of degree larger than 2).
where as is a bound for the sth objective, used as a parameter.
Another approach to handle the multi objective optimization
issue is to adapt the VNS algorithm, as well as the local search. In 4. Multiobjective optimization for vertex related values
Algorithm 1, line 3 must be replaced by “Note i(G) the value
associated to G” and line 5 by “Let G0 A tðGÞ \ Gnr so that IðG0 Þ 4 IðGÞ ”. In some situations a global description of a graph (by the mean
Algorithm 3 should be replaced by Algorithm 4. of invariants) is not useful. For example, in the study of social
networks one could focus only on individual vertices, not the
Algorithm 4. VNS for multi objective optimization in AGX. whole graph. Instead of using graph invariants, one then would
1: Define kmax the maximum perturbation magnitude use vertex-related quantities, such as the degree d(v) (number of
2: Let G be an initial graph ▹Usually a random graph vertices adjacent to v) or the transmission tv. As it is the case for
3: Set Gn ¼ fGg the set of best known solutions invariants, vertex quantities are often better described in relation
4: Set k’1 with other quantities. The problem of multi objective optimization
5: repeat could be extended to values computed for each vertex.
6: Randomly chose G A Gn
4.1. Algorithmic issues
7: G0 ’PERTURBðG; kÞ
0
8: G0 ’LSðG0 Þ
In terms of optimization, the solution of such a problem is a set of
9: if G″ is not dominated by any graph G A Gn then vertices associated to their corresponding graph. When working on
10: Set Gn ’fG″ g [ Gn individual vertices, the objective function, i.e. the vector I(G) is no
11: Remove all dominated solutions from Gn more appropriate, but should be replaced by a set of vectors, I(v),
12: Set k’1 8 v A G. The approach that consists in solving the multi objective
13: else optimization problem by a parametric optimization of a single
14: Set k’ðk þ1Þmodðkmax Þ objective function cannot be applied anymore. If some vertices do
15: until the stopping criterion is reached not respect the constraint added for parameterization, the whole

Please cite this article as: Caporossi G. Variable Neighborhood Search for extremal vertices : The AutoGraphiX-III system. Computers
and Operations Research (2016), http://dx.doi.org/10.1016/j.cor.2015.12.009i
 G. Caporossi / Computers & Operations Research ∎ (∎∎∎∎) ∎∎∎–∎∎∎
Fig. 2. Representation of the Pareto front, maxðcv Þ on the x-axis and maxðt v Þ on the y-axis for n¼ 10.
graph is avoided, even if some other vertices of that same graph
could be pareto optimal. The optimization would then be achieved
over a subset of the realizable solutions, which is clearly suboptimal.
Therefore, it is mandatory to develop a proper strategy in order to
solve this problem. An approach to handle this problem is to modify
the definition of domination (used in Algorithm 1, line 5) and
eventually adapt the local search algorithm.
Definition 1. Domination in the strong sense: We say that the
graph G2 dominates the graph G1 if (v A VðG2 Þ so that IðvÞ 4 IðuÞ
8 u A V ðG1 Þ.
This definition is very restrictive. In the context of a local search, it
leads to poor results because it is very easy to find a graph G which
is not dominated by any G0 A tðGÞ. In that case the local search stops
even if G is not a good solution.
Definition 2. Domination in the weak sense: We say that G2
dominates G1 if ( v A VðG2 Þ such that ∄u A VðG1 Þ so that IðuÞ 4 IðvÞ.
According to this definition, graph G2 dominates graph G1 if there is a
vertex in G2 that is not dominated by any vertex of G1. However, it is
possible that a vertex v of VðG2 Þ is not dominated by any vertex of
VðG1 Þ and there also exists a vertex u of VðG1 Þ that is not dominated
by any vertex of VðG2 Þ. In such case u and v are not comparable.
Implementing this definition in Algorithm 1 would cause it to cycle.
Please cite this article as: Caporossi G. Variable Neighborhood Search for extremal vertices : The AutoGraphiX-III system. Computers
and Operations Research (2016), http://dx.doi.org/10.1016/j.cor.2015.12.009i
5
It seems impossible to extend or modify the definition of dom-
ination of a graph by another and get an efficient local search. If the
values of objective function used in the optimization refer to vertices,
then the individual vertices associated to the corresponding graph
must be considered. Therefore, we cannot conclude that a graph
dominates another, but that it dominates a vertex of a graph. The
corresponding definition could be as follows:
Definition 3. A graph G2 dominates the pair ðG1 ; uÞ if (v A VðG2 Þ so
that IðvÞ 4 IðuÞ.
By using this definition, G2 will implicitly be associated to the
vertex v, thus forming the pair ðG2 ; vÞ. So, if we use this definition
instead of Definition 2, we solve the cycling problem and would
theoretically produce the appropriate results.
None of the proposed definitions is completely appropriate.
The two first would lead to inefficient algorithms from a theore-
tical point of view. Indeed, Algorithm 4 using the Definition 3 is
technically adapted for the optimization in the case of vertex
related values, but it involves the implementation of special data
structures to identify the vertex associated to each solution. Fur-
thermore, the use of such a structure allows only a single supposed
Pareto optimal vertex after each local search, which implies a large
number of local searches in order to identify a complete set of
Pareto optimal solution.
6 G. Caporossi / Computers & Operations Research ∎ (∎∎∎∎) ∎∎∎–∎∎∎

Using a different strategy, an efficient algorithm would not only

involve a single graph, but also a set of graphs, each of which Algorithm 4 must be slightly adapted. The adapted version is
having at least one vertex that is not dominated by any vertex described in Algorithm 6.
found during that local search. As the algorithm is based upon
Algorithm 6. New VNS algorithm in AGX.
vertices and involved a set of graphs, to simplify the notation, we
will write v A G to identify a vertex v A V ðGÞ of a graph G A G. The 1: Define kmax to be the maximum perturbation magnitude
resulting algorithms are Algorithm 5 and Algorithm 6. 2: Let G be an initial graph ▹Usually a random graph
3: Set Gn ¼ fGg the set of best known solutions
Algorithm 5. New local search in AGX.
4: Set k’1
1: function LSPARETO(G) 5: repeat
2: Let G be an initial graph. 6: Randomly chose G A Gn
3: Let G’fGg be the current set of solutions 7: G0 ’PERTURBðG; kÞ
8: G’LSparetoðG0 Þ
4: Note I(v) the value associated to a vertex v A VðGÞ
be maximized
▹ To 9: if ( v A G; ∄u A Gn so that IðuÞ 4 IðvÞ
10: Set Gn ’G [ Gn
5: repeat 11: remove of Gn all graphs whose vertices are all
dominated
6:
7:
for GA G do
0
Let G A tðGÞ \
▹ Test all graphs
so that ( v A V ðG0 Þ; IðvÞ 4 IðuÞ 8 u A G
Gnr
12:
13: else
Set k’1

0 14: Set k’k þ 1modðkmax Þ

8: if G exists then
15: until the stopping criterion is reached
9: G’G [ fG0 g
10: remove of G all graphs whose vertices are all
dominated
4.2. Example

11: 0
until G does not exist.
according to t
▹ G is a set of local optimum
As in the previous example, let us consider transmission and
adjusted centrality. These values are associated to vertices, so, the
12: return G search for extremal graphs should be replaced by a search for

Fig. 3. Representation of the Pareto front, cv on the x-axis and tv on the y-axis.

Please cite this article as: Caporossi G. Variable Neighborhood Search for extremal vertices : The AutoGraphiX-III system. Computers
and Operations Research (2016), http://dx.doi.org/10.1016/j.cor.2015.12.009i
 G. Caporossi / Computers & Operations Research ∎ (∎∎∎∎) ∎∎∎–∎∎∎ 7

extremal vertices, constraining the pair (tv,cv) to be associated to Max i2(G)

the same vertex. The results obtained by AGX-III are represented
on Fig. 3. We note that there are a large number of dominated
solutions (vertices) on that figure. This is because all the vertices
that belong to the same graph as a vertex from the tentative Pareto
front are represented.
In the present case, instead of the relation (11) the relation
2t v þ cv r n2  1 ð12Þ
is obtained. This relation was proved in [15]. We notice that the
bound is much smaller because the quantities t v r maxðt v Þ and cv
r maxðcv Þ must be computed for the same vertex.
In addition to the problems handled by AGX-2, AGX-III may
handle this new class of problems. The method used is to treat
relations among different quantities through multi objective
optimization instead of parameterization. The modelling paradigm
is thus changed as well as the optimization algorithm. In addition
to its new capability, it turns out that AGX-III is more efficient,
even for the problems that may be handled by AGX-2 as well.
Max i1(G)

5. Performance comparison Fig. 4. Representation of the dominated surface with 3 solutions when both
objective functions are maximized.

The only suitable algorithms for vertex related values are

Algorithm 4, in which Definition 3 is implemented, and Algorithm 6. Table 1
The steps in Algorithm 4 implementing Definition 3 for domination Performance comparison between the dominated surface after Algorithm 4 and
Algorithm 6 were performed for each pair of invariants on 100 runs (average,
are the same as those of Algorithm 4 except that the comparison is minimum, maximum).
based upon vertex values (which implies a linear time additional
computation time, a constant time for each vertex for each objective). J(G) J(G) J(G) Kf(G) Kf(G) E(G)
This additional computation time is not significant compared to the vs vs vs vs vs vs
Kf(G) E(G) λ1 ðGÞ E(G) λ1 ðGÞ λ1 ðGÞ
computation of the objective functions.
In order to compare these algorithms, we propose to use them 105.52 80.1348 54.7297 489.368 191.199 136.434
with invariants. The comparison will not be affected, but this Algorithm 4 81.84 50.6016 54.7297 358.47 127.311 75.3986
approach does not require the use of a special data structure to 149.55 90.7557 54.7297 979.147 281.961 147.468
associate a special vertex to each graph. 810.72 116.805 54.7297 2427.87 796.545 176.962
Let us consider the following invariants: Algorithm 6 606.87 57.2514 54.7297 1438.58 755.708 176.795
863.19 118.61 54.7297 2537.91 798.032 176.971
 J(G): the Balaban index [6] is defined as follows: Best found solution 863.195 118.61 54.7297 2538.19 798.032 196.971
Number of 48 20 1 84 83 20
m X 1
JðGÞ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffi; solutions
m n þ 2ðu;vÞ A E t u  t v

where tu, the transmission of the vertex u is the sum of geodesic

distances between u and the other vertices of the graph.
 Kf(G): the Kirchhoff index (or resistance) is defined as follows:
X
n 1
1
Kf ðGÞ ¼ n ; Table 1 describes the results obtained. As an indication, the line
μ
k¼1 k
where μk is the kth eigenvalue of the laplacian matrix of G.
 E(G): the energy is defined as:
X
n the best solution from Algorithm 4, the two exceptions being
EðGÞ ¼ j λi ðGÞj ;
i¼1
where, λi ðGÞ is the ith eigenvalue of the adjacency matrix of G.
 λ1 ðGÞ: the spectral radius is the largest eigenvalue of the adja-
cency matrix of G.
The problems used to compare the algorithms were built by
considering all pairs of invariants, each of them being maximized.
To evaluate the performance of Algorithms 4 and 6, they were both
programmed in AGX-III. To avoid artifacts related to the unsure
measure of the CPU time on modern computers (specially when
multi core computers are used), it was decided that the stopping
criterion would be the number of evaluations of the objective
function (100 000 evaluations), and 100 runs were performed on
each problem for graphs with 10 vertices. Note that the test
problems used involved the maximization of two non negative
invariants, so that the dominated surface (Fig. 4) could be used as a
criterion.
Best found solution provides the best solution that was found after
various trials, even if it was not during the tests. The worst solu-
tion obtained with Algorithm 6 is quite systematically better than
JðGÞ vs λ1 ðGÞ, which is not really a multi objective problem as it
yields only one optimal solution, and the problem JðGÞ vs EðGÞ for
which 20 possible Pareto optimal solutions were found.
Comparing the Pareto fronts obtained, shows that a wide por-
tion of the front is completely missed by Algorithm 4. An expla-
nation could be that it fails to diversify its search, while this task is
made easier by the storage of a variety of solutions in Algorithm 6.
It is clear that Algorithm 6 performs much better than
Algorithm 4 when the number of Pareto optimal solutions
increases. This result was expected as Algorithm 4 cannot generate
more than one solution per local search, which is not the case for
Algorithm 6. The computational effort required for each Pareto
optimal solution is thus much smaller when Algorithm 6 is
applied.

Please cite this article as: Caporossi G. Variable Neighborhood Search for extremal vertices : The AutoGraphiX-III system. Computers
and Operations Research (2016), http://dx.doi.org/10.1016/j.cor.2015.12.009i
8 G. Caporossi / Computers & Operations Research ∎ (∎∎∎∎) ∎∎∎–∎∎∎
6. Conclusion
The main goal of this work was to extend the capabilities of
AGX to handle vertex related values. As stressed in Section 4.1, this
implies the use of a native multi objective optimization algorithm.
Apart from its capability to handle problems involving vertex
related values, one of the main contributions of AGX-III is thus its
optimization module. It turns out that the new proposed algo-
rithm performs significantly better than the natural approach that
was illustrated by Algorithm 4.
A deeper analysis of the results indicates that Algorithm 4 fails
to diversify the search and the solution often remains on the same
portion of the Pareto front. This result is not surprising and a
similar effect was already noticed in AGX when optimizing
invariants that could only take few values, such as diameter or
maximum degree (this difficulty is known as plateau). The algo-
rithm then fails to find any improved solution, because it does not
explore other solutions which are equivalent to the incumbent. In
the present context, this phenomena is the following: when a
graph G is found, Algorithm 4 simply ignores alternative graphs
G0 A tðGÞ that do not dominate G, even if some of them are not
dominated by G either (they are equivalent to the incumbent). In
Algorithm 6, those graphs are stored in the current set of Pareto
optimal solutions Gn , and their neighborhood is likely to be
explored later.
In the present paper, a VNS algorithm for attempting to find
Pareto optimal solution for vertex related values objectives is
proposed. The problem has a special characteristic as many Pareto
optimal solutions (vertices) may be associated to the same graph.
In this case, it was shown that keeping a variety of non dominated
solutions during the local search is better than the natural
approach which consists in performing a local search on a single
solution and eventually add it to the Pareto front because the
classical approach cannot add more than a single new solution to
the Pareto front after each local search.
The new algorithm may lead to a promising approach to handle
the optimization problem related to plateau, as the drawbacks
from Algorithm 4 are being rather well corrected by Algorithm 6.
Therefore, the use of Algorithm 6 can be extended to the classical
optimization in AGX. However, the case of two non comparable
solution is different from the case of two graphs with the same
objective value, as in this last situation, one needs to avoid that
both graphs are isomorphic, otherwise the set Gn could yield a
huge number of isomorphic graphs that would slow down the
optimization. A way to reduce the number of these graphs in Gn is
to check the isomorphism, which would be computationally
intensive. Instead of using an isomorphism test, it is possible to
use a discriminating invariant such as the Balaban index. A graph
would then be added to Gn only if it has the best known value for
Please cite this article as: Caporossi G. Variable Neighborhood Search for extremal vertices : The AutoGraphiX-III system. Computers
and Operations Research (2016), http://dx.doi.org/10.1016/j.cor.2015.12.009i
the objective function and different values for the discriminating
invariant.
Acknowledgments
This work was supported by NSERC (Canada) (Grant no. RGPIN-
2015-04945). The author would like to thank an anonymous
referee for his useful comments.
References
[1] Aouchiche M, Bonnefoy J-M, Fidahoussen A, Caporossi G, Hansen P, Lacheré J,
et al. Variable neighborhood search for extremal graphs. 14. The AutoGraphiX
2 system. In: Liberti L, Maculan N, editors. Global Optimization. From Theory
to Implementation. New-York: Springer Science; 2006.
[2] Aouchiche M, Caporossi G, Hansen P. Variable neighborhood search for
extremal graphs. 20. automated comparison of graph invariants. MATCH
Commun Math Comput Chem 2007;58:365–84.
[3] Aouchiche M, Hansen P. Proximity and remoteness in graphs: Results and
conjectures. Networks 2011;58:95–102.
[4] Aouchiche M, Hansen P. The normalized revised szeged index. MATCH Com-
mun Math Comput Chem 2012;67:369–81.
[5] Aouchiche M, Hansen P, Zheng M. Variable neighborhood search for extremal
graphs. 19. further conjectures and results about the randic index. MATCH
Commun Math Comput Chem 2007;58:83–102.
[6] Balaban AT. Distance connectivity index. Chem Phys Lett 1982;89:399–404.
[7] Caporossi G, Cvetkovic D, Gutman I, Hansen P. Variable neighborhood search
for extremal graphs. 2. Finding graphs with extremal energy. J Chem Inf
Comput Sci 1999;39(6):984–96.
[8] Caporossi G, Hansen P. Variable neighborhood search for extremal graphs. 1.
The AutoGraphiX system. Discret Math 2000;212:29–44.
[9] Caporossi G, Hansen P. Variable neighborhood search for extremal graphs. v.
Three ways to automate finding conjectures. Discret Math 2004;276:81–94.
[10] Caporossi G, Hansen P. A learning optimization algorithm in graph theory.
Versatile search for extremal graphs using a learning algorithm. Lecture Notes
in Computer Science 2012;7219:16–30.
[11] Caporossi G, Paiva M, Vukicević D, Segatto M. Centrality and betweenness:
Vertex and edge decomposition of the wiener index. MATCH Commun Math
Comput Chem 2012;68:293–302.
[12] Caporossi, G. AutoGraphiX web page 〈https://www.gerad.ca/Gilles.Caporossi/
agx〉; last [accessed 25.10.2015].
[13] Freeman LC. A set of measures of centrality based on betweenness. Sociometry
1977;40:35–41.
[14] Hansen P, Mladenović N. Variable neighborhood search: Principles and
applications. Eur J Oper Res 2001;130:449–67.
[15] Majstorović S, Caporossi G. Bounds and relations involving adjusted centrality
of the vertices of a tree. Graphs Comb. 2015;31:2319–34.
[16] Mélot H. Facet defining inequalities among graph invariants: the system
GraPHedron. Discret Appl Math 2008;156:1875–91.
[17] Mladenović N, Hansen P. Variable neighborhood search. Comput Oper Res
1997;24:1097–100.
[18] Vukicević D, Caporossi G. Network descriptors based on betweenness cen-
trality and transmission and their extremal values. Discret Appl Math
2013;161:2678–86.