Self-Organized Data and Image Retrieval as a Consequence of

Inter-Dynamic Synergistic Relationships in Artificial Ant Colonies

Vitorino RAMOS, Fernando MUGE, Pedro PINA
CVRM GeoSystems Centre, Technical Univ. of Lisbon (IST),
Av. Rovisco Pais, Lisbon, 1049-001, Portugal

Abstract. Social insects provide us with a powerful metaphor to create decentralized systems of simple
interacting, and often mobile, agents. The emergent collective intelligence of social insects swarm
intelligence resides not in complex individual abilities but rather in networks of interactions that exist
among individuals and between individuals and their environment. The study of ant colonies behavior and
of their self-organizing capabilities is of interest to knowledge retrieval/ management and decision support
systems sciences, because it provides models of distributed adaptive organization which are useful to solve
difficult optimization, classification, and distributed control problems, among others. In the present work
we overview some models derived from the observation of real ants, emphasizing the role played by
stigmergy as distributed communication paradigm, and we present a novel strategy (ACLUSTER) to tackle
unsupervised data exploratory analysis as well as data retrieval problems. Moreover and according to our
knowledge, this is also the first application of ant systems into digital image retrieval problems.
Nevertheless, the present algorithm could be applied to any type of numeric data.

1. Introduction: Stigmergy and Distributed Awareness

Synergy, from the greek word synergos, broadly defined, refers to combined or co-operative
effects produced by two or more elements (parts or individuals). The definition is often
associated with the quote the whole is greater than the sum of its parts (Aristotle, in
Metaphysics), even if it is more accurate to say that the functional effects produced by
wholes are different from what the parts can produce alone. Synergy is a ubiquitous
phenomena in nature and human societies alike. One well know example is provided by the
emergence of self-organization in social insects, via direct (mandibular, antennation,
chemical or visual contact, etc) or indirect interactions. The latter types are more subtle and
defined by Grass as stigmergy [5] to explain task coordination and regulation in the context
of nest reconstruction in Macrotermes termites. An example [1], could be provided by two
individuals, who interact indirectly when one of them modifies the environment and the
other responds to the new environment at a later time. In other words, stigmergy could be
defined as a typical case of environmental synergy. Grass showed that the coordination and
regulation of building activities do not depend on the workers themselves but are mainly
achieved by the nest structure: a stimulating configuration triggers the response of a termite
worker, transforming the configuration into another configuration that may trigger in turn
another (possibly different) action performed by the same termite or any other worker in the
colony. Another illustration of how stimergy and self-organization can be combined into
more subtle adaptive behaviors is recruitment in social insects. Self-organized trail laying by
individual ants is a way of modifying the environment to communicate with nest mates that
follow such trails [1]. Division of labor is also another paradigmatic phenomena of
stigmergy. Simultaneous task performance (parallelism) by specialized workers is believed
to be more efficient than sequential task performance by unspecialized workers. But by far
more crucial to the present work and aim, is how ants form piles of items such as dead
bodies (corpses), larvae, or grains of sand (fig. 1). There again, stigmergy is at work: ants
deposit items at initially random locations. When other ants perceive deposited items, they
are stimulated to deposit items next to them, being this type of cemetery clustering action,

to appear with minor revisions in, Javier Ruiz-del-Solar, Ajith Abraham and Mario Kppen (Eds.), Hybrid
Intelligent Systems, IOS Press, Frontiers of Artificial Intelligence and Applications, Vol. 87., Santiago, Chile,
Dec. 1-4, 2002.


Figure 1. From left to rigth, a sequential clustering task of corpses performed by a real ant colony. 1500 corpses
are randomly located in a circular arena with radius = 25 cm, where Messor Sancta workers are present. The
fig. shows the initial state (left), 2 hours, 6 hours and 26 hours (rigth) after the beginning of the experiment [1].

organization, and brood sorting a type of self-organization and adaptive behavior. There are
other types of examples (e.g. prey collectively transport), yet stimergy is also present: ants
change the perceived environment of other ants (their cognitive map, according to [3]), and
in every example, the environment serves as medium of communication [1]. Nevertheless,
what all these examples have in common is that they show how stigmergy can easily be
made operational. As mentioned by Bonabeau in [1], that is a promising first step to design
groups of artificial agents which solve problems: replacing coordination (and possible some
hierarchy) through direct communications by indirect interactions is appealing if one wishes
to design simple agents and reduce communication among agents. Finally, stigmergy is often
associated with flexibility: when the environment changes because of an external
perturbation, the insects respond appropriately to that perturbation, as if it were a
modification of the environment caused by the colonys activities. When it comes to
artificial agents, this type of flexibility is priceless: it means that the agents can respond to a
perturbation without being reprogrammed to deal with that particular instability. In our
context, this means that no classifier re-training is needed for any new sets of data-item types
(new classes) arriving to the system, as is necessary in many classical models, or even in
some recent ones. Moreover, the data-items that were used for supervised purposes in early
stages, can now, along with new items, be re-arranged in more optimal ways. Classification
and/or data retrieval remains the same, but the system reorganizes itself in order to deal with
new classes, or even new sub-classes. Recently, several papers (for a good revision see [1])
have highlighted the efficiency of stochastic approaches based on ant colonies for different
problem solving. Data clustering is also one of those problems in which real ants can suggest
very interesting heuristics for computer scientists. One of the first studies using the metaphor
of ant colonies related to the above clustering domain is due to Deneubourg [4], where a
population of ant-like agents randomly moving onto a 2D grid are allowed to move basic
objects so as to cluster them. This method was then further generalized by Lumer et al. [10],
applying it to exploratory data analysis, for the first time. Our aim is to improve these
models, introducing some radical changes and different ant-like heuristics, developing a
model without any local memory and/or hybridization with more classical approaches.
Moreover, the present work will be applied for the first time to image retrieval and
exploratory data analysis. The datasets represent a collection of the most representative 14
types of Portuguese grey granites (fig.3), with a total set of 237 images, each represented by
117 Mathematical Morphology [16,9,12] features. Some chinese granites were also used as a
test, since they can mislead several human experts, leading to a total of 244 images x 177
features. Sections III -IV describe the present proposal, while results are in sections V-VI.

2. Corpse Clustering and Variants into Exploratory Data Analysis

In several species of ants, workers have been reported to sort their larvae or form piles of
corpses literally cemeteries to clean up their nests. Chrtien (see [1]) has performed
experiments with the ant Lasius niger to study the organization of cemeteries. Other
experiments include the ants Pheidole pallidula reported in [4] by Denebourg et al., and
many species actually organize a cemetery. Figure 1 (section I) shows the dynamics of
cemetery organization in another species: Messor sancta. If corpses, or more precisely,
sufficiently large parts of corposes ara randomly distributed in space at the beginning of the
experiment, the workers form cemetery clusters within a few hours, following a behavior
similar to aggregation. If the experimental arena is not sufficiently large, or if it contains
spatial heterogeneities, the clusters will be formed along the edges of the arena or, more
generally, following the heterogeneities. The basic mechanism underlying this type of
aggregation phenomenon is an attraction between dead items mediated by the ant workers:
small clusters of items grow by attracting workers to deposit more items. It is this positive
and auto-catalytic feedback that leads to the formation of larger an larger clusters. In this
case, it is therefore the distribution of the clusters in the environment that plays the role of
stigmergic variable. Denebourg et al. [4] have proposed one model (hereafter called BM, for
basic model) to account for the above-mentioned phenomenon of corpse clustering in ants.
The general idea is that isolated items should be picked up and dropped at some other
location where more items of that type are present. Let us assume that there is only one type
of item in the environment. The probability P
p
for a randomly moving, unladen agent
(representing an ant in the model) to pick up an item is given by (Eq. 2.1):

2
1
1
(
(
,
\
,
,
(
j
+

f k
k
P
p

(2.1)
2
2
(
(
,
\
,
,
(
j
+

f k
f
P
d

(2.2)

where f is the perceived fraction of items in the neighborhood of the agent, and k
1
is a
threshold constant. When f << k
1
, P
p
is close to 1, that is, the probability of picking up an
item is high when there are not many items in the neighborhood. P
p
is close to 0 when f >>
k
1
, that is, items are unlikely to be removed from dense clusters. The probability P
d
for a
randomly moving loaded agent to deposit an item is given by Eq. 2.2., where k
2
is another
threshold constant: for f << k
2
, P
d
is close to 0, whereas for f >> k
2
, P
d
is close to 1. In their
simulations, Denebourg et al. [4] have used k
1
= 0.1 and k
2
= 0.3, testing the spatial sorting
organization of 400 items of two types, on a 100 x 100 grid, using 10 agents and T = 50;
5,000,000 iterations were needed to accomplish a feasible visual result. As expected, the
depositing behavior obeys roughly opposite rules.
As we shall see, the algorithms later described (as well as those proposed) in the present
work, are inspired by this idea, but rely on a more direct evaluation of f. This procedure
should, therefore, be taken as an example among many possible procedures, and changing
the detail how f is perceived does not drastically alter the results, according to Bonabeau [1].
Among other differences proposed later, are also those directly related to how the agents
move on the spatial grid. For instance, real ants are likely to use chemical or tactile cues to
orient their behavior. In their simulations, however, Denebourg et al. [4] have taken the
option of using randomly moving agents, while in here and due to our aim, we suggest the
use of ant-like spatial transition probabilities (section III), based on chemical pheromone
non-linear weighting functions. Significantly more interesting to the present proposal is
however, Lumers and Faieta model [10]. Both authors have generalized Denebourg et al.s
BM [4], to apply it to exploratory data analysis. The idea is to define a distance or
dissimilarity d between objects in the space of object attributes. Let d(o
i
, o
j
) be the distance
between two objects o
i
and o
j
in the feature space. Let us also assume that an agent is located
at site r at time t, and finds an object o
i
at that site. The local density f(o
i
) with respect to
object o
i
at site r is then given by:

( )
( )

]
]
]
,

,


) (
2
) (
,
1
1
, 0 max
r Neigh o
j i
i
sxx j
o o d
s
o f

(2.3)

f(o
i
) is a measure of the average similarity of object o
i
with the other objects o
j
present in the
neighborhood of o
i
. That is, f(o
i
) replaces the fraction f of similar objects in the BM model,
while is a factor that defines the scale of dissimilarity: it is important for it determines
when two items should or should not be located next to each other. Then, and inspired by
Denebourg et al.s functions [4] (Eqs. 2.1 and 2.2), Lumer and Faieta [10] defined picking
up probabilities similarly, where f was substituted by f(o
i
), and dropping probabilities as
P
d
=2.f(o
i
) for f(o
i
) < k
2
, and P
d
=1 for the remaining cases. Lumer and Faieta [10] have used
k
1
= 0.1, k
2
= 0.15 (while BM uses k
2
= 0.3) and = 0.5, with t
max
= 10
6
steps. In order to
illustrate the functioning of their algorithm, the authors used a simple example in which the
attribute space is R
2
, and the values of the two attributes for each object correspond to its
coordinates (x,y) in R
2
. The same distribution was later used for tests in the present work
see fig. 2a, section V. Lumer and Faieta [10] have then added three features to their system,
due to the fact that are generally more clusters in the projected system than in the initial
distribution. These features help to solve this problem, even if they are computationally
intensive and broadly bio-inspired. They are: (1) ants with different moving speeds, (2) a
short term memory, and (3), behavioral switches.

3. From Randomly Moving Agents to Bio-Inspired Spatial Probabilities

Instead of trying to solve some disparities in the basic LF algorithm by adding different ant
casts, short-term memories and behavioral switches (described in section II) which are
computationally intensive, representing simultaneously a potential and difficult complex
parameter tuning, it is our intention (within the present ACLUSTER proposal) to follow real
ant-like behaviors as possible (some other features will be incorporated, as the use of
different response thresholds to task-associated stimulus intensities, discussed later at section
IV). In that sense, bio-inspired spatial transition probabilities are incorporated into the
system, avoiding randomly moving agents, which tend the distributed algorithm to explore
regions manifestly without interest (e.g., regions without any type of object clusters), being
generally, this type of exploration, counterproductive and time consuming. Since this type of
transition probabilities depend on the spatial distribution of pheromone across the
environment, the behavior reproduced is also a stigmergic one. Moreover, the strategy not
only allows to guide ants to find clusters of objects in an adaptive way (if, by any reason, one
cluster disappears, pheromone tends to evaporate on that location), as the use of embodied
short-term memories is avoided (since this transition probabilities tends also to increase
pheromone in specific locations, where more objects are present). As we shall see, the
distribution of the pheromone represents the memory of the recent history of the swarm, and
in a sense it contains information which the individual ants are unable to hold or transmit.
There is no direct communication between the organisms but a type of indirect
communication through the pheromonal field. In fact, ants are not allowed to have any
memory and the individuals spatial knowledge is restricted to local information about the
whole colony pheromone density. In order to design this behavior, one simple model was
adopted (Chialvo and Millonas, [3]), and extended (as in [13]) due to specific constraints of
the present proposal. As described in [3], the state of an individual ant can be expressed by
its position r, and orientation . It is then sufficient to specify a transition probability from
one place and orientation (r,) to the next (r
*
,
*
) an instant later. The response function can
effectively be translated into a two-parameter transition rule between the cells by use of a
pheromone weigthing function (Eq. 3.1):

( )

(
,
\
,
(
j
+
+
1
1 W
(3.1)
( ) ( )
( ) ( )

k j
j j
i i
ik
w W
w W
P
/

(3.2)

This equation measures the relative probabilities of moving to a cite r (in our context, to a
grid location) with pheromone density (r). The parameter is associated with the
osmotropotaxic sensitivity (a kind of instantaneous pheromonal gradient following), and on
the other hand, 1/ is the sensory capacity, which describes the fact that each ants ability to
sense pheromone decreases somewhat at high concentrations. In addition to the former
equation, there is a weigthing factor w(), where is the change in direction at each time
step, i.e. measures the magnitude of the difference in orientation. As an additional condition,
each individual leaves a constant amount of pheromone at the pixel in which it is located
at every time step t. This pheromone decays at each time step at a rate k. Then, the
normalised transition probabilities on the lattice to go from cell k to cell i are given by P
ik
[3]
(Eq. 3.2), where the notation j/k indicates the sum over all the pixels j which are in the local
neighbourhood of k.
i
measures the magnitude of the difference in orientation for the
previous direction at time t-1. That is, since we use a neighbourhood composed of the cell
and its eight neighbours,
i
can take the discrete values 0 through 4, and it is sufficient to
assign a value w
i
for each of these changes of direction. Chialvo et al used the weights of w
0

=1 (same direction), w
1
=1/2, w
2
=1/4, w
3
=1/12 and w
4
=1/20 (U-turn). In addition, coherent
results were found for =0.07 (pheromone deposition rate), k=0.015 (pheromone
evaporation rate), =3.5 (osmotropotaxic sensitivity) and =0.2 (inverse of sensory
capacity), where the emergence of well defined networks of trails were possible. For a
detailed mathematical discussion of this model, and other conditions readers are reported to
[3] and [13]. In order to achieve emergent and autocatalytic mass behaviours around item
groups on the habitat, which can significantly change the expected ant colony cognitive map
(pheromonal field), instead of a constant pheromone deposition rate used in [3], a term not
constant is included. This strategy follows an idea implemented by Ramos et al. [13], while
extending the Chialvo model into digital image habitats. In here, however, this term is
naturally related with the amount of items found in one specific region. So for instance, if we
use
h
as that measure (i.e., the number of items present in one neighborhood), the
pheromone deposition rate T for a specific ant at that specific cell (at time t), should change
to a dynamic value (p is a constant = 0.0025): T = + p
h
. Notice that, if no objects are
present, results expected by this extended model will be equal to those found by Chialvo and
Millonas in [3], since
h
equals to zero.

4. Stressing the Role of Response Thresholds to Task-Associated Stimulus Intensities

In order to model the behavior of ants associated to different tasks, as dropping and picking
up objects, we suggest the use of combinations of different response thresholds. As we have
seen before, there are two major factors that should influence any local action taken by the
ant-like agent: the number of objects in his neighborhood, and their similarity (including the
hypothetical object carried by one ant). Lumer and Faieta [10], use an average similarity
(Eq. 2.3, section II), mixing distances between objects with their number, incorporating it
simultaneously into a response threshold function like the one of Denebourgs (Eq. 2.1, 2.2,
section II). Instead, in the present proposal, we suggest the use of combinations of two
independent response threshold functions, each associated with a different environmental
factor (or, stimuli intensity), that is, the number of objects in the area, and their similarity.
Moreover, the computation of average similarities are avoided in the present algorithm, since
this strategy can be somehow blind to the number of objects present in one specific
neighborhood. In fact, in Lumer and Faietas work [10], there is an hypothetical chance of
having the same average similarity value, respectively having one or, more objects present in
that region. But, experimental evidences and observation in some types of ant colonies, can
provide us with a different answer. After Wilson [17], it is knowned that minors and majors
in the polymorphic species of ants Genus Pheidole, have different response thresholds to
task-associated stimulus intensities (i.e., division of labor). Recently, and inspired by this
experimental evidence, Bonabeau et al. [2], proposed a family of response threshold
functions in order to model this behavior. According to it, every individual has a response
threshold for every task. Individuals engage in task performance when the level of the
task-associated stimuli s, exceeds their thresholds. Authors defined s as the intensity of a
stimulus associated with a particular task, i.e. s can be a number of encounters, a chemical
concentration, or any quantitative cue sensed by individuals. One family of response
functions T

(s) (the probability of performing the task as a function of stimulus intensity s),
that satisfy this requirement is given by (Eq. 4.1) [2]:

( )
n n
n
s
s
s T

(4.1)
2 2
2
5 +

n
n

(4.2)
2
1
1
(
(
,
\
,
,
(
j
+

d k
k

(4.3)
2
2
(
(
,
\
,
,
(
j
+

d k
d

(4.4)
where n>1 determines the steepness of the threshold (normally n=2, but similar results can
be obtained with other values of n>1). Now, at s = , this probability is exactly .
Therefore, individuals with a lower value of are likely to respond to a lower level of
stimulus. In order to take account on the number of objects present in one neighborhood, Eq.
4.1, was used (where, n now stands for the number of objects present in one neighborhood,
and = 5), defining (Eq. 4.2) as the response threshold associated to the number of items
present in a 3 x 3 region around r (one specific grid location). Now, in order to take account
on the hypothetical similarity between objects, and in each ant action due to this factor, a
Euclidean normalized distance d is computed within all the pairs of objects present in that 3
x 3 region around r. Being a and b, a pair of objects, and f
a
(i), f
b
(i) their respective feature
vectors (being each object defined by F features), then d = (1/d
max
).[(1/F).
i=1,F
(f
a
(i)-f
b
(i))
2
]

.
Clearly, this distance d reaches its maximum (=1, since d is normalized by d
max
) when two
objects are maximally different, and d=0 when they are equally defined by the same F
features. Then, and (Eqs. 4.3, 4.4), are respectively defined as the response threshold
functions associated to the similarity of objects, in case of dropping an object (Eq. 4.3), and
picking it up (Eq. 4.4), at site r. Note that these functions are similar to those proposed by
Denebourg et al. [4] (k
1
and k
2
, are threshold constants), while defining probabilities for
picking up or to deposit an item (Eqs. 2.1, 2.2, section II). In here, however, we use them in
reversed order, substituting f by d (where f represented, for Denebourg et al., the perceived
fraction of items in the neighborhood of one agent, having in mind a robotic
implementation). As we can observe, the probability for a specific moving loaded agent to
deposit an item at site r, is given by Eq. 4.3. When d << k
1
(i.e., d close to 0), is close to 1,
that is, the probability of dropping an item is high when the similarity between the loaded
object and one present in the region around r, is high.
Now, in every action taken by an agent, and in order to deal, and represent different
stimulus intensities (number of items and their similarity), present at each site in the
environment visited by one ant, the strategy uses a composition of the above defined
response threshold functions (Eq. 4.2, 4.3 and 4.4). These composed probabilities are
resumed in table 1, and were used as test functions in one preliminar test (section V)
proposed in [10] in order to illustrate the functioning of the algorithm.

TYPES OF HYBRID RESPONSE FUNCTIONS USED
Function Types Picking Probability Dropping Probability
#1
P
p
= (1-). P
d
= .
#2

(a) P
p
= (1-). (b) P
p
= (a) P
d
= . (b) P
d
=
#3

(a) P
p
= 1- (b) P
p
= (a) P
d
= (b) P
d
=
#4
Lumer & Faieta (see section II) Lumer & Faieta (see section II)

Table 1 Types of picking (P
p
) and dropping (P
d
) probability functions used for several tests. In #2,3 half of
the ants used one probability function (a), while the rest used the other function (b). In #4, the LF algorithm
(section II) was fully implemented and followed, but using a toroidal grid.

On the other hand, to evaluate the algorithm behavior, a simple entropy definition is
proposed. For a finite number of n type A items, placed into a finite area grid, the entropy of
A type objects can be defined as the normalized sum, over all n, of the number of empty cells
e (or occupied by objects different from A), surrounding each item A (e
max
= 8, in 3 x 3
regions), that is, E
A
= (

e
i
) / (n. e
max
). As its obvious, several configurations lead to different
values of entropy, where E
A
reaches its maximal value (E
A
= 1) when all type A items are
disconnected from each other. Disconnected clusters of type A items, lead also to an increase
in the value of entropy.

5. Results on a 4 Classes X 200 Gaussian Distributed Points Problem

As mentioned before, we decide to test the algorithm using the same problem as Lumer and
Faieta, introduced by them in [10]. This problem consists of 800 points, represented by two
features each. That is, the attribute space is R
2
, and the values of the two attributes for each
object correspond to its coordinates (x,y) in R
2
. Four clusters of 200 points each were then
generated in attribute space, with x and y distributed according to Normal (or Gaussian)
distributions N(,) of average and variance
2
- see figure 2.a) for details. The 800 data
points (items) were then assigned at random locations on a 57 x 57 non-parametric toroidal
grid, and the clustering algorithm was run with 80 ants, using the function types specified in
table 1. Generally, the following empirical rules were followed, since they lead to good
results: A=4.n
o
, n
a
=A/40, and n
a
/n
o
=0.1, where A is the grid area, n
o
is the number of objects,
and n
a
the number of ants used. As a final result, objects that are clustered together belong
generally to the same initial distribution, and objects that do not belong to the same initial
distribution are found generally in different clusters. In figure 2.b), the evolution of total
entropy (E
total
=E
A
+E
B
+E
C
+E
D
), for 10
6
iterations (as those used in [10]) was plotted for four
different type functions. It is clear to see that probabilistic functions type #3, are the worse in
terms of clustering the different items, while the rest (including the algorithm proposed by
Lumer and Faieta [10]) have similar behaviors, and indeed reduce drastically the value of
entropy of those configurations. We can also get an idea of how the new algorithm clusters
the different items, while the algorithm proceeds (fig. 2.c,d,e,f). In this case type function #1
was used, and as observed, initially randomly deposited items at t=1 in the toroidal grid, are
then at t>0 spatially distributed according to their similarities.

6. Applications into Digital Image Retrieval: A Case Study within a Granite Database

Ornamental stones are quantitatively characterised in many ways, mostly physical, namely,
geological-petrographical and mineralogical composition, or by mechanical strength.
However, the properties of such products differ not only in terms of type but also in terms of
origin, and their variability can also be significant within a same deposit or quarry [12].

a) b)

c) t = 1, E
total
= 2.910 d) t = 10,000, E
total
= 1.744

e) t = 50,000, E
total
= 1.264 f) t = 1E6, E
total
= 0.906

Algorithm. High-level description of ACLUSTER.

/* Initialization */
For every item oi
do
Place oi
randomly on grid
End For
For all agents do
Place agent at randomly selected site
End For
/* Main loop */
For t = 1 to tmax
do
For all agents do
sum = 0
Count the number of items n around site r
If ((agent unladen) and (site r occupied by item oi
)) then
For all sites around r with items present do
/* According to Eqs. 4.2, 4.4 and Table 1 (section 4) */
Compute d,, and Pp

Draw random real number R between 0 and 1
If (R Pp
) then
sum = sum + 1
End If
End For
If ((sum n/2) or (n = 0)) then
Pick up item oi

End If
Else If ((agent carrying item oi
) and (site r empty)) then
For all sites around r with items present do
/* According to Eqs. 4.2, 4.3 and Table 1 (section 4) */
Compute d,, and Pd

Draw random real number R between 0 and 1
If (R Pd
) then
sum = sum + 1
End If
End For
If (sum n/2) then
Drop item oi

End If
End If
/* According to Eqs. 3.1 and 3.2 (section 3) */
Compute W() and Pik

Move to a selected r not occupied by other agent
Count the number of items n around that new site r
Increase pheromone at site r according to n, that is:
Pr
= Pr
+[+(n/)]
End For
Evaporate pheromone by K, at all grid sites
End For
Print location of items
/* Values of parameters used in experiments */
k1
= 0.1, k2
= 0.3, K = 0.015, = 0.07, = 400,
=3.5, =0.2, tmax
= 10
6
steps.


Fig. 2. a) Distribution of points in attribute space: 4 clusters
of 200 points each were generated in attribute space, with x
and y distributed according to Normal (or Gaussian)
distributions N(,): A[x N(0.2,0.1), y N(0.2,0.1)] , B[x
N(0.8,0.1), y N(0.2,0.1)] , C[x N(0.8,0.1), y
N(0.8,0.1)] , D[x N(0.2,0.1), y N(0.8,0.1)] , for objects
type A, B, C and D, respectively. b) Total entropy, E
total
= E
a
+
E
b
+ E
c
+ E
d
, in time, as the swarm evolves new solutions in
clustering four type of objects. Four graphs are shown which
correspond to four types of Probability functions (dropping
and picking) analyzed (see table 1). c,d,e,f) Spatial
distribution of 800 items on a 57 x 57 non-parametric toroidal
grid at several time steps. At t=1, four types of items are
randomly allocated into the grid. As time evolves, several
homogenous clusters emerge due to the ant colony action, and
as expected the total entropy decreases. In order to illustrate
the behavior of the algorithm, items that belong to different
clusters (fig. 2.a), were in here represented by different
symbols: o, , and +. Type 1 probability function was used
with k
1
=0.1 and k
2
=0.3.

Though useful, these methods do not fully solve the problem of classifying a product whose
end-use makes appearance so critically important. Appearance is conditioned not only by the
kind of stone but also depends on the subjective evaluation of beauty and hence of economic
value, which are strongly influenced by supply and demand. Traditionally, the selection
process is based on visual inspection, giving a subjective characterisation of the appearance
of the materials. Thus, one suitable tool to characterise the appearance of these natural stones
is digital image analysis. In the present work we use mathematical morphology (MM, [16]) as
a feature extraction method, as used in past works by Ramos et al. [12]. Generically, the
extraction of features by means of image analysis and mathematical morphology techniques
is implemented in 2 stages: a global and a local analysis. It consists on the extraction of
features before (global analysis) and after (local analysis) the segmentation or mineral phase
classification procedures. This approach is general and can be applied not only to characterise
slab surfaces of granites as also other types of ornamental stones. The method fully described

Figure 3 Spatial distribution of 244 images (representing 14 types of Portuguese Granites + 2 types of
Chinese Granites), at t=1,000,000. Each image (point in the environment) is composed by 117 morphological
and intensity features. Type 1 probability function was used with k
1
=0.1 and k
2
=0.3.

in [12] uses the opening MM operator (erosion followed by dilation) and the closing (dilation
followed by erosion), since they have granulometric properties [11] once are increasing,
extensive (closing) and anti-extensive (opening) and idempotent. From here, a size-intensity
diagram [9], can be extracted and used as a feature vector for each image. Data was collected
for 14 types of Portuguese grey granites, with a total set of 237 images, each represented by
117 MM features. Some chinese granites were also used as a test, since they can mislead
several human experts, leading to a total of 244 images x 177 features. Since we had 244
items (images) to self-organize by the swarm, 24 ants were used (see section V), on a 31 x 31
non-parametric toroidal grid. Fig. 3 shows the final result at t=10
6
, as well as the type of
granite textures clustered.

7. Conclusions

We have presented in this paper a new ant-based algorithm named ACLUSTER for data
unsupervised clustering and data exploratory analysis, while sketching a clear parallel
between a mode of problem-solving in social insects and a distributed, reactive, algorithmic
approach. Some of the mechanisms underlying corpse clustering , brood sorting and those
that can explain the workers behavioral flexiblity, as regulation of labor and allocation of
tasks have first been introduced. As in similar past works applied to document clustering and
text retrieval [14], the role of response thresholds to task-associated stimulus intensities were
stressed as an important part of the strategy, and incorporated into the algorithm by using
compositions of different response functions. These compositions allows the strategy not
only to be more accurate relatively to behaviours found in nature as avoids short-term
memory based strategies, and the use of several artificial ant types (using different speeds),
present in some recent approaches. Behavioral switches as used in [10], were also avoided,
in order to maintain simplicity and to avoid complex parameter settings to be performed by
the domain expert. At the level of agent moves in the grid, a truly stigmergic model was
introduced (section III) in order to deal with clusters of objects, avoiding randomly moves
which can be counterproductive in the distributed search performed by the swarm, and
adopted by all past models. In fact, the present algorithm, along with [14], were the first to
introduce pheromone traces on agents to deter random explorations and encourage objects
cluster formation, a successful feature not implemented even in some recent proposals [7,6].
Results speak for themselves (fig.2,3). While achieving similar results compared to Lumers
model [10], as pointed by the spatial entropy of solutions at each time iteration (fig.2b), the
present algorithm is by far more simple. Moreover, for some of response thresholds
compositions used, results are superior while using the present algorithm for the majority of
time iterations, that is, entropy is always lower, even if at the end they tend to the same
value. As a final advantage, ACLUSTER does not require any initial information about the
future classification, such as an initial partition or an initial number of classes. This novel
strategy was then applied for the first time to digital image retrieval via k-NNR. Generally,
similar types of images tend to be homogeneously clustered together. But more impressive is
that this type of stigmergic map can be used to classify new images, arriving to the system at
any moment. In fact, using 50 randomly chosen images as a test set (from the initial 244;
several sub-sets were used), an average classification and retrieval rate of 94% was achieved
by using k-NNR classification methods (Nearest Neighbor Rule i.e., a label of an unknown
item is determined by the label of his first k neighbors; k=3 was used). Finally, and as
verified by other tests [15] on ACLUSTER, a robust nonstop classifier could be achieved,
which produces class decisions on a continuous stream data, allowing for continuous
mappings. As we know, many categorization systems have the inhability to perform
classification and visualization in a continuous basis or to self-organize new data-items into
the older ones (evenmore into new labels if necessary), unless a new training happens. This
disadvantge is also present in more recent approaches using Self-Organizing Maps [8], as in
Kohonen maps. While a benchmark comparison of the above cited methods should be
interesting to explore, the ability of ACLUSTER to perform continuous mappings and the
incapacity of the latter to conceive it, tend to difficult any serious comparison.

References & Acknowledgements: The first author wishes to thank to FCT-PRAXIS XXI (BD20001-99) -
Ministrio da Cincia e do Ensino Superior, for his Research Fellowship.

[1] Bonabeau E., M. Dorigo, G. Thraulaz. Swarm Intelligence: From Natural to Artificial Systems. Santa Fe
Institute in the Sciences of the Complexity, Oxford University Press, New York, Oxford, 1999.
[2] Bonabeau, E., Thraulaz, G., Denebourg, J.-L., Quantitative Study of the Fixed Response Threshold
Model for the Regulation of Division of Labour in Insect Societies. Roy. Soc. B, 263, pp. 1565-1569, 1996.
[3] Chialvo, Dante R., Millonas, Mark M. How Swarms Build Cognitive Maps. In Luc Steels (Ed.), The
Biology and Technology of Intelligent Autonomous Agents, (144) pp. 439-450, NATO ASI Series, 1995.
[4] Deneubourg, J.-L., Goss, S., Franks, N., Sendova-Franks A., Detrain, C., Chretien, L. The Dynamic of
Collective Sorting Robot-like Ants and Ant-like Robots, SAB90 - 1
st
Conf. On Simulation of Adaptive
Behavior: From Animals to Animats, J.A. Meyer and S.W. Wilson (Eds.), 356-365. MIT Press, 1991.
[5] Grass, P.-P. Termitologia, Tome II Fondation des Socits. Construction. Paris; Masson, 1984.
[6] Handl, J., Meyer B. Improved Ant-based Clustering and Sorting in a Document Retrieval Interface, in J.J.
Merelo, J.L.F. Villacaas, H.-G. Beyer, P. Adamidis (Eds.), PPSN VII - 7
th
Int. Conf. on Parallel Problem
Solving from Nature, Granada, Spain, Sept. 2002.
[7] Hoe, K.M., Lai, W.K., Tai, T.S.Y. Homogeneous Ants for Web Document Similarity Modeling and
Categorization, in M. Dorigo, G. Di Caro and M. Sampels (Eds.), 3
rd
Int. Works. on Ant Algorithms,
Springer-Verlag, Brussels, Sept. 2002.
[8] Kohonen, T. Self-Organizing Maps, Springer-Verlag, Berlin, 1997.
[9] Lotufo, R.A., Trettel, E., Integrating Size Information into Intensity Histogram, in Maragos P., Schafer
R,W., Butt M.A.(Eds.), Mathematical Morphology and its Applications to Image and Signal Processing, pp. 281-
288, Kluwer Academic Publishers, Boston, 1996.
[10] Lumer E. D. & Faieta B. (1994), Diversity and Adaptation in Populations of Clustering Ants. In Cliff, D.,
Husbands, P., Meyer, J. and Wilson S. (Eds.), in From Animals to Animats 3, Proc. of the 3
rd
Int. Conf. on the
Simulation of Adaptive Behavior. Cambridge, MA: The MIT Press/Bradford Books, 1994.
[11] Matheron G. Random sets and Integral Geometry, Wiley, New York, 1975.
[12] Ramos, Vitorino, Pina, P. and Muge, F. From Feature Extraction to Classification: A Multidisciplinary
Approach applied to Portuguese Granites, 11
th
Scan. Conf. on Image Analysis, pp. 817-824, Greenland, 1999.
[13] Ramos, Vitorino, and Almeida, F. Artificial Ant Colonies in Digital Image Habitats - A Mass Behaviour
Effect Study on Pattern Recognition, in Marco Dorigo, Martin Middendorf and Thomas Stzle (Eds.), Proc. of
ANTS'00 2
nd
Int. Workshop on Ant Algorithms, pp. 113-116, Brussels, Belgium, 7-9 Sep. 2000.
[14] Ramos, Vitorino and Merelo, Juan J. Self-Organized Stigmergic Document Maps: Environment as a
Mechanism for Context Learning, in E. Alba, F. Herrera, J.J. Merelo et al. (Eds.), AEB02 - 1
st
Int. Conf. on
Metaheuristics, Evolutionary and Bio-Inspired Algorithms, pp. 284-293, Mrida, Spain, 6-8 Feb. 2002.
[15] Ramos, Vitorino., Swarms on Continuous Data, to appear in Ahmad Lotfi, Bob John and Jon Garibaldi
(Eds.), Recent Advances in Soft Computing, Springer-Verlag, Nottingham, UK, 12-13 Dec. 2002.
[16] Serra J., Image Analysis and Mathematical Morphology, Academic Press, London, 1982.
[17] Wilson, E.O., The Insect Societies, Belknam Press, Cambridge, 1971.