Ramos, v. merelo, j. 2002: self organized stigmergic document maps-environment as a mechanism for context learning

AEB02 – MÉRIDA, FEBRERO 2002

1

Self-Organized Stigmergic Document Maps:
Environment as a Mechanism for
Context Learning
Vitorino Ramos1 and Juan J. Merelo2

Abstract— Social insect societies and more specifically ant
colonies, are distributed systems that, in spite of the
simplicity of their individuals, present a highly structured
social organization. As a result of this organization, ant
colonies can accomplish complex tasks that in some cases
exceed the individual capabilities of a single ant. 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
to tackle unsupervised clustering as well as data retrieval
problems. The present ant clustering system (ACLUSTER)
avoids not only short-term memory based strategies, as
well as the use of several artificial ant types (using different
speeds), present in some recent approaches. Moreover and
according to our knowledge, this is also the first
application of ant systems into textual document clustering.
Keywords— Ant Systems, Unsupervised Clustering, Data
Retrieval, Data Mining, Distributed Computing, Document
Maps, Textual Document Clustering.

I. STIGMERGY: FROM LOCAL PERCEPTIONS TO GLOBAL
ADAPTIVE SOLUTIONS

S

YNERGY, 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 [5]. Synergy is a
ubiquitous phenomenon 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 Grassé as stigmergy [10,11]
to explain task coordination and regulation in the
1

CVRM-IST Geo-Systems Center, Instituto Superior Técnico, Av.
Rovisco Pais, 1049-001, Lisbon, PORTUGAL, vitorino.ramos
@alfa.ist.utl.pt.
2
Grupo GeNeura, Dpto. de Arquitectura y Tecnología de
Computadores, Facultad de Ciencias, Campus Fuentenueva, s/n 18071,
Granada, SPAIN, jmerelo@geneura.ugr.es.

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. Grassé 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 selforganization can be combined into more subtle adaptive
behaviors is recruitment in social insects, as in nest
cleaning by some workers [1].
Division of labor is another paradigmatic phenomenon
of stigmergy. Simultaneous task performance
(parallelism) by specialized workers is believed to be
more efficient than sequential task performance by
unspecialized workers [12]. Parallelism avoids task
switching, which costs energy and time. A key feature of
division of labor is its plasticity [21]. Division of labor is
rarely rigid. The ratios of workers performing the
different tasks that maintain the colony’s viability and
reproductive success can vary in response to internal
perturbations or external challenges. 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 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 Chialvo and Millonas [3,15,16]), 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 et al. [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

2

RAMOS and MERELO: SELF-ORGANIZED STIGMERGIC DOCUMENT MAPS

(a)

(b)

(c)

(d)

Fig. 1. From (a) to (d), 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 figure shows the initial state (a), 2 hours (b), 6 hours (c)
and 26 hours (d) after the beginning of the experiment (from [4]).

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 colony’s
activities. In other words, the colony can collectively
respond to the perturbation with individuals exhibiting
the same behavior. 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 in the colony evolution in his exploration of
the search-space, can now, along with new items, be rearranged in more optimal ways. Classification and/or
data retrieval remains the same, but the system organizes
itself in order to deal with new classes, or even new subclasses. This task can be performed in real time, and in
robust ways due to system’s redundancy. Recently,
several papers have highlighted the efficiency of
stochastic approaches based on ant colonies for problem
solving. This concerns for instance combinatorial
optimization problems like the Traveling Salesman
problem, the Quadratic Assignment problem, Routing
problem, the Bin Packing problem, or Time Tabling
problems. Numerical optimization problems have been
tackled also with artificial ants, as well as Robotics.
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 [7], 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 and

Faieta [14], applying it to exploratory data analysis, for
the first time. In 1995, the two authors were then beyond
the simple example, and applied their algorithm to
interactive exploratory database analysis, where a
human observer can probe the contents of each
represented point (sample, image, item) and alter the
characteristics of the clusters. They showed that their
model provides a way of exploring complex information
spaces, such as document or relational databases,
because it allows information access based on
exploration from various perspectives. However, this
last work entitled “Exploratory Database Analysis via
Self-Organization”, according to [1], was never
published due to commercial applications. They applied
the algorithm to a database containing the “profiles” of
1650 bank customers. Attributes of the profiles included
marital status, gender, residential status, age, a list of
banking services used by the customer, etc. Given the
variety of attributes, some of them qualitative and others
quantitative, they had to define several dissimilarity
measures for the different classes of attributes, and to
combine them into a global dissimilarity measure (in,
pp. 163, Chapter 4 [1]). 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 document filtering and document
exploratory data analysis. The datasets represent a
collection of 931 words extracted from a Spanish
newspaper. But let us first review some models.
II. CORPSE CLUSTERING , BROOD SORTING MODELS 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. Chrétien [32] has
performed experiments with the ant Lasius niger to
study the organization of cemeteries. Other experiments
include the ants Pheidole pallidula reported in [7] 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. [7] have proposed two closely related
models to account for the two above-mentioned
phenomenon of corpse clustering and larval sorting in
ants. Although the model of clustering reproduces


3

experimental observations more faithfully, the second
one gives rise to more applications. Both models rely on
the same principle, and, in fact, the clustering model is
merely a special case of the sorting model. 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 Pp for a
randomly moving, unladen agent (representing an ant in
the model) to pick up an item is given by:

 k 
Pp =  1 
k + f 
 1


2

(2.1)

where f is the perceived fraction of items in the
neighborhood of the agent, and k1 is a threshold
constant. When f << k1, Pp is close to 1, that is, the
probability of picking up an item is high when there are
not many items in the neighborhood. Pp is close to 0
when f >> k1, that is, items are unlikely to be removed
from dense clusters. The probability Pd for a randomly
moving loaded agent to deposit an item is given by:

 f 
Pd = 
k + f 

 2


2

(2.2)

where k2 is another threshold constant: for f << k2, Pd is
close to 0, whereas for f >> k2, Pd is close to 1. In their
simulations, Denebourg et al. [22] have used k1 = 0.1
and k2 = 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. In
order to evaluate f, Denebourg et al., having a robotic
implementation in mind, assumed that f is computed
through a short-term memory that each agent possesses:
an agent keeps track of the last T time units, being f
simply the number N of items encountered during these
last T time units, divided by the largest possible number
of items that can be encountered during T time units. If
one assumes that only 0 or 1 object can be found within
a time unit, then f=N/T. Their simulations [7], show how
small evenly spaced clusters emerge within a relatively
short time and then merge, more slowly, into fewer
larger clusters, achieving a spatial distribution of objects
very similar to those found in nature (fig. 1).
But, as mentioned above, this procedure lends itself
more easily to a robotic implementation. 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 [2]. 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. [7] 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.
In order to consider sorting, let us now assume that
there are two types, A and B, of items present in the
environment. The principle is the same as before, but
now f is replaced by fA and fB, the respective fractions of
items of types A and B encountered during the last T
time units. Even if several applications could be derived
from here (e.g., segregation phenomenon: Melhuish et
al., [1], pp. 178), the model is unable to reproduce
exactly the brood sorting patterns observed in
Leptothorax ants, where brood sorting is organized into
concentric areas of different brood types.
Significantly more interesting to the present proposal is
however, Lumer’s and Faieta model [14]. Both authors
have generalized Denebourg et al.’s [7] BM 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. For instance, in the BM, two objects
oi and oj can only be either similar or different, so that a
binary distance can be defined, where, for example, d(oi,
oj) = 0 if oi and oj are identical objects, and d(oi, oj) = 1
if oi and oj are not identical objects. Obviously, the very
same idea can be extended to include more complicated
objects, that is, objects with more attributes, and/or more
complicated distances. It is classical in data analysis to
have to deal with objects that can be described by a
finite number n of real-valued attributes (features), so
that objects can be seen as points in Rn, and d(oi, oj) is
the euclidean norm, or any other usual metric, such as
the infinite norm ||…||∞ or the Mahalanobis metric.
The algorithm introduced by Lumer and Faieta [14]
(hereafter LF) consists of projecting the space of
attributes onto some lower dimensional space, typically
of dimension z = 2, so as to make clusters appear with
the following property: intra-cluster distances (i.e.,
attribute distances between objects within clusters)
should be small with respect to inter-clusters distances,
that is, attribute distances between objects that belong to
different clusters. Such a mapping should, therefore,
keep some of the neighborhood relationships present in
the higher-dimensional space (which is relatively easy
since, for instance, any continuous mapping can do the
job) without creating too many new neighbors in m
dimensions, m < n, that would be false neighbors in n
dimensions (which is much less trivial since projections
tend to compress information and may map several wellseparated points in the n-dimensional space onto one
single point in the m-dimensional space). Now, the LF
algorithm works as follows. Let us assume that z = 2.
Instead of embedding the set of objects into R2 or a
subspace of R2, they approximate this embedding by
considering a grid, that is, a subspace of Z2, which can
also be considered a discretization of a real space. Ants
that are moving in this discrete space can directly
perceive a surrounding region of area s2 (a square
Neigh(s x s) of s x s sites surrounding site r). Direct
perception allows a more efficient evaluation of the state
of the neighborhood than the memory- based procedure
used in the BM algorithm: while the BM was aimed to a

4


robotic implementation, the LF algorithm is to be
implemented in a computer, with significantly fewer
material constraints. Let d(oi, oj) be the distance between
two objects oi and oj in the feature space. Let us also
assume that an agent is located at site r at time t, and
finds an object oi at that site. The “local density” f(oi)
with respect to object oi at site r is given by (Eq. 2.3):


1

f (oi ) = max 0, 2
s



 d (oi , o j ) 

1 −

α 
o j ∈Neigh( sxx ) ( r ) 


∑

f(oi) is a measure of the average similarity of object oi
with the other objects oj present in the neighborhood of
oi. That is, f(oi) replaces the fraction f of similar objects
in the BM model, while α 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. For example, if α is too large, there is not
enough discrimination between different items, leading
to the formation of clusters composed of items which
should not belong to the same cluster. If, on the other
hand, α is too small, distances between items in attribute
space are amplified to the point where items which are
relatively close in attribute space cannot be clustered
together because discrimination is too high. Then, and
inspired by Denebourg et al.’s functions [7] (Eqs. 2.1
and 2.2), Lumer and Faieta [14] defined picking up and
dropping probabilities as follows:



k1
Pp (oi ) = 
 k + f (o ) 

i 
 1
Pd (oi ) = 2 f (oi ), if
Pd (oi ) = 1, if

2

f (oi ) < k 2

(2.4)

(2.5)

f (oi ) ≥ k 2

where k1 and k2 are two constants that play a role similar
to k1 and k2 in the BM. Lumer and Faieta [14] have used
k1 = 0.1, k2 = 0.15 (while BM uses k2 = 0.3) and α = 0.5,
with tmax = 106 steps. In order to illustrate the
functioning of their algorithm, the authors used a simple
example in which the attribute space is R2, and the
values of the two attributes for each object correspond to
its coordinates (x,y) in R2. Four clusters of 200 points
each were generated in attribute space, with x and y
distributed according to Normal (or Gaussian)
distributions N(μ,σ) of average μ and variance σ2 (the
same distribution was later used for tests in the present
work – figure 2, section V). The data points (items) were
then assigned at random locations on a 100 x 100 nontoroidal grid, and the clustering algorithm was run with
10 ants. As a 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.
Lumer and Faieta [14] 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:
• Ants with different moving speeds. The swarm is
distributed uniformly in the interval [1, vmax] of possible
speed behaviors, where v is the number of grid units
walked per time unit by an ant along a given grid axis
(vmax = 6). The speed v influences the tendency of an ant
to either pick-up or drop an object through a f(oi)
function similar to Eq. 2.3, where α is replaced by the
term α(1+((v-1)/ vmax)). That is, fast moving ants are not
as selective as slow ants in their estimation of the
average similarity of an object to its neighbors. The
diversity of ants allows to form clusters over various
scales simultaneously: fast ants form coarse clusters on
large scales, i.e. drop items approximately in the right
coarse grained region, while slow ants take over at
smaller scales by placing objects with more accuracy.
• A short term memory. Ants can remember the last m
items they have dropped along with their locations. Each
time an item is picked up, the ant compares the
properties of the item with those of the m memorized
items and goes towards the location of the most similar
item instead of moving randomly. This behavior leads to
a reduction in the number of statistically equivalent
clusters, since similar items have a lower probability of
initiating independent clusters, as argued in [14].
• Behavioral switches. The system exhibits some kind of
self-annealing global behavior since items are less and
less likely to be manipulated as clusters of similar
objects from. Both authors have added the possibility for
agents to start destroying clusters if they have not
performed any deposit or pick up actions for a given
number of time steps. This procedure allows a “heating
up” of the system to escape local non-optimal spatial
configurations.
Finally, Lumer and Faieta [14] suggest that their
algorithm is halfway between a cluster analysis – insofar
as elements belonging to different concentration areas in
their n-dimensional space end-up in different clusters –
and a multi-dimensional scaling, in which an intracluster structure is constructed. Note that in the present
example, the exact locations of the various clusters on
the two-dimensional space are arbitrary, whereas they
usually have a meaning in classical factorial analysis. As
mentioned by Dorigo et al. [8], in a lot of cases,
information about the locations of the clusters is not
necessary or useful (especially in the context of textual
databases) and relaxing the global positioning
constraints allows to speed-up the clustering process
significantly.
III. 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


5

as possible (some other features will be incorporated, as
the use of different response thresholds to taskassociated 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 individual’s 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 due to specific constraints
of the present proposal.
As described by Chialvo and Millonas in [3], the state
of an individual ant can be expressed by its position r,
and orientation θ. Since the response at a given time is
assumed to be independent of the previous history of the
individual, it is sufficient to specify a transition
probability from one place and orientation (r,θ) to the
next (r*,θ*) an instant later. In previous works [15,16],
transition rules were derived and generalized from noisy
response functions, which in turn were found to
reproduce a number of experimental results with real
ants. The response function can effectively be translated
into a two-parameter transition rule between the cells by
use of a pheromone weigthing function:


σ
W (σ ) = 1 +
 1 + γσ







β

(3.1)

This equation measures the relative probabilities of
moving to a cite r (in our context, to a pixel) with
pheromone density σ(r). The parameter β is associated
with the osmotropotaxic sensitivity, recognised by
Wilson as one of two fundamental different types of ants
sense-data processing. Osmotropotaxis, is related to a
kind of instantaneous pheromonal gradient following,
while the other, klinotaxis, to a sequential method
(though only the former will be considered in the present
work as in [3]). Also it can be seen as a physiological
inverse-noise parameter or gain. In practical terms, this
parameter controls the degree of randomness with wich
each ant follows the gradient of pheromone. On the

other hand, 1/γ is the sensory capacity, which describes
the fact that each ant’s ability to sense pheromone
decreases somewhat at high concentrations. In addition
to the former equation, there is a weigthing factor w(∆θ),
where ∆θ 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 η 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 Pik (in, [3]):

Pik =

W (σ i )w(∆ i )
∑ j / k W (σ j )w(∆ j )

(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 wi for each of these
changes of direction. Chialvo et al used the weights of
w0 =1 (same direction), w1 =1/2, w2 =1/4, w3 =1/12 and
w4 =1/20 (U-turn). In addition, coherent results were
found for η=0.07 (pheromone deposition rate), k=0.015
(pheromone evaporation rate), β=3.5 (osmotropotaxic
sensitivity) and γ=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]. Except when indicated, these values will remain in
the following framework. As an additional condition,
each individual leaves a constant amount η of
pheromone at the pixel in which it is located at every
time step t. Simultaneously, the pheromone evaporates
at rate k, i.e., the pheromonal field will contain
information about past movements of the organisms, but
not arbitrarily in the past, since the field forgets its
distant history due to evaporation in a time τ ≅ 1/k. As
in [3], toroidal boundary conditions are imposed on the
lattice to remove, as far as possible any boundary effects
(e.g. one ant going out of the image at the south-west
corner, will probably come in at the north-east corner).
In order to achieve emergent and autocatalytic mass
behaviours around item groups on the habitat, instead of
a constant pheromone deposition rate η used in [6], a
term not constant is included. This upgrade can
significantly change the expected ant colony cognitive
map (pheromonal field). The strategy follows an idea
implemented by Ramos et al. [20], while extending the
Chialvo model into digital image habitats. In here,
however, this term should naturally be related with the
amount of items 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 = η + p∆h. Notice that, if no objects are
present, results expected by this extended model will be

6


equal to those found by Chialvo and Millonas in [3],
since ∆h equals to zero. In this case, this is equivalent to
say that only the swarm pheromonal field is affecting
each ant choices, and not the environment - i.e. the
expected network of trails depends largely on the initial
random position of the colony, and in clusters formed in
the initial configurations of pheromone, through relative
distances. On the other hand, if this environmental term
is added, a stable configuration will appear, which is
largely independent on the initial conditions of the
colony, and becomes more dependent on the nature of
items itself.
IV. 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 [14], use an average similarity (Eq. 2.3, section
2), mixing distances between objects with their number,
incorporating it simultaneously into a response threshold
function like the one of Denebourg’s (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
Faieta’s work [14], 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 [23], it is known 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. ([24,25]),
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 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. A response
threshold θ, expressed in units of stimulus intensity, is
an internal variable that determines the tendency of an
individual to respond to the stimulus s and perform the
associated task. More precisely, θ is such that the
probability of response is low for s<<θ and high for
s>>θ. One family of response functions Tθ (s) (the
probability of performing the task as a function of

stimulus intensity s), that satisfy this requirement is
given by (Bonabeau et al., [24,25]):

Tθ (s ) =

sn
sn +θ n

(4.1)

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 = θ , this probability is
exactly ½. Therefore, individuals with a lower value of θ
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 θ = 5), defining χ (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):

χ=

n2
n 2 + 52

(4.2)

 k 
δ = 1 
k +d 
 1


2

(4.3)

 d 
ε =
k +d 

 2


2

(4.4)

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 fa(i), fb(i) their respective feature vectors (being each
object defined by F features), then d =
(1/dmax).[(1/F).∑i=1,F(fa(i)-fb(i))2]½. Clearly, this distance
d reaches its maximum (=1, since d is normalized by
dmax) when two objects are maximally different, and d=0
when they are equally defined by the same F features.
Then, δ and ε (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. [7] (k1 and k2, 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, which is not the case in here). Let us
now review the behavior of one of these functions. The
probability δ for a specific moving loaded agent to
deposit an item at site r, is given by Eq. 4.3. When d <<
k1 (i.e., d close to 0), δ 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. Similarly, the probability δ
for a specific moving loaded agent to deposit an item at
site r, is low, when d >> k1 (i.e., d close to 1), δ is close


7

to 0, that is, items are unlikely to be deposited together
since they are very different. Now, in order to deal and
represent different stimulus intensities (number of items
and their similarity), the strategy uses a composition of
the above defined response threshold functions (Eq. 4.2,
4.3 and 4.4). These composed functions are used in
every action taken by an agent, present at each visited
site in the environment.
TYPES OF HYBRID RESPONSE FUNCTIONS USED

Function
Types
#1
#2
#3
#4

Picking
Probability
Pp = (1-χ).ε
(a)
Pp = (1-χ).ε
(b)
Pp = ε
(a)
Pp = 1-χ
(b)
Pp = ε
Lumer & Faieta
Eq.(2.4)

Dropping
Probability
Pd = χ.δ
(a)
Pd = χ.δ
(b)
Pd = δ
(a)
Pd = χ
(b)
Pd = δ
Lumer & Faieta
Eq.(2.5)

Table 1 – Types of picking (Pp) and dropping (Pd) 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.

These composed probabilities are resumed in table 1,
and were used as test functions in the “4 classes x 200
Gaussian distributed points” problem (fig. 2) proposed
by Lumer and Faieta [14] (section II), in order to
illustrate the functioning of the algorithm. 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
(emax = 8, in 3 x 3 regions), that is, EA = (∑ ei) / (n. emax).
As its obvious, several configurations lead to different
values of entropy, where EA reaches its maximal value
(EA = 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.
V. 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 [14]. This problem consists of 800 points,
represented by two features each. That is, the attribute
space is R2, and the values of the two attributes for each
object correspond to its coordinates (x,y) in R2. Four
clusters of 200 points each were then generated in
attribute space, with x and y distributed according to
Normal (or Gaussian) distributions N(μ,σ) of average μ
and variance σ2 - see figure 2 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.no, na=A/40, and na/no=0.1, where A

Algorithm.
High-level description of the ACLUSTER algorithm proposed

/* 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,χ, ε 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,χ, δ 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(σ) and Pik
Move to a selected neighboring site 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+[η+(n/α)]
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, η = 0.07, α = 400,
β =3.5, γ=0.2, tmax = 106 steps.
_________________________________________________________

Fig.2. 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(μ,σ): 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.

8


Fig.3. Total entropy, Etotal = Ea + Eb + Ec + Ed, 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).

is the grid area, no is the number of objects, and na 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 6.1.b, the evolution of total entropy
(Etotal=EA+EB+EC+ED), for 106 iterations (as those used
in [14]) 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 [14]) 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. 4). In this case, type function #1
was used. This gives an description of how, initially
randomly deposited items at t=1, are spatially distributed
according to their similarities, by the proposed
algorithm. Finally, note that a toroidal grid was used.
VI. LATENT SEMANTIC ANALYSIS AND DOCUMENT
FILTERING: PRELIMINARY RESULTS ON NEWSPAPERS

t = 1, Etotal = 2.910

t = 50,000, Etotal = 1.264

t = 1E6, Etotal = 0.906
Fig. 4. 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 (see fig. 2), were represented by
different symbols: o, ∆, • and +. Type 1 probability function was used
with k1=0.1 and k2=0.3.

A major challenge in cyberspace in general, or in
document filtering specifically, is to automate the
delivery of relevant information to individual users. In
the last few years, several systems have tried to cope
with information filtering in several ways, like
collaborative filtering (webrings), tracking the user
behaviour while he reads (anatagonomy), using genetic
algorithms to evolve profiles, using ontologies to
represent page content and user interest, but current
results are not very reliable. Other approaches use Self
Organizing Maps (SOM, Kohonen), as in Honkela and
Kohonen’s WEBSOM [13].
The lack of accurate retrieval in information filtering
may due to the lexical matching algorithm used in most
of them. For instance, synonymy and polysemy are very
important factors, but it seems that most systems does
not consider this fact [18]. One of the algorithms that
has been showed a good accuracy is Latent semantic
analysis (LSA, [6]), based on the assumption that
meaning of words can be represented in a
multidimensional space. This algorithm has been
recently and succesfully extended into LSAmercury, an
information filtering engine, by Quesada, Merelo et al.
[18], allowing users to edit their own profiles. Profiles,
queries and documents are all represented as LSA
vectors, which add flexibility to the operations that the
user can perform on her profile, while using
simultaneously non-textual matching and previous
actions informations to infer user preferences.
The present work uses LSA as a feature extraction
method, in order to map 931 words of an article at a
Spanish newspaper. In the LSA model [6, 13], terms and
documents are represented by an m x n incidence matrix
A. Each of the ni unique terms in the document
collection are assigned a row in the matrix, while each
document is assigned a column. SVD is applied to the
resulting matrix, and the main "axes" are them obtained.
Words are projected onto those axes, resulting similar
vector values for words with a similar meaning.


9

Fig. 5 – Spatial distribution of 931 items (words taken from an article at a Spanish newspaper) on a 61 x 61 non-parametric toroidal grid, at
t=106. 91 ants used type 2 probability response functions, with k1=0.1 and k2=0.3. Some independent clusters examples are: (A) anunció, bilbao,
embargo, titulos, entre, hacer, necesídad, tras, vida, lider, cualquier, derechos, medida.(B) dirigentes, prensa, ciu. (C) discos, amigos, grandes.
(D) hechos, piloto, miedo, tipo, cd, informes. (E) dificil, gobierno, justicia, crisis, voluntad, creó, elección, horas, frente, técnica, unas, tarde,
familia, sargento, necesídad, red, obra. (F) voz, puenlo, papel, asseguró. (G) nuestro, europea, china, ahora, poder, hasta, mucho, compañía,
nacionalistas, cambio, asesinado, autor, nuevo.(H) rodríguez, vez, tramitación, gran. (I) se, declara, junto, administración.(J) final, visita,
cataluña, puerta, final, jurisprudencia, todas.(K) fallo, ejército, bajo, real, situación, mission, liga, teatro, decision, queda, nacionalismo,
pasado, director, plan, manos. (L) unica, blancos, ibarra, intensidad, nuevas, las, persona, parlamento, españoles, tarde, seis, otros, euro,
elecciones, servicios, podría, otra, tiene, nada, posibilidad, hablar, porque, música, puntos, compromiso, dentro, doctrina, fiscal, abc, derecho,
atentado, sistema, carrera, razón, televisión, semanas, mundo, natural, mitad. (M) mayo, parís, ciento, consejo, reconoció, me, pero, lo,
ocasión. (N) incluso, pnv, luis, momentos, miembros, regimen, ee.(O) cabeza, ex.(P) oea, municipals, mujer, ayuntamiento, cosas, toda,
novedades, debate, firmado. (Q) domidomingo, estado, otro, primeros, estamos, no.

Thus, each word uses a 50 feature vector. Since we had
931 items (words) to self-organize by the swarm, 91 ants
were used (see section V), on a 61 x 61 non-parametric
toroidal grid. Figure 3 shows the final result at t=106.
VII. CONCLUSIONS
We have presented in this paper a new ant-based
algorithm named ACLUSTER for data unsupervised
clustering and data exploratory analysis. The aim of this
paper was to draw 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 worker’s behavioral flexiblity, as
regulation of labor and allocation of tasks have first been
analysed. The role of response thresholds to taskassociated stimulus intensities was then stressed as an
important part of the strategy, and were in fact
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, as well as the use of several artificial
ant types (using different speeds), present in some recent
approaches. Behavioral switches as used by Lumer and
Faieta [14], 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 adopted
and extended 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 many past models. Results
speak for themselves. While achieving similar results
compared to Lumer’s and Faieta model [14], as pointed
by the spatial entropy of solutions at each time iteration,
the present algorithm is by far more simple. Moreover,
for some of response thresholds compositions used,
results are better 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
document word clustering. While maintaining some
coherence for some words and for some clusters, results,
however, are far from being optimal. This probably
occurred due to the relatively quality of LSA features
regarding this specific newspaper article. At this
moment of the research ACLUSTER has not been
extensively tested within this type of features, and our
conclusions are limited because of that.

10


VIII. REFERENCES
[1] Bonabeau E., M. Dorigo, G. Théraulaz. 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., “From Classical Models of
Morphogenesis to Agent-based Models of
Pattern Formation”, Artificial Life, 3, pp. 191209, 1997.
[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. 439450, NATO ASI Series, 1995.
[4] Chrétien, L., “Organisation Spatiale du Matériel
Provenant de l’excavation du nid chez Messor
Barbarus et des Cadavres d’ouvrières chez
Lasius niger (Hymenopterae: Formicidae)”,
Ph.D. dissertation, Université Libre de
Bruxelles, 1996.
[5] Corning, P. “The Synergism Hypothesis: On the
Concept of Synergy and it´s Role in the
Evolution of Complex Systems”, Journal of
Social and Evolutionary Systems, 21 (2), 1998.
[6] Deerwester, S., Dumais, S.T., Furnas, G.W.,
Landauer, T.K., Harshman, R., “Indexing by
Latent Semantic Analysis”, Journal of the
American Society for Information Science,
41(6), pp. 391-407, 1990.
[7] 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”, SAB’90 - 1st Conf.
On Simulation of Adaptive Behavior: From
Animals to Animats, J.A. Meyer and S.W.
Wilson (Eds.), 356-365. Cambridge, MA: MIT
Press, 1991.
[8] Dorigo, M., Bonabeau, E., Theraulaz, G., “Ant
Algorithms and Stigmergy”, Future Generation
Computer Systems, 16, Elsevier, pp. 851-871,
2000.
[9] Franks, N.R. and Sendova-Franks, A.B., “Brood
Sorting by Ants: Distributing the Workload
over the Work Surface”, Behav. Ecol.
Sociobiol., 30, 109-123, 1992.
[10] Grassé, P.-P. “Termitologia, Tome II”
Fondation des Sociétés. Construction. Paris;
Masson, 1984.
[11] Grassé, P.-P. “La Reconstruction du nid et les
Coordinations
Inter-Individuelles
chez
Bellicositermes Natalensis et Cubitermes sp. La
théorie de la Stimergie: Essai d’interprétation
du Comportement des Termites Constructeurs”
Insect Soc., 6, pp. 41-80, 1959.
[12] Jeanne, R.L. “The Evolution of the
Organization of Work in Social Insects”,
Monit. Zool. Ital., 20, pp. 119-133, 1986.
[13] Kaski, S., Honkela, T., Lagus, K., Kohonen, T.,
“WEBSOM Self Organizing maps of
Documents Collections”, Neurocomputing,
Vol. 21, pp. 101-117, 1998.

[14] 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 3rd Int. Conf. on the Simulation
of Adaptive Behavior. Cambridge, MA: The
MIT
Press/Bradford Books, 1994.
[15] Mark M. Millonas. “Swarms, Phase transitions,
and Collective Intelligence”. In C.G. Langton
(Ed.), Artificial Life III, pp. 417-445, Santa Fe
Institute Studies in the Sciences of the
Complexity, Vol. XVII, Addison-Wesley
Reading, Massachussetts,1994.
[16] Mark M. Millonas. “A Connectionist-Type
Model of Self-Organized Foraging and
Emergent Behavior in Ant Swarms”, Journal
Theor. Biology, (159) pp.529, 1992.
[17] Oprisan, S.A., V. Holban, B. Moldoveanu,
“Functional Self-Organization Performing
Wide-Sense Stochastic Processes”, Phys. Lett.
A, 216, pp. 303-306, 1996.
[18] Quesada, J.F., Merelo, J.J., Castellano, J.G.,
Garcia, M., Cillero, M., “Personalization as
Disambiguation: LSAmercury, a Information
Filtering Engine based on Latent Semantic
Analysis”, preprint submitted to Elsevier
Science, 11 May, 2001.
[19] Ramos, V., The Biological Concept of Neoteny
in Evolutionary Colour Image Segmentation Simple Experiments in Simple Non-Memetic
Genetic Algorithms, in Egbert J., W. Boers et
al. (Eds.), Applications of Evolutionary
Computation, Lecture Notes in Computer
Science, Vol. 2037, pp. 364-378, SpringerVerlag, Berlin-Heidelberg, April 18-20, 2001.
[20] Ramos, V., 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 Stüzle (Eds.), Proc. of
ANTS'2000 – 2nd Int. Workshop on Ant
Algorithms (From Ant Colonies to Artificial
Ants), pp. 113-116, Brussels, Belgium, 2000.
[21] Robinson, G.E. “Regulation of Vision of Labor
in Insect Societies”, Annu. Rev. Entomol. 37,
pp. 637-665, 1992.
[22] Steinberg, M.S., T.J. Poole, “Strategies for
Specifying Form and Pattern: Adhesion-Guided
Multicellular Assembly”, Phil. Trans. R. Soc.
Lond. B, 295, pp. 451-460, 1981.
[23] Wilson, E.O., The Insect Societies, Belknam
Press, Cambridge, 1971.
[24] Bonabeau, E., Sobkowski, A. Theraulaz, G.,
Denebourg, J.-L., Adaptive Task Allocation
Inspired by a Model of Division of Labor in
Social Insects, in Lundh, D. et al. (Eds.):
Biocomputing and Emergent Computation:
Proceedings of BCEC’97, pp. 36-45, 1997.
[25] Bonabeau, E., Théraulaz, G., Denebourg, J.-L.,
“Quantitative Study of the Fixed Response
Threshold Model for the Regulation of Division
of Labour in Insect Societies”. Proceedings
Roy. Soc. London B, 263, pp. 1565-1569, 1996.

Ramos, v. merelo, j. 2002: self organized stigmergic document maps-environment as a mechanism for context learning

More Related Content

Viewers also liked (11)

Similar to Ramos, v. merelo, j. 2002: self organized stigmergic document maps-environment as a mechanism for context learning (20)

More from ArchiLab 7 (20)

Recently uploaded (20)

Ramos, v. merelo, j. 2002: self organized stigmergic document maps-environment as a mechanism for context learning