Multiagent and Grid Systems – An International Journal 3 (2007) 87–104 87
IOS Press
A biology-inspired model for the automatic
dissemination of information in P2P networks
C. Gue´ret∗, N. Monmarche´ and M. Slimane
Universite´ Franc¸ois Rabelais Tours, Laboratoire d’Informatique, 64, Avenue Jean Portalis, 37200 Tours France
E-mails: {gueret,monmarche,slimane}@univ-tours.fr
Received 6 May 2006
Revised 14 August 2006; 15 September 2006
Accepted 9 October 2006
Abstract. The Personal Intelligent Agents Framework (PIAF) is a framework developped in the context of research on smart
web browsers. Compared to a classical browser, this smart browser should help the user to find and share information using the
Internet network. This paper presents algorithms used in PIAF for the selective dissemination of information in a P2P network.
They use ideas from epidemic systems and are based on an artificial ant colony model. The main contribution made to the domain
of selective information diffusion is the use of estimated profiles defined by pheromone trails associated to connections between
peers.
Keywords: P2P network, gossip algorithm, artificial ants
1. Introduction
In the last decade, we have seen an explosion of electronical devices designed to deal with information. Electronic
documentation has changed the way to write, archive and diffuse content. Collaboration, an important activity in
corporate places [1], also evolved with the development of the Internet network. Nowadays, thanks to Computer
Supported Collaborative Work (CSCW) software, the physical position of collaborators is no longer an issue.
A typical scenario of collaboration involves four elements: users and information they hold, a network and web
browsers. Figure 1 shows the relations between these elements.
Interactions observed between these elements are representative of three main activities of a user: searching for
information in the network, sharing information with other users or exploiting previously acquired information for
his own purpose. For instance, respectively using a web search engine such as Google, participating in a social portal
and using bookmarks to record adresses. Actually, CSCW may be based on the use of various kinds of software,
from e-mails to social portals [2], to provide users with ways of exchanging information. However, in spite of their
popularity, CSCW suffers some limitations [3,4].
– lack of mutual awareness: efficient content sharing implies a global knowledge of the others’ needs. In order
not to annoy everyone, each user should know what his/her collaborators are interested in. But users may
be interested in different domains or make different punctual searches from time to time. Maintaining such
knowledge can be a difficult task. In particular, a high number of collaborators can prohibit the memorization
of other’s center of interests.
∗Corresponding author. Tel.: +33 2 47 36 14 29; Fax: +33 2 47 36 14 22.
ISSN 1574-1702/07/$17.00  2007 – IOS Press and the authors. All rights reserved

88 C. Gue´ret et al. / A biology-inspired model for the automatic dissemination of information in P2P networks
user information network browser
Fig. 1. Representation of collaborative scenario with users, information they hold, a network and web browsers.
user information network browser
Fig. 2. Interactions using an intelligent web browser.
– lack of motivation: users might not be motivated enough to use a software helping them share resources. Such
software may involve, for instance, sending emails to members of the network or using dedicated tools to
share needs and/or findings. In both cases the user has to make an effort and, maybe, change his habits. Such
constraints may weaken his motivation and dissuade him from making use of the software.
– difficulties in conveying one’s wishes: it can be hard for a user to precisely define what they are interested in. If
we take the example of web browsing, users are most likely to jump from page to page looking for interesting
links rather than following a precise and predetermined path.
Out of a CSCW context, an isolated user may also suffer similar problems: expressing his needs to a web search
engine and cope in with the feeling of being “lost in cyberspace” [5] can be a tricky challenge for him.
In usual web-enabled activities, the web browser is only a passive intermediary displaying web pages. Users do
not have direct interactions with it (see Fig. 1). Turning this web browser into an intelligent assistant could help the
user to cope with his problems. Particularely, this enhanced browser could take part in communications and dialog
directly with users, changing the model of interactions (see Fig. 2).
This idea has crossed the mind of many researchers since the 90’s and led to the production of various kinds
of assistive software. Among the most representative solutions proposed, there are: agents guiding users on a
website [6,7], recommender software helping to discover new websites [8,9], collaborative browsing and social
filtering tools [10], or remembrance agents [11,12]. All those solutions share a same limitation: the need to use
them and, eventually, learn how to do it. The collaboration software is a passive tool used by a necessarily active
user. Actually, this is one of the problems of CSCW assistive software were supposed to help users to cope with.
Idealy, an intelligent assistant could be expected to be more pro-active than any other software and, in fine, able to
interact with the user just like an other user would do.
Following this idea, Zhong [13] recently proposed the “Wisdom Web” as a new way to consider information
search. Instead of formulating queries, a user may talk with the “Wisdom Web” just like he would do with any
of his acquaintances. As a reply, this intelligent form of web may send requests to other “classical” web servers
and aggregate answers. Gathered data concerning the user should also be used to provide a personalized response.
In a more concise analysis, Zhong identified six fundamental abilities for the Wisdom Web: efficiently identify
and use different information sources (autonomic Web support), understand semantics, use meta-knowledge to
link ideas, situate things in time, supply personnalized answers and show a sense of humor. An other interesting
approach, related to information management/sharing, is the concept of a “Social Semantic Desktop” which turns
the personnal desktop into a collaboration environment [14]. Such environment might help users to manage their
personal information and share it with others. Technologies needed to implement it would come from the domain
of social networks (for social and knowledge aspects), P2P services (for distributed heterogeneous information

C. Gue´ret et al. / A biology-inspired model for the automatic dissemination of information in P2P networks 89
management) and semantic desktops (for personal information management) [15]. “GoogleDesktop”, the desktop
searching tool created by Google.com, can be seen as a first step toward this new desktop.
The developpement of this kind of intelligent web-enabled services has been recently identified has a new
research domain named “Web Intelligence” (WI). The purpose of WI is to “explore the fundamental roles as well
as practical impacts of Artificial Intelligence (AI) and advanced Information Technology (IT) on the next generation
of Web-empowered products, systems, services, and activities.” [13] and represents the general context of the
development of the Personal Intelligent Agent Framework (PIAF) presented in this paper. This framework is focused
on information flow between users and designed according to three objectives respectively dealing with problems of
mutual awareness, motivation and expressiveness experienced in CSCW use:
1. Detect users sharing some points of interest and help them to get into contact;
2. Operate transparently for the user and being non intrusive. For instance, using Zero-input interfaces [16];
3. Work without having the user to define a profile to specify his center of interests.
Algorithms used in PIAF are developped around the methaphora of a group of people talking in a room. It can be
assumed that everybody may talk or listen (it is a P2P model), nobody has announced to anyone what he is interested
in (no profiles have been published) and discussions consist in free exchanges (no queries are generated). Two
phenomenoms can be observed : first, people gather information listening to conversations in the crowd. Secondly,
people tend to congregate according to keywords they heard and create groups of individuals sharing common
interests. In PIAF, artificial ants are in charge of recreating the first phenomenom while a dedicated component
dynamically checks and optimizes the connections in the P2P overlay network. Each peer is assimilated to an
artificial nest exchanging food (information) with other nests. The resulting dynamic information network could be
used as a new source of information for the “wisdom web” or a way for the users to transparently exchange resources
in a “social semantic desktop”. PIAF is a fully decentralised information sharing system: to operate, it has to be
installed on each participating computer. The use of this artificial ant paradigm along with the absence of any kind
of central server makes our approach to information diffusion in P2P networks asynchronous and parallel.
The remainder of this paper is organized as follows. The Section 2 gives an overview of the related litterature for
information exchange in a P2P network. Algorithms for the communication layer are presented in Section 3 and
followed by details concerning experimentation environment. Experimental results are discussed in Section 4. A
conclusion on the perspectives for future work takes place in Section 5.
2. Related litterature
As a general definition, exchanging data consists in sending a message from a server to a client. A network of
potential collaborators is a so-called P2P network were each user may act as a client (looking for a shared data) or
a server (informing users about what he shares). In such decentralized networks, the duality of roles highlight the
problem for a client of finding a server while looking for a particular information. This problem can be solved using
a centralized server to index shared data [17] however this kind of hybrid architecture has been proved not to scale.
Instead, fully decentralized P2P networks are a prefered solution. For a server, the dual problem of finding clients
to send an information to is not easier to deal with.
Practically, depending on the initiator of the transfer, it is possible to make a distinction between “push” and
“pull” strategies [18] for data exchange. In the “pull” case, the client sends a request to the server which in turn
sends it’s reply, whereas in the “push” case the server directly sends the message to the client. For instance, raising
a question is a classical way to pull for an answer while, although they did not ask for it, the editor of a newspaper
pushes information to readers. In a P2P network, push and pull are respectively illustrated by query processing and
selective dissemination.
– Query processing
Query processing is a classical pull mechanism: a user sends a query and the system returns a list of peers to
contact. In a structured overlay, the location of peers and data is predifined. Identifiers assigned to each peer
and shared data are used to structure the network. The set of all identifiers constitutes the search space for
requests. Usually, it is partitioned among peers with a Distributed Hash Table [19] or a hierarchical ordering

90 C. Gue´ret et al. / A biology-inspired model for the automatic dissemination of information in P2P networks
of identifiers [20]. According to its identifier, a query is routed from peer to peer until is objective is fulfiled.
On the other hand, an unstructured overlay does not rely on a predefined architecture. A query is blindly (e.g.
without information on the underlying topology) routed in the network until a result is obtained [21–23].
Query processing in a structured overlay has the advantage to be very fast and not subject to false positive
answers. Nonetheless the use of identifiers does not fit well with content-based searches and the maintenance
of the overlay structure generates costly network traffic.
– Selective dissemination of information (SDI)
SDI consists in the dissemination of information in a network. A selection is applied to decide wich users the
server will sends messages to. Publish/subscribe mechanism (pub/sub for short) is an example of SDI. Pub/sub
are event based systems: clients subscribe to the event service and servers publish notifications which will be
dispatched to attended recipients. To subscribe, a client defines a profile to specify constraint about events he is
interested in. This profile can be seen as a long term or continuous query [24] and is pulled into pub/sub system.
Depending if this profile defines constraints concerning the topic or the content of published events, the pub/sub
is considered to be topic-based or content-based. Instead of directly dispatching notifications between peers,
common implementations of pub/sub in P2P networks relies on superpeers [21] as event brokers. Clients send
notifications and subscriptions events to the broker they are connected to. Events are then dispatched between
brokers before being locally distributed to clients. In the Siena [25] pub/sub, clients also send “Advertise”
events to notify which kind of event they are likely to publish and help brokers to route subscription events
better.
Using probabilistic multicast protocols like “Gossip” [26] to send messages is an other way to perform SDI.
“Gossiping” consists in sending a message to a subset of the connected peers, according to a given probability.
Depending on the objective for transfers, several strategies have been proposed to compute this probability
of transmission [27]. This mechanism has the advantage to constitute a simple and efficient way to diffuse
messages in a network [28] and is usually used for data aggregation tasks (in sensors networks, for instance).
If the transmission probability is set to 1 (e.g. if all connected peers receive the message), gossip turns into the
“flooding” protocol used, for instance, in firsts versions of Gnutella.
Query processing and pub/sub both relies on the specification of a query. This constraint does not fit in with the
third objective of the PIAF framework (working without queries or profiles). A gossip algorithm is then a more
suitable solution. The objective of PIAF is to dispatch information according to users’ center of interests. Thus, the
transmission probability must be based on the evaluation of the similarity between messages to transfer and a model
for the centers of interests of potential receivers. For a given peer, the higher this probability is, the better are his
chances to be elected as a receiver.
There are two solutions commonly used to gather the model of a given peer interests (sometimes referenced to as
a “profile” or an “expertise”). In a “push” strategy, this profile is published by the peers [29]. That is, periodically,
a given peer sends to his neighbors an up-to-date version of his profile. Whereas in a “pull” strategy, the profile is
sent on demand. For instance, using agents randomly walking in the network and collecting profiles from visited
peers [30]. The model presented here introduces a new strategy inspired by the ideas of overhearing [31] and use of
information trails [32] in a network. The assumption is made that whenever a peer sends a message over the network,
he gives an hint about what he is interested in. Hence, instead of inquiring about profiles, this data can be guessed
looking at trails laid from previous exchanges. Because much information is exchanged on the links, estimated
profiles are actually a synthesis of different user’s center of interests. Therefore, instead of precisely refering to
user’s interests, these profiles define a kind of global memory of communications in the network.
PIAF’s information diffusion algorithms are based on an artificial ant paradigm. This class of algorithm has been
introduced by Dorigo to find the shortest paths in a graph [33] and, later on, successfully applied to combinatorial
problems [34] or network routing [35,36]. The principle of artificial ants is to translate into algorithmic models some
of the real ants’ biological principles. Particularely, many ant species are known to use chemical trails to perform
navigation and forraging. In this case, the involved volatile substances, called pheromones, are a kind of indirect
communication mean between workers of the colony. The good behavior of this class of algorithm, especially within
distributed environment, has led us to use similar design. The same idea has also been explored in the context of
content-based searches in unstructured P2P networks [37].

C. Gue´ret et al. / A biology-inspired model for the automatic dissemination of information in P2P networks 91
The SDI model proposed here can be compared to the idea of Autonomous Gossiping (A/G) proposed by Datta
et al. [38]. A/G, as introduced in the context of mobile ad-hoc networks (MaNETs), is based on epidemic and
economic ideas. It consists on spreading information to interested neighbors (supposed to be epidemic) and avoiding
the sending of information to non-interested (non-epidemic) peers. Local to each peer, information items are in
competition for limited resources (memory, storage space, . . .). Hence, uninteresting information items are dropped
to free room for information of more interest. The main difference between PIAF and A/G is the use of estimated
profiles rather than predefined ones.
Unlike social filtering techniques, selective information dissemination is not aimed at identifying information a
user might be interested in. For instance, purchase recommendations consists in finding items similars to ones added
into the virtual cart by the user. Instead, information dissemination consists in informing others users that somebody
has put this book into the cart. Selection is applied in order not to advertise everyone but, hopefully, only potentially
interested users.
However, the set of information items gathered could be used by a collaborative filtering system. A recommender
agent using PIAF could notify the user with a message like “I detected you are writting a document about ants,
maybe those resources could be of some interest for you:” followed by a list a website adresses, book references or
any other kind of shared data.
To our knowledge, PIAF is the only framework for selective information dissemination designed to operate without
requiring that the user defines his preferences for a subject or another.
3. Architecture of PIAF
The personal intelligent agents framework (PIAF) is made of two structural levels. Each one is dedicated to a
particular aspect:
– The Interface layer is the place for all the components related to Human interactions.
– The Communication layer contains all the components related to the diffusion of information and topology
management.
Let us consider the example of a user browsing the Web; each time this user bookmarks a page, thus indicating
his interest in it, a component within the interface layer will generate an information item. Among other data,
this information item will embed the uniform ressource locator (URL) of the page. This item is sent down to the
communication layer in order to be shared with other peers. Then, another PIAF receiving this event will inform the
user using an interface agent.
This paper will be focused on the components of the communication layer. The global design of this layer is based
on an analogy between peers exchanging information and ants transporting food from nest to nest. The network is
defined as a set of nodes N = {ni} and a set of edges C = {(i, j)}. An edge (i, j) corresponds to a connection from
a node ni to an other node nj . C(t) ⊂ C is the subset of connections present at an instant t. The neighborhood of a
node ni is defined as Vi(t) = {nj | (i, j) ∈ C(t)}.
Artificial pheromones defined on a vector space Rn are used as a memory for information exchanged on the
network. One of them is associated to each information item and others are associated to connections between peers.
Pheromones associated to information I are designed as τ(I) while τ i→j(t) is the pheromone for the connection
(i, j). Associated to a connection, pheromones are used as a memory for information received from other peers.
Therefore, a given peer ni will store incoming pheromone vectors τ i←j·(t), ∀j ∈ Vi(t) and update them when it
receives information items.
The only assumption made for pheromone vectors is the existence of a similarity function s : R n × Rn
→ [0; 1].
It can be used to evaluate the similarity between two connections, as well as the similarity of a connection related to
an information.
The structure of the communication layer is represented on Fig. 3. It is made of different components periodically
activated by an event-driven system:
– Every T f units of time one of the ants (tries to) move an information grabbed from the buffer storage;
– Every T c units of time the connection manager checks the neighborhood efficiency and eventually modifies it;

92 C. Gue´ret et al. / A biology-inspired model for the automatic dissemination of information in P2P networks
Table 1
Relative values for period bounds
Tmin Tmode Tmax
T f T (1 − σ) T T (1 + σ)
T c T (kc − σ) kcT T (kc + σ)
T b T (kb − σ) kbT T (kb + σ)
Ta T (ka − σ) kaT T (ka + σ)
Nest
Kernel
Data management
Information manager
Storage Buffer
Topology management
Connections manager
Address book
Ants
Outcoming data Incoming data
Fig. 3. Components in the communication layer of PIAF.
– Every T a units of time the address book sends to connected peers a request message to update his addressbook.
Activities of these components will be detailed further on in following sections.
The Information manager is in charge of storage zones for information items in the PIAF. “Buffer” and “Storage”
blocs are two containers for information items. Wherever it comes from the bot or from the network, an information
item received by the information manager will be duplicated and stored in both containers. The content of “Storage”
is used by personnal agents from interface layer as a list of information items received. The “Buffer” storage is just
a transition area where items are stored waiting for an ant to take them.
The result is synthesized in the Table 1. σ is a factor used to define how wide is the triangle set around T . The
variable T itself is used as a general scale for time. The three coefficients k c,kb and ka respectively define the
relative mean value for activities for topology checks, information generation and address book updates.
3.1. Activity of ants
Ants work as follows when disseminating an information I: every T f unit of time, the ant will push this
information from its nest ni to another nest nj chosen in his neighborhood Vi(t). This activity consists in first
choosing a destination and then updating pheromones.
3.1.1. Selecting a destination
A stochastic algorithm is used to select nj within Vi(t). According to a similarity threshold smin, neighbors are
first sorted in two groups depending if they are likely to be interested by I or not. Those groups are respectively
defined as Vi(I, t) and Vi(I, t).
Vi(I, t) = {nj ∈ Vi(t) | s(τi←j(t), τ(I))
smin} (1)
Vi(I, t) = {nj ∈ Vi(t) | nj /∈ Vi(I, t)} (2)
Results from this classification are recorded in circular memory buffers E i→j(t) (“E” stands for Evaluation).
They will be used later on for topology management in order to estimate the efficiency of a connection. One buffer

C. Gue´ret et al. / A biology-inspired model for the automatic dissemination of information in P2P networks 93
1 1 1 1 10 0 0 0 0 0 0
Curseur
Fig. 4. Example of buffer E
i→j
(t) with 5 out of 12 positive evaluations.
ni n j n i n j n i n j
nk
(a) n i has an informat ion (repre-
sent ed as t he square ) and sends
it t o n j
(b) n j receives t he informat ion
and records t hat n i has sent it t o
him
(c) A new peer n k connect s t o n i .
Because the informat ion is now on
n j and it can not ret urn to n i , n k
will not receive it
Fig. 5. Example case where sending twice is useful.
is associated to each connection in the network. If a peer is classified as being interested, a 1 will be recorded in the
buffer whereas a value of 0 is written. The Fig. 4 shows the example of a buffer sized 12 with the insertion cursor set
on the fifth position. In this case, the peer would have been in V i(I, t) five times during the last twelve evaluations.
The objective of diffusion is to send an information to a maximum of (interested) peers. One of the current
strategies to achieve this consists in ensuring that an information is not sent to a same peer twice. Hence, the list of
peers visited is stored in information messages. Although, in order to contact some new peers, it may be useful to
send a same information again to a peer even if it has already received it (see Fig. 5 for an example case). Thus,
instead of a list, a circular buffer is embedded with every information to record the indexes of last visited peers. In
the example of Fig. 5, setting the buffer size to 0 allows n j to resend the information to ni. The content of the buffer
for the information I at the instant t is designed as seen(I, t). In previous versions of this algorithm [39,40], the list
of visited peers was not associated to an information item but memorized by ants. Information dissemination were
less efficient than with this newer version.
A subset of Vi(I, t) and Vi(I, t) are defined where visited peers seen(I, t) are excluded. They contain the list of
potential destinations for the ant.
V goodi (I, t) = Vi(I, t) \ seen(I, t) (3)
V badi (I, t) = Vi(I, t) \ seen(I, t) (4)
The probability Pi→j(I, t) for a peer to be elected as a destination by the ant depends on the group it was assigned
to (see Eq. (5)). For an interested peer, this probability is proportional to its relative similarity with I . Meanwhile,
non-interested peers may be equiproportionally chosen. Sending to an either interested or not-interested peer is a
matter of exploitation versus exploration. Exploitation can help having an optimal information flows but, on the
other hand, exploration is needed to find news peers to connect to. Therefore, a trade-off must be found to allow
trying to send information to some other neighbors even if they does not seem to be interested. To do so, ants have
a notion of “free will”. With a probability η, an ant may choose to select a destination interested or not. Also, it has
n+ times more chances to stay at the nest rather than sending the information to a non-interested peer.

94 C. Gue´ret et al. / A biology-inspired model for the automatic dissemination of information in P2P networks
Pi→j(I, t) =





















s(τi←j(t),τ(I))
∑
z∈V
good
i
(I,t)
s(τi←z(t),τ(I))
if nj ∈ V goodi (I, t),
(
1 − ηδ
(
|V goodi (I, t)| > 0
)) 1
|V badi (I, t)| + 1 + n+
if nj ∈ V badi (I, t)
(
1 − ηδ
(
|V goodi (I, t)| > 0
)) 1 + n+
|V badi (I, t)| + 1 + n+
ifi = j
(5)
If n+ = 0, the loopback solution has the same chances to be elected than other destinations.
With δ(·) the Kronecker symbol evaluated to 1 if the expression is satisfied and 0 otherwise. This definition is
used in the case of the group of potentially interested peers being empty. In such a case, the ant must choose one of
the non-interested peers. Because staying at the nest is considered as selecting a peer from uninterested group, such
definition is useless for the first equation.
3.1.2. Updating pheromones
On his way to the peer nj it has chosen, the ant will lay down pheromones. The amount of pheromones is defined
by a factor ρ(I, t) used for both evaporation and deposit of pheromones.
τj←i(t) = (1 − ρ(I, t)) · τj←i(t) + ρ(I, t) · τ(I) (6)
The formulation corresponds to a translation in the vector space R n of pheromones. This fits with the idea of a
peer moving from some interest to another. The translation coefficient is proportional to the number of rounds of
the information compared to his Time To Live (TTL). The number of rounds is the number of transmissions for an
information, also called number of “hops”. The TTL is the maximum allowed number of hops before the information
is dropped. The lower the rounds count is, the higher this coefficient is (see Eq. (7)). A regulation parameter α is
also used to adjust the evolution of ρ(I, t).
ρ(I, t) = ρmax exp
(
−α round(I, t)
TTL(I) − 1
)
(7)
One can adjust α according to a minimum desired value ρmin for ρ(I, t). Since an information is dropped when
round(I, t) = TTL(I) and both round(I, t) and TTL(I) are discrete counters, min (ρ(I, t)) = ρmax exp−α.
min (ρ(I, t)) = ρmax exp−α ⇔ α = ln
(
ρmax
ρmin
)
(8)
3.2. Activity of connections manager
3.2.1. Evaluates the efficiency of neighborhood
The efficiency of a neighborhood depends on the utility of the connections. For each peer n j in its neighborhood
Vi(t), the peer ni estimates the utility Ui→j(t) as the number of positive evaluations divided by the number of
evaluations performed. That is, the number of times a 1 appears in E i→j(t) divided by the size of the buffer.
Ui→j(t) =
number of 1 in Ei→j(t)
size of Ei→j(t) (9)
From this utility factor, a drop score S d is defined.
∀nj ∈ Vi(t), Sdi→j(t) = − log
(
Ui→j(t) + ε
1 + ε
)
(10)
Since Sd ∈
[
0,− log
(
ε
ε+1
)]
, the adjusting parameter ε defines the maximum value for the scores. During
experimentations, it was fixed to 0.01. Some tests with different values of ε values proved its influence on global
performances is limited.

C. Gue´ret et al. / A biology-inspired model for the automatic dissemination of information in P2P networks 95
ni n j
Vj (t) \ { ni}
τjf i (t) τjf k * (t)
Fig. 6. Example case for computing recommandations.
3.2.2. Optimize the neighborhood
First, a connectable peer is picked from the address book. The selection is performed randomly according to the
recommendations Sr of entries in the address book. Only peers which are not already connected may be contacted.
If the selection returns no results, nothing is done concerning connections until the next check procedure. On the
other hand, if an available contact has been found, the peer will try to connect to it.
Peers are limited to a maximum of Vmax connections, if |Vi(t)| = Vmax contacting a new peer involves dropping
an other connection to make room for it. If needed, a probabilistic selection is applied to the drop scores to decide
which peer to disconnect from. However, only connections with a utility beyond a threshold β might be dropped. It
defines a tolerance level for connections and allows to avoid dropping useful ones. For each neighbor, the probability
to drop the connection i, j is P di→j(t). In Eq. (11), V
′
i (t) = {nj ∈ Vi(t) | Ui→j(t)  β} is the subset of neighbors
for which this utility is not high enough.
∀nj ∈ Vi(t), P di→j(t) =



Ui→j(t)
∑
nk∈V
′
i
(t)
Ui→k(t)
if Ui→j(t)  β
0 if Ui→j(t) > β
(11)
Before a connection is dropped, the concerned peer gives his advice and says if it agrees or not. This mechanism
limits the risk for a peer of being left alone in the network, without any connection. Such a scenario could happen
if the connection to drop is the only one the peer has. Thus, an other threshold V min is introduced. It defines a
minimum size for the neighborhood. If |V j(t)| = Vmin, the peer nj will not give his agreement to ni to drop the
connection i, j.
Finally, if a peer has been picked up from the address book and if V i(t) < |Vmax|, the connection is established.
3.3. Activity of addressBook
Every T a unit of time, the peer updates the content of its address book. This update aims at discovering new
peers to connect to as well as updating data concerning those already known. For a given peer n i, updates consists
in sending to each of his neighbors an address book request. Asynchronously, at an instant t, receivers n j send
a response with the address of one of their own neighbors. The choice is probabilistic, peers of V j(t) likely to
share most centers of interests will have more chances to be elected. The Fig. 6 represents data involved in one the
executions of this procedure.
The recommendation score from j for the connection i, k ∗ at instant t is Srj:i→k∗ (t).
Srj:i→k∗ (t) = s(τj←i(t), τj←k∗ (t)), nk∗ ∈ Vj(t) \ {ni} (12)
k∗ is the index of a peer chosen randomly from V j(t) \ {ni}. Each peer of this set has a probability P rj→k(i, t) to
be chosen depending on its relative score.
P rj→k(i, t) =
Srj:i→k(t)
∑
nl∈Vj(t)\{ni}
Srj:i→l(t)
=
s(τj←i(t), τj←k(t))
∑
nl∈Vj(t)\{ni}
s(τj←i(t), τj←l(t))
, ∀nk ∈ Vj(t) \ {ni} (13)
The answer from nj to ni is then a message containing the index k∗ and the value of Srj:i→k∗(t). Once ni
receives this message, it updates its address book. If nk∗ was unknown, it is first added as a new entry. Then,
the recommendation score is recorded but the origin of this recommendation, that is the value of j, is not kept.
Only values of Sr do matter. Hence, later on in this paper, the notation S ri→j(t) will be used to denote the last
recommendation ni got for the connection (i, j).

96 C. Gue´ret et al. / A biology-inspired model for the automatic dissemination of information in P2P networks
Table 2
Average similarities within the artificial dataset
Simimarity Domain1 Domain2 Domain3 Domain4
Domain1 0.74 0.25 0.08 0.08
Domain2 0.25 0.74 0.11 0.17
Domain3 0.08 0.11 0.75 0.07
Domain4 0.08 0.17 0.07 0.75
4. Simulation experiments
Simulation experiments were used to test the performances of the information dissemination and topology man-
agement algorithms in PIAF. The use of a simulation allowed to stress the system in some particular conditions that
would be hard to find in real situations.
To simulate the presence of a user, we added a special process to generate information items. This process,
designed by the term “Bot”, is an agent periodically extracting an information item from the artificial dataset in
order to publish it on the network. The activity period for this agent is set to T b units of time. We suppose peers are
interested in only one topic and only publish information items from this particular domain. This unique center of
interest is referenced to as D(i) for a peer ni. This domain is unknown to the other elements of PIAF (particularely
ants). It is only used as a control parameter for evaluating efficiency of algorithms presented here.
For a same item, bots may publish a given number of duplicates. This could help having a more efficient
dissemination by increasing the chances of receiving this particular information item. Simulations presented in this
paper are performed in the worst conditions, when a bot publishes no duplicates.
During the simulation, the tested network was made of 20 peers. Interests domains were distributed in order to
constitute for each Domain a group of 5 interested peers. No assumption is made concerning the nature of information
modeled by pheromones vectors. For every algorithm, only the similarity between two vectors is considered. For
the tests presented in this paper, the similarity function used is the cosine (see Eq. (14)).
∀τ, τ ′ ∈ Rn, s(τ, τ ′) = < τ · τ
′ >
‖τ‖ · ‖τ ′‖
(14)
An artificial data set has been used to initiate information flows in the network. This dataset is made of four
different domains a peer may be interested in. Each domain contains 100 information items. The size n of the vector
space Rn has been set to 100. This value could be interpreted as a dictionnary of 100 words if the values of τ(I)
correspond to words frequencies in a document. Table 2 indicates the average similarity observed between every
element of a class compared to other elements of another class.
The value in each cell of Table 2 corresponds to the result of the Eq. (15)
Sim(Domainx,Domainy) =
1
|Domainx|
∑
I∈Domainx


1
|Domainy \ {I}|
∑
I′∈Domainy\{I}
s(τ(I), τ(I ′))

 (15)
We suppose transfer delays in sending an information from a peer to an other peer is negligable compared to
delays between two different transfers. Activity periods are then the only time-related factor in simulations. Every
activity period is set to a random value following a triangular distribution. The Fig. 7 depicts such a distribution
considering T ∈ [Tmin, Tmax] with a mean value of Tmode. All the simulations presented here were performed with
kc = 4, ka = 2, kb = 1, σ = 0.3 and T = 20.
Algorithms have been implemented using the discrete event simulator OmnetPP [41]. Results presented here after
have been averaged on 20 executions of algorithms.
4.1. Estimators
Four estimators are used to evaluate the efficiency of the PIAF information dissemination strategy. Two of them
are related to the first aspect which is the distribution of information while the two others are related to the topology
management. They will be successively presented in the following subsections.

C. Gue´ret et al. / A biology-inspired model for the automatic dissemination of information in P2P networks 97
density
parameter
Tmin TmaxTmode
2
Tmax − Tmin
Fig. 7. Probability density function for activity periods.
4.1.1. Efficiency of diffusion algorithm
The quality of diffusion mechanism is estimated with a “completeness” and “precision” estimators. Their definition
are similar to that of factors of “recall” and “precision” commonly used in Information Retrieval [42] problems but
have a different meaning.
Completeness is used to estimate coverage of diffused information items. Considering a particular peer n i
interested in items relevant of D(I), this factor is defined as the ratio between items from D(I) stored in his storage
zone and the total number of items stored. At a network scale, global completeness is defined as the average of
completeness factor of each peer (see Eq. (16)).
Completeness(t) = 1
|N |
∑
ni∈N
number of items related to D(i) in storage zone of n i
number of items related to D(i) available in the network (16)
Intuitively, the best solution is reached when Completeness(t)  1. That is when, at t, every peer gathered
(almost) every information item he is interested in. Ensuring that every peer recovers all of available information
items would be a simple way to achieve it. Yet, in this case peers would also get a lot of information items they are
not interested in.
Precision is used to moderate the result given by the completeness estimator. It is defined as the ratio between the
number of interesting elements gathered by a peer and the total number of information present in his storage zone.
As for completeness, the global precision equals to the average of all peer’s precision factor (see Eq. (17)).
Precision(t) = 1
|N |
∑
ni∈N
number of items related to D(i) in storage zone of n i
number of items in storage zone of n i
(17)
From an algorithmic point of view, optimal solutions are then reached when both precision and completeness
are close to 1. From a user point of view, this conclusion slightly differs. Up to them, the point is to see wether
the storage zone contains interesting items or not (i.e. if precision is  1). Because they can not know how many
interesting items are actually available in the network, users can not evaluate completeness.
4.1.2. Efficiency of topology management algorithm
It has been observed that a network of collaborator exhibits small world properties: the network is made of many
dense groups loosely connected to each other. Those groups appear as people congregate when sharing common
interests. Their density is measured by a clustering coefficient γ i(t) (see Eq. (18)).
γi(t) =
total edges in within peers in Vi(t)
total possible edges in Vi(t)
(18)
Considering a peer ni, γi(t) quantify how dense the neighborhood V i(t) of ni is. If γi(t)  1, ni is part of a
dense group (Fig. 8 shows three examples of clustering coefficient values). The global clustering coefficient γ(t) for
the network is defined as the average of clustering coefficient values for each peer.
The network clustering coefficient is defined as the average of the clustering coefficient of each peer.

98 C. Gue´ret et al. / A biology-inspired model for the automatic dissemination of information in P2P networks
x x x
i (t) = 0 i (t) = 35 = 0 .6 i (t) = 1
Fig. 8. Three clustering coefficient values.
Clustering(t) = 1
|N |
∑
ni∈N
γi(t) (19)
Althought the goal of rewiring process is to cluster peers according to their mutual interests, the clustering
coefficient does not take in account neighbors’ interests. Whether n i is surrounded by peers with similar interests or
not will not affect the value of γi(t).
Following the idea of Schmitz [30], we might have used D(i) to compute a weighted clustering coefficient but
D(i) is only a parameter used for simulations. Algorithms developped in PIAF are supposed not to be aware of
it. Instead, we have choosen to keep up with a classical definition of clustering coefficient and introduce a fourth
estimator named “satisfaction”.
A peer is considered to be satisfied if all of the connections within its neighborhood are useful. Hence, the
satisfaction of ni at an instant t is defined as the average value of utilities for connections in V i(t). From a network
point of view, the global satisfaction is defined as the average of each peer’s satisfaction.
Satisfaction(t) = 1
|N |
∑
ni∈N


∑
nj∈Vi(t)
Ui→j(t)
|Vi(t)|

 (20)
One of the conclusions of these two estimators is that a peer having an high clustering coefficient along with a low
satisfaction is located in an environment were all others peers have centers of interests different from him. With the
metaphora of a meeting, this could be something like a monomaniac of fishing surrounded by people only talking
about, say, astrophysics. An interesting side effect is that even if the peer is not satisfied by its neighborhood, it
my be used as an intermediary between other users. Its role could then be compared to that of referals in referal
networks.
As for Completeness, these two estimators might not be of interest for users. Particularely, they would be more
preocupied by their own satisfaction in using the computer rather than the satisfaction of the program. Results from
Clustering can be indirectly used if the list of connected peers is made available for users. Thanks to this list, they
may discover people with similar center of interests and establish new contacts.
4.2. Selection of parameters to study
Experimental results presented here are focused on parameters involved in the diffusion of information. Efficiency
of diffusion is driven by probabilities of transmission, P i→j(I, t). If this probability is close to 0, there will be no
tranfer and if it comes close to 1 all the peers will be flooded. The parameters directly influencing this probability
are smin, n+, Vmin, Vmax and η. Nonetheless, all of them do not need to be studied: it is possible to define relations
between them, thus allowing to fix one or many of them and adjust the values of others. In order to take η in account
while establishing the relations between parameters, we will consider that |V goodi (I, t)| > 0 and |V badi (I, t)| > 0.
For the case of a peer nj ∈ V goodi (I, t) we have:
Pi→j(I, t)|nj∈V goodi (I,t) = η
s(τi←j(t), τ(I))
∑
z∈V
good
i
(I,t)
s(τi←z(t), τ(I))
(21)
the definition of nj ∈ V goodi (I, t) implies that smin  s(τi←j(t), τ(I))  1, hence:

C. Gue´ret et al. / A biology-inspired model for the automatic dissemination of information in P2P networks 99
η smin
|V goodi (I, t)|
 Pi→j(I, t)|nj∈V goodi (I,t)  η
1
|V goodi (I, t)|smin
(22)
because 2  |V goodi (I, t)| + |V badi (I, t)|  Vmax, we have 1  |V
good
i (I, t)|  Vmax − 1, then:
η smin
Vmax − 1
 Pi→j(I, t)|nj∈V goodi (I,t)  η
1
smin
(23)
If nj ∈ V badi (I, t):
Pi→j(I, t)|nj∈V badi (I,t) = (1 − η)
1
|V badi (I, t)| + 1 + n+
(24)
considering 1  |V badi (I, t)|  Vmax − 1, we establish that:
(1 − η)
1
Vmax + n+
 Pi→j(I, t)|nj∈V badi (I,t)  (1 − η)
1
2 + n+ (25)
Finally, for the loopback probability:
(1 − η) 1 + n
+
Vmax + n+
 Pi→i(I, t)  (1 − η)
1 + n+
2 + n+
(26)
The values of parameters Vmin and Vmax depends of external constraints (limitation of bandwith use, number of
ethernet port available, . . .) so they are considered to be constant. Relations between parameters are established on
the hypothesis that 2  |V goodi (I, t)| + |V badi (I, t)|  Vmax. Using Vmin = 2 ensures this inequality is always
verified and with Vmax = 4 peers have enough free connections to form the caves previously presented.
For these tests, we have fixed ρmin = 0.05, ρmax = 0.8 respectively to be sure that every information item left
a minimal trail after being transfered and to avoid the risk of losing accumulated information stored in pheromones
associated to connections (see Eq. (7)). We also set TTL(I) = 5 to match with the minimum life time span needed
for an information to visit all the peers in a cluster and seen(I, t) = 1 so information items can not be sent-back to
the last sender.
4.3. Optimization of parameters
Once Vmax fixed, all other boundaries are defined by the pair of parameters {η, n+} or {η, smin}. We have
choosen to search for optimal values for these parameters following a three-steps procedure:
1. considering predifined values for η and n+, find an optimal value for smin;
2. considering predifined values for η and smin, find an optimal value for n+;
3. considering predifined values for n+ and smin, find an optimal value for η.
The criteria used is the value reached by estimators when no more information items may be exchanged between
peers. That is when all the bots have stopped their production and all buffer storages are empty.
The influence of smin is studied with arbitrarely fixed values for other parameters: η = 0.6 andn+ = {2, 6, 10, 20}.
It can be observed on Fig. 9 that, globaly, all the estimators falls when smin > 0.7. Because for smin  0 all the
neighbors will be estimated to be interested (thus usefull) and that for smin  1 none of them will be satisfied,
increasing smin decreases Satisfaction(t). Increasing smin, with smin  0.7, also allows to filter information items
better (increasing Precision(t)) while it does not help to get more of them (no variation of Completeness(t)). This
confirms that the higher smin is set, the better are the chance for peers to receive only interesting items. But their
chances to receive nothing at all are higher too.
The conclusion drawn from those first tests is that estimators are affected by the value of n+. Actually, their
values changes but their evolution is nearly the same. In every case, choosing smin  0.7 appears to be an optimal
setting.
Second and third results are drawn on Figs 10(a) and 10(b). They respectively represents the influence of n + for
{η = 0.6, smin = 0.7} and the impact of η with smin = 0.7 and n+ = 6.

100 C. Gue´ret et al. / A biology-inspired model for the automatic dissemination of information in P2P networks
0
0.2
0.4
0.6
0.8
1
0 0.2 0.4 0.6 0.8 1
E
st
im
at
or
s
va
lu
es
smin
clustering
precision
completeness
satisfaction
0
0.2
0.4
0.6
0.8
1
0 0.2 0.4 0.6 0.8 1
E
st
im
at
or
s
va
lu
es
smin
clustering
precision
completeness
satisfaction
0
0.2
0.4
0.6
0.8
1
0 0.2 0.4 0.6 0.8 1
E
st
im
at
or
s
va
lu
es
smin
clustering
precision
completeness
satisfaction
0
0.2
0.4
0.6
0.8
1
0 0.2 0.4 0.6 0.8 1
E
st
im
at
or
s
va
lu
es
smin
clustering
precision
completeness
satisfaction
(a) Evolut ion of est imators wit h n + = 2 (b) Evolut ion of est imators wit h n + = 6
(c) Evolut ion of est imators wit h n + = 10 (d) Evolut ion of est imators wit h n + = 20
Fig. 9. Estimation of an optimum for s
min
considering predefined values for η and n+.
0
0.2
0.4
0.6
0.8
1
2 4 6 8 10 12 14 16 18 20
E
st
im
at
or
s
va
lu
es
n+
clustering
precision
completeness
satisfaction
0
0.2
0.4
0.6
0.8
1
0 0.2 0.4 0.6 0.8 1
E
st
im
at
or
s
va
lu
es
eta
clustering
precision
completeness
satisfaction
(a) Influence of n + wit h η = 0.6 and smin = 0.7 (b) Influence of η with n + = 6 and smin = 0.7
Fig. 10. Estimation of optimums values for n+ and η considering an optimal value of smin.
The first conclusion from Fig. 10(a) is that Precision is continuously increasing on the tested interval. This was
predictable: because kc = 4, the first activity event for a connection manager will approximatively occur 4 times
later than the first activity event for a bot. Then, prior to any other transfer, a peer will first receive an item produced

C. Gue´ret et al. / A biology-inspired model for the automatic dissemination of information in P2P networks 101
0
0.1
0.2
0.3
0.4
0.5
0.6
0 500 1000 1500 2000 2500
C
lu
st
er
in
g
co
ef
fic
ie
nt
Units of time
normal
random
0
0.1
0.2
0.3
0.4
0.5
0.6
0 500 1000 1500 2000 2500
P
ee
rs
a
ve
ra
ge
s
at
is
fa
ct
io
n
Units of time
normal
random
0
0.2
0.4
0.6
0.8
1
0 500 1000 1500 2000 2500
P
re
ci
si
on
Units of time
normal
random
0
0.2
0.4
0.6
0.8
1
0 500 1000 1500 2000 2500
C
om
pl
et
en
es
s
Units of time
normal
random
(a) Evolut ion of clust ering coefficient (b) Evolut ion of sat isfact ion
(c) Evolut ion of precision (d) Evolut ion of complet eness
Fig. 11. “Normal” versus “random” versions of the algorithm.
by it’s bot. If n+ → ∞, the probability to send an item to an interested peer tends to 0 and that of staying at the
nest tends to 1 − η. By the way, all peers are initialy uninterested because of the absence of pheromone trails on
connections. Thus, the higher n+ is, the higher are the chances for items produced by a bot to be the only items
stored in storage zones. Secondly, the Fig. 10(a) has two particular points. For n+ = 6, precision is  0.82 and
clustering is maximum while for the second situed at n+ = 10, precision  0.88 and satisfaction is maximum.
To decide which value to use a choice must be made between clustering and satisfaction. The remarks formulated
concerning the potential interest for users in estimators results allows to sort them. Listed by decreasing importance,
we have Precision(t), Clustering(t), Satisfaction(t) and Completeness(t). According to this, clustering is more
important than satisfaction so n+ = 6 appears to be the best choice.
Results from Fig. 10(b) confirms that for smin = 0.7 and n+ = 6 the optimal value for η is 0.6. This set of optimal
values for smin, n+ and η can be interpreted as it: the minimum of similarity for a connection has to be equal to the
average similarity between items of a same domain; ants should have 40% of chances to perform exploration rather
than optimizing information flow.
4.4. Comparitive study of performances
Algorithms presented in this paper deals with the automonous selective dissemination of information. Because it
works without knowing users’ centers of interest, it can be compared with a random diffusion. For this particular

102 C. Gue´ret et al. / A biology-inspired model for the automatic dissemination of information in P2P networks
Fig. 12. Partial view of a caveman graph.
case, instead of using profiles hints from pheromones trails, ants compute a probability of selection P ′i→j(I, t)
equivalent for every possible direction (see Eq. (27)). Althought they randomly select destinations, ants used in
the random version of the algorithm still evaluates neighbors efficiency. This data is then used for the (unchanged)
rewiring algorithm of Connections Manager.
∀nj ∈ Vi(t) ∪ {ni} , P
′
i→j(I, t) =
1
|Vi(t)| + 1
(27)
The comparative results of our algorithm versus its randomized version are drawn on Fig. 11. With a slight
advantage for the normal version, both heuristics lead to equivalent results for completeness (see Fig. 11(d)). In
both cases, peers will gather an equivalent amount of interesting items. Clearly in favor of “normal”, the results of
precision proves that fewer uninteresting information items are received using pheromones trails (see Fig. 11(c)).
Results of Fig. 11(a) can be appreciated considering that, on the basis of shared interests, 4 groups (“caves”) can
be established in the test network. According to the definition of γ it appears that the density is maximum when all
peers are connected together. By the way, if all groups are connected like this, the four caves are isolated from each
other. This situation does not comply to the connectivity condition of a “small world”. Thus, caves must somehow
stay connected. Let us suppose that for each cave, a connection within the cave is replaced by a connection to an
other cave. The graph formed is a ring of caves connected by one link; a “caveman graph” (see Fig. 12).
It has been proved (in [43, p. 105]), that such a caveman graph has a clustering coefficient equal to 1− 6
k2−1
+O( 1
k3
)
with k = Vmax. Hence, for Vmax = 4, the most clustered graph has a clustering value of 0.6 limited by an amount of
O(1/k3). Higher values would be synonym of disconnected caves. This, with a final value of 0.5 network topology
produced by the “normal” algorithm is fair and better than the “random” version. Satisfaction results (Fig. 11(b))
indicate that, on average, peers are satisfied by 60% of their neighbors. This is more than twice the results expected
expected with a random diffusion.
5. Conclusion
In this paper, we introduced a new kind of selective information dissemination for P2P networks. Instead of
using publish/subscribe paradigm, this algorithm based on epidemic paradigm and on artificial ants model allows the
autonomous diffusion of information. Profiles commonly used to define peer’s centers of interest are remplaced by
estimated profiles. Those estimated profiles are computed from information flow and, because they are associated
to connections between peers, turn the overlay P2P network into a global memory of information exchanged.
PIAF allows the design of zero-input information sharing system and frees the user from having to define a profile.
Results presented in this paper, assessed its fair performances compared to the ones of a random diffusion. The main
drawback of this model is the important number of parameters used. Because they all have related implications,
global optimums are difficult to define. Nonetheless, under some conditions, local optimums can be found.
Futur developments will be focused on the implementation of software components from the interface layer.
Having this done, tests with real users (instead of bots) could be performed. The application of this algorithm to
classification problems is also undergoing with some early interesting results.
Acknowledgements
Some of the icons used in figures are part of the Tango project (http://tango.freedesktop.org) and distributed under
a “Creative Commons Attribution Share-Alike” licence.

C. Gue´ret et al. / A biology-inspired model for the automatic dissemination of information in P2P networks 103
References
[1] F. Bergenti, M. Mari and M. Garijo, Collaborator – Enabling Enterprise Collaboration Through Agents, in: Proceedings of the 2nd
International Workshop on Agent-based Computing for Enterprise Collaboration (WETICE-2004), University of Modena, Italy, 14–16
June 2004, pp. 41–46, IEEE.
[2] O. Buyukkokten, Orkut online community, www.orkut.com, Janvier 2004.
[3] J. Grudin, Why Cscw Applications Fail: Problems in the Design and Evaluation of Organization of Organizational Interfaces, in:
Proceedings of the 1988 ACM conference on Computer-supported cooperative work, Portland, Oregon, United States, 1988, pp. 85–93.
[4] R. Taher, Recherche d’information collaborative, Ph.D. dissertation, Laboratoire CLIPS-IMAG, Grenoble, 5 Mars 2004.
[5] GVU’s WWW Surveying Team, Gvu 8th www user survey’. http://www.cc.gatech.edu/gvu/user survey/survey-1997-10/, October-
November 1997.
[6] H. Lieberman, Letizia: An Agent that Assists Web Browsing, in: Proceedings of the Fourteenth International Joint Conference on Artificial
Intelligence (IJCAI-95), Morgan Kaufmann, 1995, pp. 924–929.
[7] A. Wexelblat and P. Maes, Footprints: History-Rich Tools for Information Foraging, in: Proceedings of the ACM SIGCHI Conference on
Human Factors in Computing Systems (CHI’99), David Lawrence Convention Center, Pittsburgh, Pennsylvania, USA, May 15–20 1999.
[8] H. Lieberman, C. Fry, and L. Weitzman, Exploring the web with reconnaissance agents, Communications of the ACM 44(8) (August 2001),
69–75.
[9] Inc Alexa Internet, Alexa web search, http://www.alexa.com/.
[10] H. Lieberman, N.W. Van Dyke and A.S. Vivacqua, Let’s Browse: A Collaborative Web Browsing Agent, in: Proceedings of the 4th
international conference on Intelligent user interfaces (IUT’99), Los Angeles, California, United States, January 05–08 1999, pp. 65–68.
[11] B. Rhodes and T. Starner, The Remembrance Agent: A Continuously Running Automated Information Retrieval System, in: The Proceedings
of The First International Conference on The Practical Application of Intelligent Agents and Multi Agent Technology (PAAM ’96),
London, UK, April 1996, pp. 487–495.
[12] B. Rhodes, The remembrance agent, www.remem.org, February 2004.
[13] Web Intelligence, N. Zhong, J. Liu and Y. Yao, eds, Springer-Verlag, 2003. ISBN: 3-540-44384-3.
[14] Nepomuk, Nepomuk – the social semantic desktop. nemopuk.semanticdesktop.org.
[15] S. Decker and M. Frank, ‘The social semantic desktop’, Technical Report 2004-05-02, DERI - Digital Enterprise Research Institute, (May
2004).
[16] T. Sharon, H. Lieberman and T. Selker, A Zero-Input Interface for Leveraging Group Experience in Web Prowsing, in: Proceedings of the
8th international conference on Intelligent user interfaces, Miami, Florida, USA, 2003, pp. 290–292.
[17] Napster, Napster file sharing, http://www.napster.com, 2003.
[18] M.J. Franklin and S.B. Zdonik, “Data in Your Face”: Push Technology in Perspective, in: Proceedings ACM SIGMOD International
Conference on Management of Data, 1998, pp. 516–519.
[19] IonStoica, R. Morris, D. Karger, M.F. Kaashoe and H. Balakrishnan, Chord: A Scalable Peer-to-Peer Lookup Service for Internet
Applications, in: SIGCOMM’01, August 27–31 2001.
[20] C.G. Plaxton, R. Rajaraman and A.W. Richa, Accessing Nearby Copies of Replicated Objects in a Distributed Environment, in: Proceedings
of ACM Symposium on Parallel Algorithms and Architectures (SPAA), June 1997.
[21] Clip2: Gnutella protocol specifications. http://www.clip2.com.
[22] I. Clarke, O. Sandberg, B. Wiley and T.W. Hong, Freenet: A Distributed Anonymous Information Storage and Retrieval System, in:
Proceedings of the Workshop on Design Issues in Anonymity and Unobservability, Berkeley, CA, July 2000.
[23] Q. Lv, P. Cao, E. Cohen, K. Li and S. Shenker, Search and Replication in Unstructured Peer-to-Peer Networks, in: Proceedings of the
16th international conference on Supercomputing, New York, New York, USA, 2002, pp. 84–95. ACM Press New York, NY, USA.
ISBN:1-58113-483-5.
[24] M. Koubarakis and C. Tryfonopoulos, Distributed Resource Sharing using Self-Organized Peer-to-Peer Networks and Languages from
Information Retrieval, in: Invitational Workshop on Self-* Properties in Complex Information Systems, Bertinoro, Italy, 31 May and 2
June 2004.
[25] A. Carzaniga, D.S. Rosenblum and A.L. Wolf, Design and evaluation of a wide-area event notification service, ACM Trans. on Computer
Systems 19(3) (August 2001), 332–338.
[26] A.J. Demers, D.H. Greene, C. Hauser, W. Irish, J. Larson, S. Shenker, H.E. Sturgis, D.C. Swinehart and D.B. Terry, Epidemic algorithms
for replicated database maintenance, Operating Systems Review 22(1) (1988), 8–32.
[27] M. Burmester, T. Van Le and A. Yasinsac, Adaptive gossip protocols: Managing security and redundancy in dense ad hoc networks,
Journal of Ad Hoc Networks 4(3) (2006), 504–515.
[28] A.-M. Kermarrec, L. Massoulie and A.J. Ganesh, ‘Reliable probabilistic communication in large-scale information dissemination systems’,
Technical Report MSR-TR-2000-105, Microsoft Research, One Microsoft Way, Redmond, WA 98052, (October 2000). 20.
[29] P. Haase and R. Siebes, Peer Selection in Peer-to-Peer Networks with Semantic Topologies, in: Proceedings of the 13th International World
Wide Web Conference, New York City, NY, USA, 2004.
[30] C. Schmitz, Self-Organization of a Small World by Topic, in: First International Workshop on Peer-to-Peer Knowledge Management
(P2PKM), August 2004.
[31] P. Busetta, L. Serafini, D. Singh and F. Zini, Extending multi-agent cooperation by overhearing, Technical Report 0101-01, Istituo Trentino
di Cultura, (January 2001).
[32] D.W. Payton, Discovering Collaborators by Analyzing Trails Through an Information Space, in: AAAI Fall Symposium on Artificial
Intelligence and Link Analysis, October 23–25 1998.
[33] M. Dorigo, V. Maniezzo and A. Colorni, The Ant System: Optimization by a colony of cooperating agents, IEEE Transactions on Systems,
Man, and Cybernetics Part B: Cybernetics 26(1) (1996), 29–41.

104 C. Gue´ret et al. / A biology-inspired model for the automatic dissemination of information in P2P networks
[34] E. Bonabeau, M. Dorigo and G. Theraulaz, Swarm Intelligence: From Natural to Artificial Systems, Oxford University Press, New York,
1999.
[35] G. Di Caro and M. Dogiro, Antnet: Distributed stigmergetic control for communications networks, Journal of Artificial Intelligence
Research 9 (1998), 317–365.
[36] F. Ducatelle, G. Di Caro and L.M. Gambardella, Using ant agents to combine reactive and proactive strategies for routing in mobile
ad hoc networks, International Journal of Computational Intelligence and Applications (IJCIA) 5(2) (2005), 169–184. Special Issue on
Nature-Inspired Approaches to Networks and Telecommunications.
[37] O. Babaoglu, H. Meling and A. Montresor, Anthill: A Framework for the Development of Agent-Based Peer-to-Peer Systems, in:
Proceedings of the 22th International Conference on Distributed Computing Systems (ICDCS ’02), Vienna, Austria, July 2002.
[38] A. Datta, S. Quarteroni and K. Aberer, Autonomous gossiping: A self-organizing epidemic algorithm for selective information dissemi-
nation in wireless mobile ad-hoc networks, LNCS 3226 (2004), 126–143.
[39] C. Gue´ret, N. Monmarche´ and M. Slimane, Spreading Information Within a p2p Network using Artificial Ants, in: Heighth Annual Meeting
on Health, Science and Technology, 2004.
[40] C. Gue´ret, N. Monmarche´ and M. Slimane, ‘Aide a` la navigation sur internet: utilisation de fourmis artificielles pour l’´echange
d’informations dans un re´seau p2p’, in: Actes du congre´ de la ROADEF, 2005.
[41] Andra´s Varga, Omnet++ discrete event simulation environment. www.omnetpp.org, 2002.
[42] C.J. van Rijsbergen, Information Retrieval, Butterworth, 1979.
[43] D.J. Watts, Small Worlds – The Dynamics of Networks between Order and Randomness, Princeton University Press, November, 24 2003.
ISBN: 0691117047.
Authors’ Bios
Christophe Gue´ret is a Ph.D. Student at the Laboratoire d’informatique de Tours, Universit e´ Franc¸ois Rabelais
Tours. He graduaded from the Engineers School in Computer Science for Industry of Tours, France. His research
interests are focused on social networks, P2P networks and biology-inspired modelisation.
Nicolas Monmarche´ is an assistant professor at the Laboratoire d’informatique de Tours, Universit e´ Franc¸ois
Rabelais Tours. He graduaded from the Engineers School in Computer Science for Industry of Tours, France. His
research interests are focused on ants-inspired algorithms, genetic algorithms and virtual reality.
Mohamed Slimane is a professor in Computer Science at the Laboratoire d’informatique de Tours, Universit e´
Franc¸ois Rabelais Tours. His research interests are focused on hidden Markov models and other kind of stochastics
algorithms.