Linked Data Meets Computational Intelligence
Position paper
Christophe Gue´ret
De Boelelaan 1105
1081 HV Amsterdam, Nederlands
Abstract
The Web of Data (WoD) is growing at an amazing rate and
it will no longer be feasible to deal with it in a global way,
by centralising the data or reasoning processes making use of
that data. We believe that Computational Intelligence tech-
niques provides the adaptiveness, robustness and scalability
that will be required to exploit the full value of ever growing
amounts of dynamic Semantic Web data.
Introduction
The Web of Data (WoD) is growing at an amazing rate as
more and more data-sources are being made available on-
line in RDF, and linked. Because of its size and dynamicity
it is no more possible to consider the WoD as a large and
static information system. It should instead be considered a
complex system in constant evolution. Thus, new algorithms
will have to be developed to cope with this new context.
The decentralized nature, and future, of the WoD
As the WoD grows we believe a centralized approach where
data is aggregated into large databases will no longer be pos-
sible. This assertion is based on the following 3 observa-
tions:
1. The massive amount of data and data sources
Data stores have become increasingly efficient and can
now deal with billions of triples but the number of differ-
ent data sources is also growing rapidly. As the number of
data stores increases, incoherences, uncertainty and con-
tradictions are more likely to appear. As an illustration
of these size issues, Table 1 shows examples of publicly
available datasets.
The execution of federated queries over live SPARQL
endpoints is known to be extremely expensive, because
known optimizations (for example to deal with joins) do
not work in the distributed case. Instead, snapshots are
taken at intervals, dumped into gigantic repositories and
made available in database style for querying. But both
the exponential increase of figures and the risk for con-
flicts creates a serious threat to the viability of this ap-
proach.
Copyright c
2010, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
Data source Number of triples
US Census data 1 billion
DBPedia 274 million
MusicBrainz 36 million
DBLP 10 million
WordNet 2 million
Table 1: Example of publicly available data sets and their
respective size.
2. Privacy and coherence problems of centralisation
Data providers may see privacy issues in providing their
data to a central store. This is particularly true for in-
formation related to social networks which is, by na-
ture, highly personal. People providing social informa-
tion about themselves and their friends want to keep con-
trol over their data. For these reasons, decentralisation is
considered to be the future of social networks (Yeung et
al. 2009).
Additionally, the centralisation of data also raise compat-
ibility issues. Data sources using different schemas are
better off kept separated and made compatible through
a translation layer rather than translated into a single
schema and merged into one data store. A Peer Data Man-
agement System (PDMS) (Tatarinov et al. 2003) is an ex-
ample of such a decentralised system. A PDMS is defined
by a set of autonomous peers hosting data according to
different schemas they relate through semantic mappings.
For new peers willing to share data, PDMS has the advan-
tage of a low entry cost, which is limited to the definition
of mappings with one of the peers already present.
One may also consider having data physically centralised
but kept de-centralised. This model is that of a single data
store hosting several named graphs or that of a cloud-
based architecture hosting several stores. Although pre-
serving the provenance/context information, this model
leads to having the data being served by a single store.
As such, privacy concerns may still be present.
3. Opaque data locality
The recent development of cloud computing has high-
lighted an interesting fact about linked data consumers
and providers: their interest is in having data available

somewhere and, more importantly, always accessible -
not in knowing where or how it is stored or processed.
And this is exactly what cloud computing provides: an
abstraction over physical constraints related to data serv-
ing. Once put in a cloud, such as Amazon EC2 (Amazon
Web Services 2009), data is made available in a scalable
fashion. The addition/removal of nodes in the cloud is
a transparent operation for the user which sees the data
provider as a single entity.
Cloud computing comes close to the idea of Tuple-space
based middleware such as the Triple Space Comput-
ing (Fensel 2004). Such middleware creates a persistent
shared information space every peer can read from and
write data to in an asynchronous way. Tasks such as sort-
ing & storing that data or dealing with failure of stor-
age components are in the charge of autonomic entities
that exhibit several self-* behaviours (Kephart and Chess
2003).
We believe that the field of Computational Intelligence
provides a set of techniques that can deal with the decen-
tralised nature of the future WoD. Before discussing how
Computational Intelligence applies to the Semantic Web, we
briefly introduce the field.
Computational intelligence
Computational Intelligence (CI) is a branch of AI focused
on heuristic algorithms for learning, adaptation and evolu-
tion. Evolutionary computation and collective intelligence
are among the most popular paradigms of CI. They provide
particular advantages :
 Learning and adaptation: the performance of CI algo-
rithms improves during their execution. As good results
are found, they learns why these results are good and/or
how to find similar ones. This learning can also cope with
changing conditions and consequently adaptat.
 Simplicity: CI design principles are simple. For instance,
an evolutionary algorithm is based on the survival of the
fittest: in an iterative process, solutions are guessed, ver-
ified and deleted if they are not fit. The expected result,
that is to find an optimal solution to the problem, comes
as a consequence of the basic mecanism. This bottom-up
approach differs from the complex top-down approaches
commonly used to deal with hard problems.
 Interactivity: CI techniques are used in a continously
running, typically iterative, process. This implies two
things: First, at any-time, the best result found so far can
be returned as an answer to the posed problem. Secondly,
CI algorithms can be made interactive and incorporate a
user into the loop. This interaction is of great benefit when
the quality of a result is difficult to appreciate in an auto-
mated way (Takagi 2001).
 Scalability, robustness and parallelisation: all of these
three advantages result from using a population of co-
evolving solutions instead of focusing on only one. Be-
cause each member of the population is independent, al-
gorithms are easy to parallelise. Also, the bad perfor-
mance of some members will be compensated by the
global efficiency of the entire population, making the pop-
ulation robust against individual failures.
It is widely recognised that new adaptive approaches to-
wards robust and scalable reasoning are required to exploit
the full value of ever growing amounts of dynamic Semantic
Web data (Fensel and Harmelen 2007). Computational In-
telligence provides the features that algorithms dealing with
the WoD will have to exhibit.
This paper
This paper highlights some of the research that has been
conducted so far to apply CI to Semantic-Web problems.
Although this combination is a recent field of study, some
achievements have already been made: the eRDF evolution-
ary algorithm makes it possible to query live SPARQL end-
points without any pre-processing, such as the creation of an
index or the retrieval of data dumps; by gossiping seman-
tic information a PDMS can self-organise and improve its
interoperability at large; a swarm of agents taking care of
the localisation of data within a Triple Space produces opti-
mised data clusters and query routes.
All of these applications and some others will be detailed
in the rest of the paper. A particular emphasis will be put on
the joint work currently conducted by the CI and Knowledge
Representation and Reasoning groups of the Vrije Univer-
siteit of Amsterdam. We are particularly interested in fos-
tering emergence and self-organisation (Wolf and Holvoet
2005) and turning data-related tasks into anytime optimisa-
tion problems.
Emergence and self-organisation in
decentralised knowledge systems
One can define emergence as the apparition of global be-
haviour/structure/pattern from local interaction between el-
ements of a system (Wolf and Holvoet 2005). The use of
emergence can be compared to the divide&conquer-like ap-
proach commonly used for distributed computing. In di-
vide&conquer, a supervisor divides a global task into sub-
task, send it to a set of computing resources and coordinate
the results (Figure 1).
Peer Peer Peer
Coordinator Result
Figure 1: Divide and conquer resolution. A central coordi-
nator direct peers toward a specific goal
Architectures focused on emergence define the local in-
teraction between entities. The global goal to achieve is un-
known to the entities, nor does one entity have as a com-
plete view of the solution at any moment. The expected goal
emerges from the local interactions (Figure 2).

Peer
Peer
Peer ResultResult
Result
Figure 2: Global properties emerge from local interaction
among the peers
In the following, we introduce work that has been done to
design emergence of semantic interoperability, structure
(semantic overlays and data clusters) and knowledge (new
triples) in distributed semantic networks.
Semantic interoperability
A set of different data sources is not guaranted to share a
common schema for their data. The semantic interoperabil-
ity of such a set requires setting up schema mediation allow-
ing any participant to communicate with any other.
Semantic gossiping (Cudre´-Mauroux 2006) as been pro-
posed as an alternative to the common strategy of defin-
ing a global schema every participant should map to. Lo-
cal agreements in the form of pair-wise schema mappings
are established among the peers. Gossiping (Demers et al.
1987) is a probabilistic message passing algorithm where
peers exchanges part of their knowledge with their neigh-
bours. By exchanging information about the local agree-
ments they have established, peers are able to improve their
own set of mappings and thus enhance their interoperability
with other peers. As a global result of these local enhance-
ments, semantic interoperability can be achieved at network
scale (Cudre-Mauroux and Aberer 2004).
Self-organising structure
The use of non-centralised RDF spaces poses the problem
of the localisation of data. Every peer is a potential provider
of any given triple. Overlay structure and optimized data
distribution have to be used to provide a more efficient use
of this decentralised data.
A common example of self-organising optimized data dis-
tribution is that of Distributed Hash Tables (DHT) that as-
sign data to a particular peer based on a specific identifier
space. The SwarmLinda project combines a Tuple space
approach with an artificial ants algorithm to create a self-
organising system showing the emergence of data clusters
and optimised routes for querying that data. Artificial ants
applies a sorting algorithm inspired by the brood sorting
ability of real ants: they take some triples at one location
and move them to an other one where similar triples are
found (Graff 2008). When they are moved, traces are left by
the triples on the network connections used for their journey.
These traces are used later on to do query routing (Koske
2009).
The PIAF system (Gue´ret, Monmarche´, and Slimane
2007) also make use of traces left by pieces of informa-
tion moved from peer to peer. Instead of being used for
query routing, these traces are used to efficiently dissemi-
nate information. Ants carying a piece of information will
follow paths on which traces are the most similar to the car-
ried item.
Note that collective intelligence has already proven to be
efficient for network routing (Dorigo and Caro 1998) and
data clustering (Deneubourg et al. 1990). The work reported
in this section shows that these techniques can be combined
with Semantic Web technology in order to make a better use
of decentralised data.
Emergence of Knowledge
Knowledge can also by designed to emerge from local in-
teraction between different RDF spaces. We make a distinc-
tion between two types of knowledge one can expect to see
emerging in a network:
 The computation of a closure under RDFS semantics, for
instance, resulting in the creation of new triples thus ma-
terializing information that was implicit.
 In a more generic way, the application of rules lead to the
creation of new explicit knowledge that was not antici-
pated nor implicit.
Both are essentially the result of some forward-chaining
reasoning based on axioms of different expressivity. Be-
cause of this, it is possible to design a single architecture
able to exhibit both type of emerging properties.
In our work on swarm-based reasoning, reasoning tasks
are distributed among a number of agents, i.e. autonomous
micro-reasoning processes that are referred to as “beasts”,
which explore data distributed across distinct locations. In
contrast to batch-oriented reasoning, their exploration is a
continuous process. The generic model hence defined is that
of an constantly evolving anthill inhabited, used and modi-
fied by restless beasts.
 Implicit knowledge: In (Dentler, Gue´ret, and Schlobach
2009), we applied this model to the computation of clo-
sure over a decentralised RDF graph. Such a graph is
seen as a network, where each subject and each object is
a node and each property an edge. A path is composed of
several nodes that are connected by properties, i.e. edges.
The beasts, each representing an active reasoning rule,
which might be (partially) instantiated, move through the
graph by following their paths. Swarms of independent
light- weight beasts travel from RDF node to RDF node
and from location to location, checking whether they can
derive new information according to the information that
they find on their way. Whenever a beast traverses a path
that matches the conditions of its rule, it locally adds a
new derived triple to the graph. Given an added transition
capability between (sub-)graphs, it can be shown that the
method converges towards closure.
 Explicit knowledge: when dealing with truly decen-
tralised data, that each user keeps locally, there is more
to be done than computing the closure. Beasts can be de-
signed to apply arbitrary rules and extend personal data.
Bibliographic data is an example of such data that would
ideally be maintained by its authors and directly reasoned

over. By looking at the publications made by different
people, a social network can be infered when assuming
that two authors of a same paper know each other. Beasts
can be instanciated with this simple rule and sent walking
over the graph. Whenever an author of a paper and then a
second author of the same paper are found, a knows rela-
tion is created between the two. After some time, a social
network emerges.
An important aspect of swarm based reasoning is that
it is not the data that is sent to a reasoner but it’s instead
the reasoning process that moves towards the data. The
amount of data disclosed by every peer is minimal (only
what is needed by the beast) and is kept in control by the
user (nothing prevents a beast to authenticate itself to the
peer). The progress of this work can be followed online at
http://www.beast-reasoning.net/.
Anytime resolution of complex optimisation
problems across decentralised data sources
Several problems within the Semantic Web can be expressed
as optimisation problems. For instance, ontology mapping
and SPARQL query answering can be seen as a combinato-
rial optimisation problems which consist of assembling an
optimal set of entities. CI techniques are suited for this type
of task, in particular, when the search space is very large
(e.g. costly to explore exhaustivly) or changing. The al-
gorithms presented in this section are all population-based.
That is, they optimise a set of candidate solutions instead
of optimising only one. The way this set is used differs
from one algorithm type to the other, depending on whether
a Evolutionary algorithm or a Particle Swarm Optimisation
is used.
Evolutionary algorithms define a competition metaphor:
only the best solution(s) can survive and be improved. The
basic loop (see Figure 3) consists of the creation of new can-
didate solutions, their evaluation and selection of the best
candidates for the next loop (Eiben and Smith 2003).
Guesssolutions Evaluatesolutions Select asubset Altersolutions
Best so far
Figure 3: A basic evolutionary loop
Particle Swarm Optimisation is instead inspired from bird
swarms and fish schooling were all the candidate solutions
survive and evolve. The competition taking place in the EAs
is replaced by social metaphors were the the best candidate
solution attracts others to it and leads the swarm (Engel-
brecht 2006). At every iteration, individuals move towards
some direction, evaluate the quality of their current position
and compare it with how they perceive the quality of their
neighbours’ positions. Then, a decision is taken on which
direction to go for the next step (see Figure 4).
We now present population-based algorithms applied to
Iteration n+1Iteration n
Figure 4: The population follows its best neighbour. The
globally best individual is greyed. When this best changes,
the entire population changes direction.
the classic problems of ontology matching and query an-
swering over an RDF dataset.
Ontology matching
Ontology matching, that is establishing mappings between
concepts of different ontologies, is a common task Seman-
tic Web developers are confronted with. Considering two
ontologies of n1 and n2 each, the task consists in finding
which concept pairing is the best out of the n1n2 possible
pairs. Several algorithms exist to explore this search space
efficiently and output a set of optimal mappings. These can
be found on the Ontology Alignment Evaluation Initiative
website1.
The continuous optimisation feature of evolutionary al-
gorithms and particle swarm techniques enables the pro-
gressive construction of a set of mappings. GAOM (Wang,
Ding, and Jiang 2006) and GOAL (Martinez-Gil, Alba, and
Montes 2008) are two initatives making use of a genetic
algorithm (GA). They both use a feature-based similarity
function to compare two ontologies. This function is a
weighted combination of measures used to evaluate a map-
ping according to different aspects such as the edit distance
between labels or the number of shared instances. Dur-
ing the evolutionary process, candidate solutions having the
highest similarity score are kept and improved. Going from
one generation to the next one, bad mappings are dropped
and replaced by new ones whose quality will be evaluated in
the next evaluation phase of the evolutionary loop.
The MapPSO (Bock and Hettenhausen 2008) algorithm
follows a similar strategy based on a similarity function but
makes use of PSO instead of a GA. A set of candidate so-
lutions fly above a search space defined by the similarity
function in a quest for positions with the highest correct-
ness. The main difference between the GA approach is the
social behaviour of individuals in a PSO. Instead of a fight-
ing for survival, they co-operate and share knowledge about
their environment.
The shared advantage of using a Genetic Algorithms or a
PSO algorithm for ontology mapping is the ability to incor-
porate new similarity measures at limited cost. This is be-
cause similarities are computed only on candidate solutions
and are not pre-computed over the entire search space.
1http://oaei.ontologymatching.org/

RDF query answering
Giving an answer to a query expressed over an RDF dataset
is basically a pattern matching problem. One has to find
triples matching a particular set of graph patterns. The clas-
sical way of dealing with this is to retrieve, and then merge,
partial results for every single graph pattern to match.
We introduced evolutionary RDF (eRDF) query answer-
ing (Oren, Gue´ret, and Schlobach 2008) as an alternative
to these approaches. Complete variable assignments are
randomly generated and checked against the basic graph
patterns defined by the query. The quality of a solu-
tion is proportional to the correctness of instanciated graph
patterns. This evaluation leaves room for customisation
as any metric can be used to appreciate the correctness
of a given instanciated graph pattern. For instance, a
generated triple <uri,foaf:name,‘‘Amsterdam’’>
can be checked for existance with a boolean outcome
(0=non existant, 1=exist). It can also be compared to an
other triple <uri,foaf:name,‘‘A’dam’’> using an
bounded string-based distance and a real-valued outcome.
As for the ontology matching case, changes on the mea-
sure for the validity of a candidates solutions have a limited
impact on the performances. Based on such well-defined,
and user-specified, notions of similarity eRDF returns “per-
fect” answers if possible, and approximate answers if nec-
essary. The ongoing work on eRDF and can be followed
online on http://www.erdf.nl.
A speciality of eRDF is to allow distributed queries over
live data-sources as only very simple unary queries are
needed (assert correctness). Additionally, eRDF can issue
all of its queries in a fully parallel fashion. There is no
theoretical restriction on the number of data-sources and
their data-size only marginally increases individual response
times. Of course, increasing data-size in combination with
a constant population size will increase convergence time.
However, given the any-time character of evolutionary meth-
ods good answers are still returned comparatively quickly.
This makes eRDF an interesting alternative for exploration
and discovery for the Web of Data. In order to illustrate this
feature, we recently developped on top of eRDF the Like?
discovery engine (Gue´ret, Groth, and Schlobach 2009). Us-
ing that engine a user can submit a query to discover what is
like something else, the likeness being related to the number
of shared properties. Like? is a simple application: once a
user has indicated a keyword and pressed “go”, a query is
sent to Sindice in order to find a resource related to that key-
word. The set of triples describing that resource are turned
into a SPARQL query sent to eRDF which, in turns, streams
the results found. These results are resources sharing at least
one property with the query resource. This application is ac-
cessible at http://www.like.nu.
As demonstrated by the Like? application, eRDF is able
to query not only the sets presented in Table 1 but also the
billion triple set provided in the context of the Billion Triple
Challenge2 of ISWC2009.
2http://challenge.semanticweb.org/
Conclusion
In this position paper, we described how the WoD can and
needs to be seen as an ever changing, dynamic, complex
system. We believe that it will no longer be feasible to deal
with the WoD in a global way, by centralising the data or
reasoning processes making use of that data. New adaptive,
robust and scalable techniques are required to exploit the
full value of ever growing amounts of dynamic Semantic
Web data. We argue for using Computational Intelligence
techniques as the basis for these new algorithms and gave
an overview of the research that has been conducted in this
direction.
References
Amazon Web Services. 2009. Public data sets on aws.
http://aws.amazon.com/publicdatasets/.
Bock, J., and Hettenhausen, J. 2008. MapPSO results for
oaei 2008. In Shvaiko, P.; Euzenat, J.; Giunchiglia, F.;
and Stuckenschmidt, H., eds., OM, volume 431 of CEUR
Workshop Proceedings. CEUR-WS.org.
Cudre-Mauroux, P., and Aberer, K. 2004. A necessary
condition for semantic interoperability in the large. In In
Ontologies, DataBases, and Applications of Semantics for
Large Scale Information Systems (ODBASE.
Cudre´-Mauroux, P. 2006. Emergent semantics. Ph.D. Dis-
sertation, Lausanne. Prix EPFL de doctorats - EPFL doc-
torate award (2007).
Demers, A.; Greene, D.; Hauser, C.; Irish, W.; Larson, J.;
Shenker, S.; Sturgis, H.; Swinehart, D.; and Terry, D. 1987.
Epidemic algorithms for replicated database management.
In Proceedings of the 6th Annual ACM Symposium on Prin-
ciples of Distributed Computing, 1–12.
Deneubourg, J. L.; Goss, S.; Franks, N.; Sendova-Franks,
A.; Detrain, C.; and Chre´tien, L. 1990. The dynamics
of collective sorting robot-like ants and ant-like robots. In
Proceedings of the first international conference on simu-
lation of adaptive behavior on From animals to animats,
356–363. Cambridge, MA, USA: MIT Press.
Dentler, K.; Gue´ret, C.; and Schlobach, S. 2009. Semantic
web reasoning by swarm intelligence. In Proceedings of
the 5th International Workshop on Scalable Semantic Web
Knowledge Base Systems (SSWS2009), collocated at the
8th International Semantic Web Conference (ISWC2009).
Dorigo, M., and Caro, G. D. 1998. Antnet: Distributed
stigmergetic control for communications networks. Jour-
nal of Artificial Intelligence Research 9:317–365.
Eiben, A., and Smith, J. 2003. Introduction to evolutionary
computing. Springer.
Engelbrecht, A. P. 2006. Fundamentals of Computational
Swarm Intelligence. John Wiley & Sons.
Fensel, D., and Harmelen, F. v. 2007. Unifying reason-
ing and search to web scale. IEEE Internet Computing
11(2):96–95.
Fensel, D. 2004. Triple-space computing: Semantic web
services based on persistent publication of information. In-
telligence in Communication Systems 43–53.

Graff, D. 2008. Implementation and evaluation of a swarm-
linda system. Technical report, Freie Universitat Berlin.
Gue´ret, C.; Groth, P.; and Schlobach, S. 2009. erdf: Live
discovery for the web of data. ISWC2009 Billion Triple
Challenge.
Gue´ret, C.; Monmarche´, N.; and Slimane, M. 2007. Shar-
ing resources in a p2p network with artificial ants. Jour-
nal of Mathematical Modelling and Algorithms (JMMA)
6:345–360.
Kephart, J. O., and Chess, D. M. 2003. The vision of
autonomic computing. Computer 36(1):41–50.
Koske, S. 2009. Swarm approaches for semantic triple
clustering and retrieval in distributed rdf spaces. Technical
report, Freie Universitat Berlin.
Martinez-Gil, J.; Alba, E.; and Montes, J. F. A. 2008. Op-
timizing ontology alignments by using genetic algorithms.
In Gue´ret, C.; Hitzler, P.; and Schlobach, S., eds., Nature
inspired Reasoning for the Semantic Web (NatuReS2008),
volume 419 of CEUR Workshop Proceedings. CEUR-
WS.org.
Oren, E.; Gue´ret, C.; and Schlobach, S. 2008. Anytime
query answering in rdf through evolutionary algorithms. In
7th International Semantic Web Conference (ISWC2008).
Takagi, H. 2001. Interactive evolutionary computation:
Fusion of the capabilities of EC optimization and human
evaluation. Proceedings of the IEEE 89(9):1275–1296. In-
vited Paper.
Tatarinov, I.; Ives, Z. G.; Madhavan, J.; Halevy, A. Y.; Su-
ciu, D.; Dalvi, N. N.; Dong, X.; Kadiyska, Y.; Miklau, G.;
and Mork, P. 2003. The piazza peer data management
project. SIGMOD Record 32(3):47–52.
Wang, J.; Ding, Z.; and Jiang, C. 2006. Gaom: Genetic
algorithm based ontology matching. In APSCC ’06: Pro-
ceedings of the 2006 IEEE Asia-Pacific Conference on Ser-
vices Computing, 617–620. Washington, DC, USA: IEEE
Computer Society.
Wolf, T. D., and Holvoet, T. 2005. Emergence versus self-
organisation: Different concepts but promising when com-
bined. 1–15. Springer-Verlag.
Yeung, C.; Liccardi, I.; Lu, K.; Seneviratne, O.; and
Berners-Lee, T. 2009. Decentralization: The future of on-
line social networking. In W3C Workshop on the Future of
Social Networking Position Papers.