QUERY ROUTING IN NOMADIC ENVIRONMENTS

WITH PUMAS

Angela Carrillo-Ramos

, Marlène Villanova-Oliver

, Jérôme Gensel

Hervé Martin

and Miguel Torres

LIG Laboratory, STEAMER Team, 681 rue de la Passerelle, 38402 Saint Martin D’Hères, France

Pontificia Universidad Javeriana, ISTAR Team, Carrera 7 No. 40 – 62, Bogotá, Colombia

Keywords: P2P systems, Agents, Query Routing, nomadic environment, adaptation.

Abstract: PUMAS is a framework based on agents and a Peer to Peer (P2P) approach, which allows nomadic users to

access several information sources through different types of devices (eventually mobile). PUMAS provide

the nomadic user with information adapted to her/his preferences and to the characteristics of the context of

use (i.e. location, connection moment, inherent characteristics of the access device). In order to deliver

adapted results, PUMAS relies on mechanisms for enriching the initial query, by means of adding criteria

based on user preferences and context of use. This phase of query enrichment precedes a query routing

process composed of three activities: the query analysis, the selection of sources that are able to answer to

the query and the redirection of the query towards these sources. We present the algorithms and the

knowledge managed by PUMAS agents in these activities and we illustrate the query routing process that

PUMAS executes in order to deliver adapted information.

1 INTRODUCTION

Mobile Devices (MD) such as PDA, phones, laptops,

etc., are used to access distant information sources,

and are also used as information sources by

themselves. In this context, to guarantee access to

several information sources and to share these

resources among users, architectures and

technologies with high communicative potential are

required. This potential is one of the main

characteristics of Peer to Peer (P2P) systems.

The handling of queries that come from nomadic

users becomes an increasingly complex process

since the data that can satisfy the query can be held

by different information sources. Additionally to the

number of potential sources, the heterogeneity of the

source’s structure and access methods can be an

issue for their exploitation. Finally, it is difficult to

provide the user with the best-adapted information to

his needs and to the characteristics of her access

device. These issues give rise to the need to

previously analyze the queries, in order to identify

the sources that will give the most appropriate

answers to the query (this supposes that knowledge

about the information managed by each source is

available), then to evaluate and finally to integrate

the results. These activities correspond to a

traditional query routing process (Xu et al., 1999).

However, these activities do not consider explicitly

the adaptation of information. Our approach

addresses this issue and proposes a query enrichment

process, which comes upstream from the routing

process. This enrichment “augments” the initial

query by adding criteria that consider both user

preferences and context of use of the current session.

In our work, this context is composed of information

about the features of the MD, the location, the access

rights and the activities of the user (Carrillo, 2007).

This proposition of query enrichment is based on

several factors that can influence the routing (in the

sense that the sources are the addressees of the

queries and that they can change): i) Changes of the

user’s location can produce changes in terms of

access and information needs (Thilliez et al., 2004).

We believe that this aspect is also sometimes valid

when there is a device change; ii) User preferences

are context-aware (Carrillo, 2007). When the query

is submitted, all context changes involve potential

consequences on the query, and consequently on the

sources. iii) The technical constraints of the MD of a

nomadic user can give rise to problems of

information display, which are difficult to anticipate.

125

Carrillo-Ramos A., Villanova-Oliver M., Gensel J., Martin H. and Torres M. (2008).

QUERY ROUTING IN NOMADIC ENVIRONMENTS WITH PUMAS.

In Proceedings of the Fourth International Conference on Web Information Systems and Technologies, pages 125-130

DOI: 10.5220/0001513401250130

 SciTePress

In order to considerate these factors, we propose

PUMAS (Peer Ubiquitous Multi-Agent System), an

agent based framework that aims at providing a

nomadic user with adapted information according to

her preferences and to the context of use. PUMAS

also offers means for questioning several sources,

which correspond to Information Systems (IS)

executed on servers, or simple files stored on other

MD. PUMAS agents perform the enrichment

mechanism of the initial query by adding criteria

taking into account this way both user preferences

and context of use. This phase of query enrichment

leads to a routing process.

In PUMAS, users communicate with each other

by means of a hybrid P2P architecture. We have

followed the recommendations of (Shizuka et al.,

2004) which consider that the use of Hybrid P2P

architectures is appropriated for: i) the prevention of

security problems related to agent mobility; ii) point

to point or broadcast communication among agents;

iii) management of agent state and provided services

(e.g., “connected”, “disconnected”, etc.). The main

characteristics of the hybrid P2P systems in PUMAS

are (Carrillo et al., 2005): i) a MD can communicate

with a specific IS by giving to the Router Agent (RA)

the ID of the required IS as a parameter in the query.

The RA then transmits the query to this IS

(communication from agent to agent).; ii) an agent is

autonomous to connect/disconnect from PUMAS

whenever the agent requires it.

PUMAS is composed of four Multi-Agent

Systems (MAS) in charge of connection,

communication, information sources management

and information adaptation. In a general way, the

MAS are similar to P2P systems in several points:

each agent is a peer in the sense that it can perform

its own tasks independently of the server and of the

other agents. P2P systems (Röhm et al., 2000) are

characterized by: i) direct communication among

peers without communication through a specific

server; ii) the autonomy of a peer for performing its

assigned tasks. Considering a P2P approach, a MAS

must represent the required knowledge for each

agent to perform the different roles that it can

assume (e.g., client, server) (Panti et al., 2002).

This paper is organized as follows: Section 2

defines the query routing process in P2P systems.

Section 3 depicts a brief PUMAS framework

overview and is dedicated to the query routing

process in PUMAS. In particular, we present the

algorithms and the knowledge managed in the

activities of this process: the query analysis and its

redirection to the sources able to answer it. Finally,

we present the conclusion and future work.

2 QUERY ROUTING PROCESS IN

P2P SYSTEMS

(Xu et al., 1999) define query routing as a general

problem which is based on two main activities: i)

Query Evaluation using the most relevant sources,

and ii) Result integration of the data generated by

different sources. In order to perform these two

activities, several issues must be considered: i)

Source selection, which consists in the analysis of

the user’s query in order to determine the sources

that are able to fulfill the query. In order to resolve

this problem, the system must know the kind and

structure of the information managed by each

source; ii) Query Evaluation which is performed by

the sources selected at previous step. Since there can

be heterogeneous data representations among

different sources, the original query must be

translated into terms that the questioned sources can

understand; iii) Result Integration: since results can

be returned by different sources, they must be

integrated into a single one that will be returned to

the user.

In this paper, we focus on the first identified

problem (Source Selection). We have classified the

possible solutions to this problem into three main

categories: i) solutions that use a gathering of peers

for query answering, ii) those based on filters and

data matching between terms of a query and the data

of a source and, iii) those based on Trust and

Reputation strategies.

Considering the gathering of peers that are able

to answer to the query, (Brunkhorst et al., 2003)

propose to specify the peers associated to a super-

peer and the information managed by each peer

within the context of the P2P system based on

schemas. The super-peers form a small subset of

peers in which each peer has specific responsibilities

and the capability of gathering with others in order

to perform tasks such as query routing, mediation in

conflict resolution and teamwork. These authors

have defined a set of indexes necessary to perform

the query routing among the (super-)peers of the

system. (Kokkinidis et al., 2006) propose to gather

peers that share the same information schemas in

order to define plans for query answering in a

distributed way. In this proposal, the super-peers are

responsible for the query routing, and the peers are

responsible for handling the queries.

(Koloniari et al., 2004) define

query routing as a

mechanism for determining the location of nodes

containing documents, which can fulfill the query.

This location is based on a mechanism of document

and filter matching. The filters are specialized data

WEBIST 2008 - International Conference on Web Information Systems and Technologies

126

structures that gather general information of large

collections of documents, and which redirect the

queries only towards the nodes that contain relevant

information. Each node has two types of filters: the

local filter, which gathers general information of

documents that are stored locally in the node, and

the merged filters, which gather general information

of documents of the neighbor nodes. When a query

arrives to a node, this node verifies its local filter

and uses the merged filters in order to redirect the

query only towards the nodes whose filters

correspond to the query. (Xu et al., 1999) present

some strategies based on techniques of clustering for

the sources’ selection. These strategies are achieved

in two steps: i) construction of knowledge about the

sources that compose each cluster and about

information managed by each one. ii) Ranking of

sources: the position of each source in the ranking is

determined considering the total of the matching

scores obtained by the clusters to which the source

belongs and the clusters‘ size.

In order to optimize the query routing process,

(Agostini et al., 2004) propose to use several metrics

related to: the reliability of the sources, their

capabilities to satisfy the user’s information needs

and their reliability for delivering information to

users. Agostini et al. propose a strategy of Trust and

Reputation allowing the selection of peers that are

able to answer the query. The trust assigned to an

agent is the degree of security estimated regarding

the quality of its answers. Trust is based on its

reputation. The reputation of an agent is the opinion

that other agents have about it; this opinion is

obtained considering the results provided by an

agent to previous inquiries and additional criteria

such as the time taken to answer, the exhaustiveness

of the answers, the relevance of the answers, etc.

Reputation is a value in a qualitative or quantitative

scale calculated by each agent. A system of peer

reputation compiles, distributes and manages

information about previous behaviors of peers. The

strategy of Trust and Reputation proposed by

Agostini et al. is based on the following process: a

component named “seeker” selects the peers that are

able to answer to the query Q with the highest

probability considering established criteria (e.g. first

to answer, quickest to answer, most reliable answer,

most trusted peer, etc.). In order to decide, the seeker

creates and manages a list <p

, p

,…p

> of trusted

peers for sending the query. The list is sorted using a

decreasing trust level. The seeker’s strategy to

answer to a query is as follows: first, the seeker asks

, then p

, and so on, until it receives relevant

answers based on the established criteria. It is

important to notice that the list of trusted peers can

evolve in time.

In a nomadic environment, information required

by a user can evolve in function of the context of

use. For example, a user can ask for a list of

restaurants and for getting those located in the street

where she is and which are open at the time she

formulates the query. The techniques of query

routing presented previously do not allow the

management of this kind of context evolution. The

following section presents PUMAS, a framework,

which considers these aspects by means of

enrichment mechanisms of the initial query in order

to adapt information to the user.

3 QUERY ROUTING IN PUMAS

During an information search process in nomadic

environments, a user can be confronted with several

problems. One can mention: i) access problems

related to the characteristics of networks and of the

user’s MD; ii) a lack of adaptation of the results

considering user’s characteristics and preferences,

and technical constraints of the user’s MD; iii) a lack

of mechanisms for searching distributed information

on several sources and devices (servers or MD). In

order to resolve these problems, we have proposed

PUMAS, a framework based on agents, which

provides nomadic users with information adapted to

their characteristics and those of their MD. PUMAS

also provides means of interrogating several sources

(e.g. Information Systems, IS). In the remainder of

this paper, we focus on the querying of sources of IS

type. The PUMAS’ architecture is composed of four

MAS (their complete description are given in

(Carrillo, 2007)): i) The Connection MAS provides

mechanisms for facilitating the connection of

different types of MD to different IS. ii) The

Communication MAS assures transparent

communication between the users’ MD and the

system, and applies a display filter (with the help of

Adaptation MAS agents) in order to present

information in an adapted way, considering the

technical constraints of the MD. iii) The Information

MAS receives queries from users, redirects them to

the IS that are able to answer them, applies a content

filter (with the help of Adaptation MAS agents),

considering the user profile, and returns the results

to the Communication MAS. iv) The Adaptation

MAS communicates with agents of the three other

MAS in order to exchange information about user,

connection, MD and communication features, etc

QUERY ROUTING IN NOMADIC ENVIRONMENTS WITH PUMAS

127

In this paper, we focus on the algorithms and the

knowledge managed for the query routing process in

PUMAS, achieved by the Router Agents (RA) which

receive enriched queries. These activities are

inspired of the work of (Xu et al., 1999). When a RA

receives a query, this agent can send it to a specific

or several ISAgents (belonging to the Information

MAS), or it can split the query in sub-queries which

are sent to one or several ISAgents.

(assert (IS

(name PharmacyIS)

(agentID PharmacyISA)

(device server)

(locationD "North Hospital")

(information_items "patient

medicines" "dose" “medicines")))

Figure 1: IS Representation: Class and Fact.

The first activity of the query routing process is the

query analysis. In this activity, the RA analyzes the

complexity of the query and classifies it in “simple”

(if it can be processed by only one ISAgent) or

“complex” (if several ISAgents are required to

process it). In this case, this activity will split the

query in sub-queries. This analysis is based on the

facts related to the IS, stored in a knowledge base

managed by the RA. A fact is a knowledge piece

which describes a real world element. In order to

clarify its definition, we also show its representation

as a UML classe (see Figure 1). The RA also

analyzes the adaptation criteria of a query (e.g.,

location, user activities), the list of the addressees of

a query, etc. After this analysis, the RA decides if it

must split the query in sub-queries. Let us suppose

for example that a doctor wants to consult the results

of the medical analysis of a patient, her diet and her

prescribed medicines. We suppose that no IS can

completely answer this query. This query is

identified as “complex”, and is therefore split by the

RA in three sub-queries, concerning respectively

“medical analysis”, “diet” and “prescribed

medicines”. Such a division is based on the stored

knowledge obtained by the RA regarding the

managed information by the different IS as a JESS

fact (see Figure 1).

In order to split the query, the RA identifies the

query items [item

, item

... item

]. In our example,

item

corresponds to “medical analysis”, item

“diet” and item

to “prescribed medicines”. Then,

this agent searches for the IS which manage an

equivalent item. Two items are “equivalent” if they

are semantically equals (i.e., if both items have the

same meaning) or are equal syntactically (e.g.,

equality in strings of characters). The applications’

designers define the equivalence relationship. They

also implement our Matching Algorithm. This

algorithm generates an Information System List (ISL)

containing tuples (IS, <items managed by the IS>):

(1) Initialize ISL answering particularly to the query (ISL)

(2) nISÅ 0 ;

(3) iÅ 1 ; // index of the items of the query

(4) While i ∈ [1, n] do // index of the items of the query

(5) jÅ 1 ; // index of the IS

(6) While j ∈ [1, s] do // index of the IS known by the

(7) mÅ 0 ; // index of the information items of

(8) While m < size (list of info items managed

by IS

) do

(9) If Compare (item

,info_item

managed

by IS

) then

(10) nIS Å nIS +1 ;

(11) ISLÅ ISL ⊕ (IS

, <item

>) ;

(12) End If

(13) m Å m + 1 ;

(14) End While

(15) j Å j + 1 ;

(16) End While

(17) i Å i + 1 ;

(18) End While

(19) If nIS is equals to 0 then

(20) query_type Å "without answer"

(21) else

(22) sidÅ countDifferent (ISL) ;

(23) If sid is equals to 1 then

(24) query_type Å "simple"

(25) else // sid is upper to 1

(26) query_type Å "complex"

(27) End If

(28) End If

(29) ISL Å Compact(ISL)

First, the ISL is empty (see line (1)). Then, the

variable nIS (containing the number of IS managing

the equivalent items to those of the query, see line

(2)), is initialized. For each item of the query (item

with i ∈ [1, n], lines (3) to (18)), it searches the IS

(IS

with j ∈ [1, s]) which manage information items

equivalents to the analyzed one. The “Compare”

method verifies the equivalence of the items (see

line (9)). If they are equivalent, the algorithm

increases the variable nIS, and adds into the ISL a

tuple whose first term corresponds to an IS

and the

second one corresponds to the analyzed item (see

lines (10) and (11)). When all the query items have

been analyzed, the algorithm verifies the value of

nIS (see lines (19) to (27)). If this value is zero, it

means that there is no IS able to manage equivalent

items from the query. In this case, the query is

tagged as “simple”. In the other case, the algorithm

analyzes the ISL in order to know the number of

different IS which manage items equivalent to those

of the query, in order to characterize the query as

"complex". This number is calculated using the

+ name: String

+ Agent ID: String

+ Device: String

+ locationD: String

+ Information_items: List

+ name: String

+ Agent ID: String

+ Device: String

+ locationD: String

+ Information_items: List

WEBIST 2008 - International Conference on Web Information Systems and Technologies

128

"countDifferent" method (see line (22)). Finally, the

algorithm uses the "Compact" method which leaves

in the ISL only one tuple associated for each IS. The

second term of this tuple corresponds to all the query

items managed by the IS. We illustrate the

“countDifferent” and “Compact” methods in order to

clarify them: Let us suppose that the RA has

identified items: i

, i

and i

, and it knows the

following IS: IS

, IS

and IS

. After the

items’ comparison (see lines (4) to (19)), we suppose

that the ISL is composed of the following tuples:

(IS

, <i

>) (IS

, <i

>) (IS

, <i

>) (IS

, <i

>) (IS

, <i

(IS

, <i

>) (IS

, <i

>) (IS

, <i

>) (IS

, <i

>) (IS

, <i

(IS

, <i

>) (IS

, <i

>) (IS

, <i

>) (IS

, <i

>) (IS

, <i

The “countDifferent” method will return the value 4

because only IS

, IS

and IS

appear in the

tuples. The “Compact” method leaves only one tuple

for each IS. That is:

(IS1, <i1, i2, i3, i5>) (IS3, <i1, i2, i3, i5 >)

(IS4, <i2, i4, i5 >) (IS5, <i1, i2, i4, i5>)

Following the algorithm, we can conclude that the

query is “complex” and the ISL is composed of:

(PharmacyIS, <"prescribed medicines">)

(NutritionistIS, <"diet">)

(MedicalAnalysisLaboratory, <"medical analysis">)

An item of the query can be managed by several IS,

then it is necessary to select the most appropriated IS

for answering them. The second activity corresponds

to the selection of the IS. In this activity, a query can

be rerouted towards a specific agent or towards a

group of agents. If the addressees of a query are

known, the selection is relatively easy. Otherwise,

the RA selects the IS and composes the network of

neighbors (considering the ISL produced during the

previous activity). This process is based on the ideas

of (Yang et al., 2004) who propose an information

retrieval process in non structured P2P systems,

where each node has a data collection shared with

other nodes. When a user submits a query, her

corresponding node becomes the sender of the query

and it can send messages (including the query) to

several of its neighbors. When a neighbor receives

the query, it handles the query using first its local

information. If the node finds some results, it returns

them to the sender node. In our proposal, a peer is

named “neighbor” of other peers, if it satisfies a set

of characteristics (criteria defined in the user

preferences) such as a close location, similar

activities, roles, knowledge, colleagues working in

the same group, etc. However, characteristics are not

restricted to proximity criteria.

Since several agents can answer to the same query,

the network of neighbors can be composed of these

agents. This gathering is useful when the RA does

not have any information about the IS or when it is

the first time that this agent works with the

neighbors. In order to avoid useless, redundant or

unusable communications, and to select the most

relevant neighbors, the RA applies query’s

adaptation criteria. For example, if the criterion is

location, the network is composed of the close

neighbors; if the user queries depend on his previous

queries, the RA must redirect them to the trusted

neighbors; if there is no criterion defined, the RA

analyzes the trust level of its neighbors. The RA

associates a trust level to each neighbor, based on

previous answers applying the Trust and Reputation

strategy proposed by (Agostini et al., 2004) (see

section 2). If a query has been split in several sub-

queries in the analysis activity, the RA analyzes the

agents that can answer to each sub-query and these

agents compose the network. For each sub-query, it

is necessary to select the IS that are able to answer.

Finally, the

network is composed of the aggregation

of the different generated sub-networks for each sub-

query. The RA considers the ISL produced in the

analysis activity. In our example, this agent selects

the PharmacyIS, the one of the Medical Analysis

Laboratory and the one of the Nutritionists. These IS

are the only ones able to answer to each one of the

three sub-queries. These sub-queries are sent to the

corresponding ISAgents. We give below the example

of the sub-query Query

produced by the IS of the

medical analysis laboratory where the ISAgent is

named MedicalAnalysisLaboratoryISA (see Figure

2)). According to an equivalent principle, Query

and Query

will be sent respectively to the agents

NutritionistISA of the NutritionistIS and the

PharmacyISA of the PharmacyIS.

(assert (Query (QueryID Query1)

(UserID “Doctor John Smith”)

(ISA MedicalAnalysisLaboratoryISA”)

(IS “MedicalAnalysisLaboratoryIS”)

(required_info “Medical

analysis”)

(parameters “patient name” “date”)))

Figure 2: Query Representation: Fact and Class.

The third activity is the redirection of queries to the

IS. In this activity, once the RA has identified the

potential IS (neighbors), it must analyze the trust

level associated to each one of them, to determine a

query’s redirection protocol. This information about

trust level can be unavailable, if it is the first time

that the RA executes this query or that it works with

these IS. In the same way, the trust levels of the

neighbors can be similar. In these conditions

+ QueryID: String

+ UserID: String

+ ISA: String

+ IS: String

+ required_info: List

+ parameters: List

Query

+ QueryID: String

+ UserID: String

+ ISA: String

+ IS: String

+ required_info: List

+ parameters: List

Query

QUERY ROUTING IN NOMADIC ENVIRONMENTS WITH PUMAS

129

(absence of established trust), the RA sends the

query in “broadcast”. When the RA has information

exploitable in terms of trust, it redirects the query in

a sequential way, beginning by the most trusted

agent. The answer to the query will be the one

generated by the first agent that fulfills it. If the RA

does not receive any answer, the user will be

informed about it. If the RA only knows the

neighbors able to answer the sub-queries (query

, …,

query

), the RA sends directly the sub-queries to

these neighbors. In our example, the RA sends

Query

to the MedicalAnalysisLaboratoryISA,

Query

to the NutritionistISA and Query

to the

PharmacyISA. The RA must collect and classify the

obtained answers from the different agents and

select the most relevant ones considering established

adaptation criteria. The scenarios that present the

sending and enrichment of query and the reception

of results can be found in (Carrillo et al., 2005).

4 CONCLUSIONS

When a user formulates queries, the results can

come from different Information Sources (IS). In

this paper, we have defined a query routing process

as a mechanism which analyzes the query, and

performs a matching (semantic or syntactic) between

query’s items and IS’ items. This matching is

achieved in order to select the IS able to answer to

user queries. After identifying items and the

recognition of the IS that are able to manage

equivalent items, this process splits the query. A

Router Agent considers adaptation criteria provided

by the user (e.g. location, user’s activities,

preferences) in order to choose the most

appropriated IS for answering the query. Finally, this

process must collect and classify the query results.

We have illustrated the query routing process by

means of an example implying the different IS of a

hospital in which a doctor asks for information about

the prescribed medicines, the medical analysis and

the diet of a patient. Our future work aims the

definition of an algorithm for the collection and

analysis of results coming from one or several IS.

REFERENCES

Agostini, A., Moro, G., 2004. Identification of

Communities of Peers by Trust and Reputation. In

AIMSA 2004, 11

Int. Conf. on Artificial Intelligence:

Methodology, Systems, and Applications. Springer, p.

85-95.Brunkhorst, I., Dhraief, H., Kemper, A., Nedjl

W., Wiesner, C., 2003. Distributed Queries and Query

Optimization in Schema-Based P2P Systems. In

DBISP2P, 1

Int. Workshop on Databases, Inf.

Systems, and P2P Computing. Springer, p. 184-199.

Carrillo-Ramos, A., Gensel, J., Villanova-Oliver, M.,

Martin, H., 2005. A Peer Ubiquitous Multi-Agent

Framework for providing nomadic users with adapted

information. In AP2PC 2005, 4

Int. Workshop on

Agents and P2P Computing. Springer, p. 159-172.

Carrillo-Ramos, A., 2007. Agents ubiquitaires pour un

accès adapté aux systèmes of information : Le

Framework PUMAS. Doctoral Thesis of the

University Joseph Fourier, Grenoble, France. (March

, 2007)

Kokkinidis, G., Lefteris, S., Christophides, V., 2006.

Query Processing in RDF/S-Based P2P Database

Systems. Semantic Web and P2P. Springer, p. 59-87.

Koloniari, G., Pitoura, E., 2004. Content-Based Routing of

Path Queries in Peer-to-Peer Systems. In EDBT 2004,

Int. Conf. on Extending DB Technology. Springer,

p. 29-47.

Panti, M., Penserini, L., Spalazzi, L., 2002. A Multi-Agent

System based on the P2P model to Information

Integration. In AAMAS 2002, 1

Int. Conf. on

Autonomous Agents and Multi-Agent Systems.

http://www.agentcities.org/EUNET/Projects/acnet_pro

j_38.pdf (Last Access: September 2006)

Röhm, U., Böhm, K., Schek, H., 2000. OLAP Query

Routing and Physical Design in a Database Cluster. In

EDBT 2000, 7

Conference on Extending Database

Technology. Springer, p. 254-268.

Shizuka, M., MA, J., Lee, J., Miyoshi, Y., Takata, K.,

2004. A P2P Ubiquitous System for Testing Network

Programs. In EUC 2004, Embedded and Ubiquitous

Computing. Springer, p. 1004-1013.

Thilliez M., Delot T., 2004. Evaluating Location

Dependent Queries Using ISLANDS. In ISSADS

2004, Symposium on Advanced Distributed Systems.

Springer, p. 126-136.

Xu, J., Lim, E., Ng, W.K., 1999. Cluster-Based Database

Selection Techniques for Routing Bibliographic

Queries. In DEXA 99, 10

Int. Conf. on Database and

Expert Systems Applications. Springer, p.100-109.

Yang, D., Xu, L., Cai, W., Zhou, S., Zhou, A., 2004.

Efficient Query Routing for XML Documents

Retrieval in Unstructured Peer to Peer Networks. In

APWeb 2004. Springer, p. 217-22

WEBIST 2008 - International Conference on Web Information Systems and Technologies

130