The Semantic Grid: A Future e-Science Infrastructure

David De Roure, Nicholas R. Jennings and Nigel R. Shadbolt1

Dept of Electronics and Computer Science,
University of Southampton,
Southampton SO17 1BJ, UK

{dder,nrj,nrs}@ecs.soton.ac.uk

Abstract
e-Science offers a promising vision of how computer and communication technology can
support and enhance the scientific process. It does this by enabling scientists to generate,
analyse, share and discuss their insights, experiments and results in an effective manner. The
underlying computer infrastructure that provides these facilities is commonly referred to as
the Grid. At this time, there are a number of grid applications being developed and there is a
whole raft of computer technologies that provide fragments of the necessary functionality.
However there is currently a major gap between these endeavours and the vision of e-Science
in which there is a high degree of easy-to-use and seamless automation and in which there are
flexible collaborations and computations on a global scale. To bridge this practice–aspiration
divide, this paper presents a research agenda whose aim is to move from the current state of
the art in e-Science infrastructure, to the future infrastructure that is needed to support the full
richness of the e-Science vision. Here the future e-Science research infrastructure is termed
the Semantic Grid (Semantic Grid to Grid is meant to connote a similar relationship to the one
that exists between the Semantic Web and the Web). In particular, we present a conceptual
architecture for the Semantic Grid. This architecture adopts a service-oriented perspective in
which distinct stakeholders in the scientific process, represented as software agents, provide
services to one another, under various service level agreements, in various forms of
marketplace. We then focus predominantly on the issues concerned with the way that
knowledge is acquired and used in such environments since we believe this is the key
differentiator between current grid endeavours and those envisioned for the Semantic Grid.

1. Introduction
Scientific research and development has always involved large numbers of people,
with different types and levels of expertise, working in a variety of roles, both
separately and together, making use of and extending the body of knowledge. In
recent years, however, there have been a number of important changes in the nature
and the process of research. In particular, there is an increased emphasis on
collaboration between large teams, an increased use of advanced information
processing techniques, and an increased need to share results and observations
between participants who are not physically co-located. When taken together, these
trends mean that researchers are increasingly relying on computer and communication
technologies as an intrinsic part of their everyday research activity. At present, the key
communication technologies are predominantly email and the Web. Together these
have shown a glimpse of what is possible; however to more fully support the e-
Scientist, the next generation of technology will need to be much richer, more flexible
and much easier to use. Against this background, this paper focuses on the

1 The authors are listed alphabetically.

The Semantic Grid
requirements, the design and implementation issues, and the research challenges
associated with developing a computing infrastructure to support future e-Science.

The computing infrastructure for e-Science is commonly referred to as the Grid
[Foster98] and this is, therefore, the term we will use here. This terminology is chosen
to connote the idea of a ‘power grid’: namely that e-Scientists can plug into the e-
Science computing infrastructure like plugging into a power grid. An important point
to note however is that the term ‘grid’ is sometimes used synonymously with a
networked, high performance computing infrastructure. While this aspect is certainly
an important enabling technology for future e-Science, it is only a part of a much
larger picture that also includes information handling and support for knowledge
processing within the e-scientific process. It is this broader view of the e-Science
infrastructure that we adopt in this document and we refer to this as the Semantic Grid
[DeRoure2001]. Our view is that as the Grid is to the Web, so the Semantic Grid is to
the Semantic Web [BernersLee99, BernersLee01]. Thus the Semantic Grid is
characterised as an open system in which users, software components and
computational resources (all owned by different stakeholders) come and go on a
continual basis. There should be a high degree of automation that supports flexible
collaborations and computation on a global scale. Moreover, this environment should
be personalised to the individual participants and should offer seamless interactions
with both software components and other relevant users2.

The grid metaphor intuitively gives rise to the view of the e-Science infrastructure as
a set of services that are provided by particular individuals or institutions for
consumption by others. Given this, and coupled with the fact that many research and
standards activities are embracing a similar view (e.g., [WebServices01]), we adopt a
service-oriented view of the Grid throughout this document (see section 2 for a more
detailed justification of this choice). This view is based upon the notion of various
entities (represented as software agents) providing services to one another under
various forms of contract (or service level agreement) in various forms of
marketplace.

Given the above view of the scope of e-Science, it has become popular to characterise
the computing infrastructure as consisting of three conceptual layers3:

• Data/computation
This layer deals with the way that computational resources are allocated,
scheduled and executed and the way in which data is shipped between the
various processing resources. It is characterised as being able to deal with
large volumes of data, providing fast networks and presenting diverse
resources as a single metacomputer. The data/computation layer builds on the
physical ‘grid fabric’, i.e. the underlying network and computer infrastructure,

2 Our view of the Semantic Grid has many elements in common with the notion of a ‘collaboratory’
[Cerf93]: a centre without walls, in which researchers can perform their research without regard to
geographical location - interacting with colleagues, accessing instrumentation, sharing data and
computational resource, and accessing information in digital libraries. We extend this view to
accommodate ‘information appliances’ in the laboratory setting, which might, for example, include
electronic logbooks and other portable devices.
3 The three layer grid vision is attributed to Keith G. Jeffery of CLRC, who introduced it in a paper for
the UK Research Councils Strategic Review in 1999.
2

The Semantic Grid
which may also interconnect scientific equipment. Here data is understood as
uninterpreted bits and bytes.

• Information
This layer deals with the way that information is represented, stored, accessed,
shared and maintained. Here information is understood as data equipped with
meaning. For example the characterisation of an integer as representing the
temperature of a reaction process, the recognition that a string is the name of
an individual.

• Knowledge
This layer is concerned with the way that knowledge is acquired, used,
retrieved, published and maintained to assist e-Scientists to achieve their
particular goals and objectives. Here knowledge is understood as information
applied to achieve a goal, solve a problem or enact a decision. In the Business
Intelligence literature knowledge is often defined as actionable information.
For example, the recognition by a plant operator that in the current context a
reaction temperature demands shutdown of the process.

There are a number of observations and remarks that need to be made about this
layered structure. Firstly, all grids that have or will be built have some element of all
three layers in them. The degree to which the various layers are important and utilised
in a given application will be domain dependent - thus, in some cases, the processing
of huge volumes of data will be the dominant concern, while in others the knowledge
services that are available will be the overriding issue. Secondly, this layering is a
conceptual view on the system that is useful in the analysis and design phases of
development. However, the strict layering may not be carried forward to the
implementation for reasons of efficiency. Thirdly, the service-oriented view applies at
all the layers. Thus there are services, producers, consumers and contracts at the
computational layer, at the information layer and at the knowledge layer (figure 1).


Information
services
Data/Computation
services
Knowledge services
E-Scientist’s Environment









Figure 1: Three layered architecture viewed as services

Although this view is widely accepted, to date most research and development work
in this area has concentrated on the data/computation layer and on the information
layer. While there are still many open problems concerned with managing massively
distributed computations in an efficient manner and in accessing and sharing
information from heterogeneous sources (see the companion paper [DeRoure02] for
3

The Semantic Grid
more details), we believe the full potential of grid computing can only be realised by
fully exploiting the functionality and capabilities provided by knowledge layer
services. This is because it is at this layer that the reasoning necessary for seamlessly
automating a significant range of the actions and interactions takes place. Thus this is
the area we focus on most in this paper.

The remainder of this paper is structured in the following manner. Section 2 provides
a motivating scenario of our vision for the Semantic Grid. Section 3 provides a
justification of the service-oriented view for the Semantic Grid. Section 4
concentrates on knowledge services. Section 5 concludes by presenting the main
research challenges that need to be addressed to make the Semantic Grid a reality.

2. A Semantic Grid Scenario
To help clarify our vision of the Semantic Grid, we present a motivating scenario that
captures what we believe are the key characteristics and requirements of future e-
Science environments. We believe this is more instructive than trying to produce an
all embracing definition.

This scenario is derived from talking with e-Scientists across several domains
including the physical sciences. It is not intended to be domain-specific (since this
would be too narrow) and at the same time it cannot be completely generic (since this
would not be detailed enough to serve as a basis for grounding our discussion). Thus
it falls somewhere in between. Nor is the scenario science fiction – these practices
exist today, but on a restricted scale and with a limited degree of automation. The
scenario itself (figure 2) fits with the description of grid applications as “coordinated
resource sharing and problem solving among dynamic collections of individuals”
[Foster01].

The sample arrives for analysis with an ID number. The technician logs it
into the database and the information about the sample appears (it had been
entered remotely when the sample was taken). The appropriate settings are
confirmed and the sample is placed with the others going to the analyser (a
piece of laboratory equipment). The analyser runs automatically and the
output of the analysis is stored together with a record of the parameters and
laboratory conditions at the time of analysis.

The analysis is automatically brought to the attention of the company
scientist who routinely inspects analysis results such as these. The scientist
reviews the results from their remote office and decides the sample needs
further investigation. They request a booking to use the High Resolution
Analyser and the system presents configurations for previous runs on
similar samples; given this previous experience the scientist selects
appropriate parameters. Prior to the booking, the sample is taken to the
analyser and the equipment recognizes the sample identification. The
sample is placed in the equipment which configures appropriately, the door
is locked and the experiment is monitored by the technician by live video
then left to run overnight; the video is also recorded, along with live data
from the equipment. The scientist is sent a URL to the results.

4

The Semantic Grid
Later the scientist looks at the results and, intrigued, decides to replay the
analyser run, navigating the video and associated information. They then
press the “query” button and the system summarises previous related
analyses reported internally and externally, and recommends other scientists
who have published work in this area. The scientist finds that their results
appear to be unique.

The scientist requests an agenda item at the next research videoconference
and publishes the experimental information for access by their colleagues
(only) in preparation for the meeting. The meeting decides to make the
analysis available for the wider community to look at, so the scientist then
logs the analysis and associated metadata into an international database and
provides some covering information. Its provenance is recorded. The
availability of the new information prompts other automatic processing and
a number of databases are updated; some processing of this new
information occurs.

Various scientists who had expressed interest in samples or analyses fitting
this description are notified automatically. One of them decides to run a
simulation to see if they can model the sample, using remote resources and
visualizing the result locally. The simulation involves the use of a problem
solving environment (PSE) within which to assemble a range of
components to explore the issues and questions that arise for the scientist.
The parameters and results of the simulations are made available via the
public database. Another scientist adds annotation to the published
information.

10
Analysis
Simulation
Video
H
iR
es

An
al
ys
er
Public
Database
An
al
ys
er
Pr
iv
at
e
D
at
ab
as
e
Sa
m
pl
e
D
at
ab
as
e
Knowledge
services:
Annotation,
Publication

Figure 2: Workflow in the scenario

5

The Semantic Grid
This scenario draws out a number of underlying assumptions and raises a number of
requirements that we believe are broadly applicable to a range of e-Science
applications:

• Storage. It is important that the system is able to store and process potentially
huge volumes of content in a timely and efficient fashion.
• Ownership. Different stakeholders need to be able to retain ownership of their
own content and processing capabilities, but there is also a need to allow
others access under the appropriate terms and conditions.
• Provenance. Sufficient information is stored so that it is possible to repeat the
experiment, re-use the results, or provide evidence that this information was
produced at this time (the latter may involve a third party).
• Transparency. Users need to be able to discover, transparently access and
process relevant content wherever it may be located in the Grid.
• Communities. Users should be able to form, maintain and disband
communities of practice with restricted membership criteria and rules of
operation.
• Fusion. Content needs to be able to be combined from multiple sources in
unpredictable ways according to the users’ needs; descriptions of the sources
and content will be used to combine content meaningfully.
• Conferencing. Sometimes it is useful to see the other members of the
conference, and sometimes it is useful to see the artefacts and visualisations
under discussion.
• Annotation. From logging the sample through to publishing the analysis, it is
necessary to have annotations that enrich the description of any digital content.
This meta-content may apply to data, information or knowledge and depends
on agreed interpretations.
• Workflow. To support the process enactment and automation, the system needs
descriptions of processes. The scenario illustrates workflow both inside and
outside the company.
• Notification. The arrival of new information prompts notifications to users and
initiates automatic processing.
• Decision support. The technicians and scientists are provided with relevant
information and suggestions for the task at hand.
• Resource reservation. There is a need to ease the process of resource
reservation. This applies to experimental equipment, collaboration (the
conference), and resource scheduling for the simulation.
• Security. There are authentication, encryption and privacy requirements, with
multiple organisations involved, and a requirement for these to be handled
with minimal manual intervention.
• Reliability. The systems appear to be reliable but in practice there may be
failures and exception handling at various levels, including the workflow.
• Video. Both live and stored video have a role, especially where the video is
enriched by associated temporal metacontent (in this case to aid navigation).
• Smart laboratory. For example, the equipment detects the sample (e.g. by
barcode or RFID tag), the scientist may use portable devices for note-taking,
and visualisations may be available in the lab.
• Knowledge. Knowledge services are an integral part of the e-Science process.
Examples include: finding papers, finding people, finding previous
6

The Semantic Grid
experimental design (these queries may involve inference), annotating the
uploaded analysis, and configuring the lab to the person.
• Growth. The system should support evolutionary growth as new content and
processing techniques become available.
• Scale. The scale of the scientific collaboration increases through the scenario,
as does the scale of computation, bandwidth, storage and complexity of
relationships between information.

3. A Service-Oriented View
This section expands upon the view of the Semantic Grid as a service-oriented
architecture in which entities provide services to one another under various forms of
contract4. Thus, as shown in figure 1, the e-Scientist’s environment is composed of
data/computation services, information services, and knowledge services. However,
before we deal with the specifics of each of these different types of service, it is
important to highlight those aspects that are common since this provides the
conceptual basis and rationale for what follows. To this end, section 3.1 provides the
justification for a service-oriented view of the different layers of the Semantic Grid.
Section 3.2 then addresses the technical ramifications of this choice and outlines the
key technical challenges that need to be overcome to make service-oriented grids a
reality. The section concludes (section 3.3) with the e-Science scenario of section 2
expressed in a service-oriented architecture.

3.1 Justification of a Service-Oriented View
Given the set of desiderata and requirements from section 2, a key question in
designing and building grid applications is what is the most appropriate conceptual
model for the system? The purpose of such a model is to identify the key constituent
components (abstractions) and specify how they are related to one another. Such a
model is necessary to identify generic grid technologies and to ensure that there can
be re-use between different grid applications. Without a conceptual underpinning, grid
endeavours will simply be a series of handcrafted and ad hoc implementations that
represent point solutions.

To this end, an increasingly common way of viewing many large systems (from
governments, to businesses, to computer systems) is in terms of the services that they
provide. Here a service can simply be viewed as an abstract characterization and
encapsulation of some content or processing capabilities. For example, potential
services in our exemplar scenario could be: the equipment automatically recognising
the sample and configuring itself appropriately, the logging of information about a
sample in the international database, the setting up of a video to monitor the
experiment, the locating of appropriate computational resources to support a run of
the High Resolution Analyser, the finding of all scientists who have published work
on experiments similar to those uncovered by our e-Scientist, and the analyser raising
an alert whenever a particular pattern of results occurs (see section 3.3 for more

4 This view pre-dates the work of Foster et al on the Open Services Grid Architecture [Foster02]. While
Foster’s proposal has many similarities with our view he does not deal with issues associated with
developing services through autonomous agents, with the issue of dynamically forming service level
agreements, nor with the design of marketplaces in which the agents trade their services.
7

The Semantic Grid
details). Thus, services can be related to the domain of the Grid, the infrastructure of
the computing facility, or the users of the Grid – i.e., at the data/computation layer, at
the information layer, or at the knowledge layer (as per figure 1). In all of these cases,
however, it is assumed that there may be multiple versions of broadly the same
service present in the system.

Services do not exist in a vacuum, rather they exist in a particular institutional
context. Thus all services have an owner (or set of owners). The owner is the body
(individual or institution) that is responsible for offering the service for consumption
by others. The owner sets the terms and conditions under which the service can be
accessed. Thus, for example, the owner may decide to make the service universally
available and free to all on a first-come, first-served basis. Alternatively, the owner
may decide to limit access to particular classes of users, to charge a fee for access and
to have priority-based access. All options between these two extremes are also
possible. It is assumed that in a given system there will be multiple service owners
(each representing a different stakeholder) and that a given service owner may offer
multiple services. These services may correspond to genuinely different functionality
or they may vary in the way that broadly the same functionality is delivered (e.g.,
there may be a quick and approximate version of the service and one that is more time
consuming and accurate).

In offering a service for consumption by others, the owner is hoping that it will indeed
attract consumers for the service. These consumers are the entities that decide to try
and invoke the service. The purpose for which this invocation is required is not of
concern here: it may be for their own private use, it may be to resell onto others, or it
may be to combine with other services.

The relationship between service owner and service consumer is codified through a
service contract. This contract specifies the terms and conditions under which the
owner agrees to provide the service to the consumer. The precise structure of the
contract will depend upon the nature of the service and the relationship between the
owner and the provider. However examples of relevant attributes include the price for
invoking the service, the information the consumer has to provide to the provider, the
expected output from the service, an indication about when this output can be
expected, and the penalty for failing to deliver according to the contract. Service
contracts can either be established by an off-line or an on-line process depending on
the prevailing context.

The service owners and service producers interact with one another in a particular
environmental context. This environment may be common to all entities in the Grid
(meaning that all entities offer their services in an entirely open marketplace). In other
cases, however, the environment may be closed and entrance may be controlled
(meaning that the entities form a private club).5 In what follows, a particular
environment will be called a marketplace and the entity that establishes and runs the
marketplace will be termed the market owner. The rationale for allowing individual
marketplaces to be defined is that they offer the opportunity to embed interactions in
an environment that has its own set of rules (both for membership and ongoing

5 This is analogous to the notion of having a virtual private network overlaid on top of the Internet. The
Internet corresponds to the open marketplace in which anybody can participate and the virtual private
network corresponds to a closed club that can interact under its own rules.
8

The Semantic Grid
operation) and they allow the entities to make stronger assumptions about the parties
with which they interact (e.g., the entities may be more trustworthy or cooperative
since they are part of the same club). Such marketplaces may be appropriate, for
example, if the nature of the domain means that the services are particularly sensitive
or valuable. In such cases, the closed nature of the marketplace will enable the entities
to interact more freely because of the rules of membership.

To summarise, the key components of a service-oriented architecture are as follows
(figure 3): service owners (rounded rectangles) that offer services (filled circles) to
service consumers (filled triangles) under particular contracts (solid links between
producers and consumers). Each owner-consumer interaction takes place in a given
marketplace (denoted by ovals) whose rules are set by the market owner (filled cross).
The market owner may be one of the entities in the marketplace (either a producer or
a consumer) or it may be a neutral third party.



service

consumer

market owner

service contract
service owner1
service owner2
marketplace3
marketplace2
marketplace1
e-Science Infrastructure
service owner3
Figure 3: Service-oriented architecture: key components



























Given the central role played by the notion of a service, it is natural to explain the
operation of the system in terms of a service lifecycle (figure 4). The first step is for a
service owner to define a service they wish to make available to others. The reasons
for wanting to make a service available may be many and varied – ranging from
altruism, through necessity, to commercial benefit. It is envisaged that in a given grid
application all three motivations (and many others besides) are likely to be present,
9

The Semantic Grid
although perhaps to varying degrees that are dictated by the nature of the domain.
Service creation should be seen as an ongoing activity. Thus new services may come
into the environment at any time and existing ones may be removed (service
decommissioning) at any time. This means the system is in a state of continual flux
and never reaches a steady state. Creation is also an activity that can be automated to a
greater or lesser extent. Thus, in some cases, all services may be put together in an
entirely manual fashion. In other cases, however, there may be a significant automated
component. For example, it may be decided that a number of services should be
combined; either to offer a new service (if the services are complementary in nature)
or to alter the ownership structure (if the services are similar). In such cases, it may be
appropriate to automate the processes of finding appropriate service providers and of
getting them to agree to new terms of operation. This dynamic service composition
activity is akin to creating a new virtual organisation: a number of initially distinct
entities can come together, under a set of operating conditions, to form a new entity
that offers a new service. This grouping will then stay in place until it is no longer
appropriate to remain in this form, whereupon it will disband.














Define how service
is to be realised

CREATION
Re-negotiation


The service creation process
how the service is to be real
description language. These
consumer (i.e., they are encap
meta-information associated
which the service can be pro
the service and what are the
the service available in the ap
advertising and registration fa

The service procurement pha
service owner and a service c
service according to a particu
points to note about this proc
service owner may be unabl
Secondly, in most cases, the
different and autonomous st
established will be some for
Establish contract between cording
owner and consumer
PROCUREMENT ENACTMENT
Contract results
Establish contract
Figure 4: Service lifecycle
covers three broad types of a
ized by the service owner usi
details are not available e
sulated by the service owner)
with the service. This indica
cured. This meta-information
likely contract options for pro
propriate marketplace. This re
cilities to be available in the m
se is situated in a particular m
onsumer establishing a contrac
lar set of terms and condition
ess. Firstly, it may fail. That
e or unwilling to provide the
service owner and the service
akeholders. Thus the process
m of negotiation – since the
10 Enact service ac
to contract ctivity. Firstly, specifying
ng an appropriate service
xternally to the service
. Secondly, specifying the
tes the potential ways in
indicates who can access
curing it. Thirdly, making
quires appropriate service
arketplace.
arketplace and involves a
t for the enactment of the
s. There are a number of
is, for whatever reason, a
service to the consumer.
consumer will represent
by which contracts are
entities involved need to

The Semantic Grid
come to a mutually acceptable agreement on the matter. If the negotiation is
successful (i.e., both parties come to an agreement) then the outcome of the
procurement is a contract between the service owner and the service consumer.
Thirdly, this negotiation may be carried out off-line by the respective service owners
or it may be carried out at run-time. In the latter case, the negotiation may be
automated to a greater or lesser extent – varying from the system merely
automatically flagging the fact that a new service contract needs to be established to
automating the entire negotiation process6.

The final stage of the service lifecycle is service enactment. Thus, after having
established a service contract, the service owner has to undertake the necessary
actions in order to fulfil its obligations as specified in the contract. After these actions
have been performed, the owner needs to fulfil its reporting obligations to the
consumer with respect to the service. This may range from a simple inform indicating
that the service has been completed, to reporting back complex content that represents
the results of performing the service. The above assumes that the service owner is
always able to honour the contracts that it establishes. However, in some cases the
owner may not be able to stick to the terms specified in the contract. In such cases, it
may have to renegotiate the terms and conditions of the contract; paying any penalties
that are due. This enforcement activity is undertaken by the market owner and will be
covered by the terms and conditions that the service providers and consumers sign up
to when they enter into the marketplace.

Having described the key components of the service-oriented approach, we return to
the key system-oriented desiderata noted in section 2. From the above discussion, it
can be seen that a service-oriented architecture is well suited to grid applications:

• able to store and process huge volumes of content in a timely fashion;
o The service-oriented model offers a uniform means of describing and
encapsulating activities at all layers in the Grid. This model then needs
to be underpinned by the appropriate processing and communication
infrastructure to ensure it can deliver the desired performance.

• allow different stakeholders to retain ownership of their own content and
processing capabilities, but to allow others access under the appropriate terms
and conditions;
o Each service owner retains control over the services that they make
available to others. They determine how the service is realized and set
the policy for accessing the service.

• allow users to discover, transparently access and process relevant content
wherever it may be located in the Grid;
o The overall system is simply viewed as a number of service
marketplaces. Any physical distribution and access problems are
masked via the service interface and the service contract. The

6 Automated negotiation technology is now widely used in many e-commerce applications
[Guttman98]. It encompasses various forms of auctions (a one-to-many form of negotiation) as well as
bi-lateral negotiations. Depending on the negotiation protocol that is in place, the negotiation can be
concluded in a single round or it may last for many rounds. Thus negotiation need not be a lengthy
process; despite the connotation from human interactions that it may be!
11

The Semantic Grid
marketplace itself has advertisement and brokering mechanisms to
ensure appropriate service owners and consumers are put together.

• allow users to form, maintain and disband communities of practice with
restricted membership criteria and rules of operation;
o Each community can establish its own marketplace. The marketplace
owner defines the conditions that have to be fulfilled before entities can
enter, defines the rules of interaction for once the marketplace is
operational, and enforces the rules through appropriate monitoring.

• allow content to be combined from multiple sources in unpredictable ways
according to the users’ needs;
o It is impossible to a priori predict how the users of a system will want
to combine the various services contained within it. Thus services must
be able to be composed in flexible ways. This is achieved by
negotiation of appropriate contracts. This composition can be done on
a one-off basis or may represent a more permanent binding into a new
service that is offered on an ongoing basis (as in the establishment of a
new virtual organisation).

• support evolutionary growth as new content and processing techniques
become available.
o Services represent the unit of extension of the system. Thus as new
content or processing techniques become available they are simply
represented as new services and placed in a marketplace(s). Also new
marketplaces can be added as new communities of practice emerge.

3.2 Key Technical Challenges
The previous section outlined the service-oriented view of the Semantic Grid.
Building upon this, this section identifies the key technical challenges that need to be
overcome to make such architectures a reality. To this end, table 1 represents the key
functionality of the various components of the service-oriented architecture, each of
which is then described in more detail in the remainder of this section.

Service Owner Service Consumer Marketplace
Service creation Service discovery Owner and consumer
registration
Service advertisement Service registration
Service contract creation Service contract creation Policy specification
Service delivery Service result reception Policy monitoring and
enforcement
Table 1: Key functions of the service-oriented architecture components

3.2.1 Service Owners and Consumers as Autonomous Agents
A natural way to conceptualise the service owners and the service consumers are as
autonomous agents. Although there is still some debate about exactly what constitutes
12

The Semantic Grid
agenthood, an increasing number of researchers find the following characterisation
useful [Wooldridge97]:

an agent is an encapsulated computer system that is situated in some
environment and that is capable of flexible, autonomous action in that
environment in order to meet its design objectives

There are a number of points about this definition that require further explanation.
Agents are [Jennings00]: (i) clearly identifiable problem solving entities with well-
defined boundaries and interfaces; (ii) situated (embedded) in a particular
environment—they receive inputs related to the state of their environment through
sensors and they act on the environment through effectors; (iii) designed to fulfill a
specific purpose—they have particular objectives (goals) to achieve; (iv)
autonomous— they have control both over their internal state and over their own
behaviour7; (v) capable of exhibiting flexible problem solving behaviour in pursuit of
their design objectives—they need to be both reactive (able to respond in a timely
fashion to changes that occur in their environment) and proactive (able to act in
anticipation of future goals) .

Thus, each service owner will have one or more agents acting on its behalf. These
agents will manage access to the services for which they are responsible and will
ensure that the agreed contracts are fulfilled. This latter activity involves the
scheduling of local activities according to the available resources and ensuring that
the appropriate results from the service are delivered according to the contract in
place. Agents will also act on behalf of the service consumers. Depending on the
desired degree of automation, this may involve locating appropriate services, agreeing
contracts for their provision, and receiving and presenting any received results.

3.2.2 Interacting Agents
Grid applications involve multiple stakeholders interacting with one another in order
to procure and deliver services. Underpinning the agents’ interactions is the notion
that they need to be able to inter-operate in a meaningful way. Such semantic
interoperation is difficult to obtain in grids (and all other open systems) because the
different agents will typically have their own individual information models.
Moreover, the agents may have a different communication language for conveying
their own individual terms. Thus, meaningful interaction requires mechanisms by
which this basic interoperation can be effected (see section 4.2 for more details).

Once semantic inter-operation has been achieved, the agents can engage in various
forms of interaction. These interactions can vary from simple information
interchanges, to requests for particular actions to be performed and on to cooperation,
coordination and negotiation in order to arrange interdependent activities. In all of
these cases, however, there are two points that qualitatively differentiate agent

7 Having control over their own behaviour is one of the characteristics that distinguishes agents from
objects. Although objects encapsulate state and behaviour (more accurately behaviour realisation), they
fail to encapsulate behaviour activation or action choice. Thus, any object can invoke any publicly
accessible method on any other object at any time. Once the method is invoked, the corresponding
actions are performed. In this sense, objects are totally obedient to one another and do not have
autonomy over their choice of action.
13

The Semantic Grid
interactions from those that occur in other computational models. Firstly, agent-
oriented interactions are conceptualised as taking place at the knowledge level
[Newell82]. That is, they are conceived in terms of which goals should be followed, at
what time, and by whom. Secondly, as agents are flexible problem solvers, operating
in an environment over which they have only partial control and observability,
interactions need to be handled in a similarly flexible manner. Thus, agents need the
computational apparatus to make run-time decisions about the nature and scope of
their interactions and to initiate (and respond to) interactions that were not foreseen at
design time (cf. the hard-wired engineering of such interactions in extant approaches).

The subsequent discussion details what would be involved if all these interactions
were to be automated and performed at run-time. This is clearly the most technically
challenging scenario and there are a number of points that need to be made. Firstly,
while such automation is technically feasible, in a limited form, using today’s
technology, this is an area that requires more research to reach the desired degree of
sophistication and maturity. Secondly, in some cases, the service owners and
consumers may not wish to automate all of these activities since they may wish to
retain a degree of human control over these decisions. Thirdly, some contracts and
relationships may be set up at design time rather than being established at run-time.
This can occur when there are well-known links and dependencies between particular
services, owners and consumers.

The nature of the interactions between the agents can be broadly divided into two
main camps. Firstly, those that are associated with making service contracts. This will
typically be achieved through some form of automated negotiation since the agents
are autonomous [Jennings01]. When designing these negotiations, three main issues
need to be considered:

• The Negotiation Protocol: the set of rules that govern the interaction. This
covers the permissible types of participants (e.g. the negotiators and any
relevant third parties), the negotiation states (e.g. accepting bids, negotiation
closed), the events that cause negotiation states to change (e.g. no more
bidders, bid accepted) and the valid actions of the participants in particular
states (e.g. which messages can be sent by whom, to whom, at what stage).
• The Negotiation Object: the range of issues over which agreement must be
reached. At one extreme, the object may contain a single issue (such as price),
while on the other hand it may cover hundreds of issues (related to price,
quality, timings, penalties, terms and conditions, etc.). Orthogonal to the
agreement structure, and determined by the negotiation protocol, is the issue of
the types of operation that can be performed on agreements. In the simplest
case, the structure and the contents of the agreement are fixed and participants
can either accept or reject it (i.e. a take it or leave it offer). At the next level,
participants have the flexibility to change the values of the issues in the
negotiation object (i.e. they can make counter-proposals to ensure the
agreement better fits their negotiation objectives). Finally, participants might
be allowed to dynamically alter (by adding or removing issues) the structure of
the negotiation object (e.g. a car salesman may offer one year’s free insurance
in order to clinch the deal).
14

The Semantic Grid
• The Agent’s Decision Making Models: the decision-making apparatus the
participants so as to act in line with the negotiation protocol in order to
achieve their objectives. The sophistication of the model, as well as the range
of decisions that have to be made, are influenced by the protocol in place, by
the nature of the negotiation object, and by the range of operations that can be
performed on it. It can vary from the very simple, to the very complex.
In designing any automated negotiation system the first thing that needs to be
established is the protocol to be used (this is called the mechanism design problem). In
this context, this will be determined by the market owner. Here the main consideration
is the nature of the negotiation. If it is a one-to-many negotiation (i.e., one buyer and
many sellers or one seller and many buyers) then the protocol will typically be a form
of auction. Although there are thousands of different permutations of auction, four
main ones are typically used. These are: English, Dutch, Vickrey, and First-Price
Sealed Bid. In an English auction, the auctioneer begins with the lowest acceptable
price and bidders are free to raise their bids successively until there are no more offers
to raise the bid. The winning bidder is the one with the highest bid. The Dutch auction
is the converse of the English one; the auctioneer calls for an initial high price, which
is then lowered progressively until there is an offer from a bidder to claim the item. In
the first-priced sealed bid, each bidder submits their offer for the item independently
without any knowledge of the other bids. The highest bidder gets the item and they
pay a price equal to their bid amount. Finally, a Vickrey auction is similar to a first-
price sealed bid auction, but the item is awarded to the highest bidder at a price equal
to the second highest bid. More complex forms of auctions exist to deal with the cases
in which there are multiple buyers and sellers that wish to trade (these are called
double auctions) and with cases in which agents wish to purchase multiple interrelated
goods at the same time (these are called combinatorial auctions). If it is a one-to-one
negotiation (one buyer and one seller) then a form of heuristic model is needed (e.g.
[Faratin99; Kraus01]). These models vary depending upon the nature of the
negotiation protocol and, in general, are less well developed than those for auctions.

Having determined the protocol, the next step is to determine the nature of the
contract that needs to be established. This will typically vary from application to
application and again it is something that is set by the market owner. Given these two,
the final step is to determine the agent’s reasoning model. This can vary from the very
simple (bidding truthfully) to the very complex (involving reasoning about the likely
number and nature of the other bidders).

The second main type of interaction is when a number of agents decide to come
together to form a new virtual organisation. This involves determining the participants
of the coalition and determining their various roles and responsibilities in this new
organisational structure. Again this is typically an activity that will involve
negotiation between the participants since they need to come to a mutually acceptable
agreement about the division of labour and responsibilities. Here there are a number
of techniques and algorithms that can be employed to address the coalition formation
process [Sandholm00; Shehory98] although this area requires more research to deal
with the envisaged scale of grid applications.

15

The Semantic Grid
3.2.3 Marketplace Structures
Marketplaces should be able to be established by any agent(s) in the system (including
a service owner, a service consumer or a neutral third party). The entity which
establishes the marketplace is here termed the market owner. The owner is responsible
for setting up, advertising, controlling and disbanding the marketplace. In order to
establish a marketplace, the owner needs a representation scheme for describing the
various entities that are allowed to participate in the marketplace (terms of entry), a
means of describing how the various allowable entities are allowed to interact with
one another in the context of the marketplace, and what monitoring mechanisms (if
any) are to be put in place to ensure the marketplace’s rules are adhered to.

3.3 A Service-Oriented View of the Scenario
The first marketplace is that connected with the scientist’s own lab. This marketplace
has agents to represent the humans involved in the experiment, thus there is a scientist
agent (SA) and a technician agent (TA). These are responsible for interacting with the
scientist and the technician, respectively, and then for enacting their instructions in the
Grid. These agents can be viewed as the computational proxies of the humans they
represent – endowed with their personalised information about their owner’s
preferences and objectives. These personal agents need to interact with other
(artificial) agents in the marketplace in order to achieve their objectives. These other
agents include an analyser agent (AA) (that is responsible for managing access to the
analyser itself), the analyser database agent (ADA) (that is responsible for managing
access to the database containing information about the analyser), and the high
resolution analyser agent (HRAA) (that is responsible for managing access to the
high resolution analyser). There is also an interest notification agent (INA) (that is
responsible for recording which scientists in the lab are interested in which types of
results and for notifying them when appropriate results are generated) and an
experimental results agent (ERA) (that can discover similar analyses of data or when
similar experimental configurations have been used in the past). The services
provided by these agents are summarised in table 2.

Agent Services Offered Services Consumed By
Scientist Agent
(SA)
resultAlert
reportAlert
Scientist
Scientist
Technician Agent (TA) MonitorAnalysis Technician
Analyser Agent
(AA)
configureParameters
runSample
ADA
ADA
Analyser Database Agent
(ADA)
logSample
setAnalysisConfiguration
bookSlot
recordAnalysis
Technician
Technician
TA
AA
High Resolution
Analyser Agent
(HRAA)
bookSlot
configureParameters
runAnalysis
videoAnalysis
monitorAnalysis
reportResults
replayExperiment
suggestRelatedConfigurations
SA
Scientist
Scientist
Scientist, Technician
Technician
SA
Scientist
Scientist
Interest Notification Agent
(INA)
registerInterest
notifyInterestedParties
findInterestedParties
Scientists, Technicians
ADA
Scientist
16

The Semantic Grid
Experimental Results Agent
(ERA)
FindSimilarExperiments HRAA

Table 2: Services in the scientist’s lab marketplace

The operation of this marketplace is as follows. The technician uses the logSample
service to record data about the sample when it arrives and the
setAnalysisConfiguration service to set the appropriate parameters for the
forthcoming experiment. The technician then instructs the TA to book a slot on the
analyser using the bookSlot service. At the appropriate time, the ADA informs the
AA of the settings that it should adopt (via the configureParameters service) and
that it should now run the experiment (via the runSample service). As part of the
contract for the runSample service, the AA informs the ADA of the results of the
experiment and these are logged along with the appropriate experimental settings
(using the recordAnalysis service). Upon receipt of these results, the ADA informs
the INA of them. The INA then disseminates the results (via the
notifyInterestedParties service) to scientists who have registered an interested in
results of that kind (achieved via the registerInterest service).

When interesting results are received, the SA alerts the scientist (via the resultAlert
service). The scientist then examines the results and decides that they are of interest
and that further analysis is need. The scientist then instructs the SA to make a booking
on the High Resolution Analyser (via the bookSlot service). When the booking is
made, the HRAA volunteers information to the scientist about the configurations of
similar experiments that have previously been run (via the
suggestRelatedConfigurations service). Using this information, the scientist sets
the appropriate configurations (via the configureParameters service). At the
appropriate time, the experiment is started (via the runAnalysis service). As part of
the contract for this service, the experiment is videoed (via the videoAnalysis
service), monitoring information is sent to the technician (via the monitorAnalysis
service) and a report is prepared and sent to the SA (via the reportResults service).
In preparing this report, the HRAA interacts with the ERA to discover if related
experiments and results have already been undertaken (achieved via the
findSimilarExperiments service).

The scientist is alerted to the report by the SA (via the reportAlert service). The
scientist decides the results may be interesting and decides to replay some of the key
segments of the video (via the replayExperiment service). The scientist decides the
results are indeed interesting and so asks for relevant publications and details of
scientists who have published on this topic. This latter activity is likely to be provided
through an external marketplace that provides this service for the wider community
(see table 4). In such a marketplace, there may be multiple Paper Repository Agents
that offer the same broad service (findRelatedPapers and findRelatedAuthors)
but to varying degrees of quality, coverage, and timeliness.

Armed with all this information, the scientist decides that the results should be
discussed within the wider organisation context. This involves interacting in the
Scientist’s Organisation Marketplace. The agents involved in this marketplace are the
research meeting convener agent (RMCA) (responsible for organising research
meetings) and the various scientist agents that represent the relevant scientists. The
17

The Semantic Grid
services provided by these agents are given in table 3. The RMCA is responsible for
determining when research meetings should take place, this is achieved via the
arrangeMeeting service through interaction with the SAs of the scientists involved.
The scientist requests a slot to discuss the latest experimental findings (via the
setAgenda service) and provides the appropriate data for discussion to the RMCA
that disseminates it to the SA’s of the relevant participants (via the
disseminateInformation service). As a consequence of the meeting, it is decided
that the results are appropriate for dissemination into the scientific community at
large.

Agent Services Offered Service Consumed By
Research Meeting Convener
Agent (RMCA)
arrangeMeeting
setAgenda
disseminateInformation
SAs
Scientist
SAs
Scientist Agent (SA) arrangeMeeting
receiveInformation
RMCA
RMCA

Table 3: Services in the scientist’s organisation marketplace

The general scientific community is represented by a series of distinct marketplaces
that are each responsible for different aspects of the scientific process. As decided
upon at the organisation’s meeting, the sample data is logged in the appropriate
international database (using the logSample service). This database has an attached
notification service at which individual scientists can register their interests in
particular types of data (via the registerInterests service). Scientists will then be
informed, via their SA, when new relevant data is posted (via the
disseminateInformation service).

Services Offered Services Consumed By
International Sample
Database Agent (ISDA)
LogSample
registerInterests
disseminateInformation
Scientist
Scientist
SAs
Paper Repository Agent
(PRA)
FindRelatedPapers
FindRelatedAuthors
SAs
SAs
Scientist Agent (SA) ReceiveRelevantData
ArrangeSimulation
Scientist
Scientist
Simulation Provider
Agent (SPA)
offerSimulationResource
utiliseSimulationResource
SA
SA
Problem Solving
Environment Agent
(PSEA)
WhatSimulationTools
simulationSettingInfo
analyseResults
Scientist
Scientist
Scientist

Table 4: Services in the general scientific community marketplace

One of the scientists who receives notification of the new results believes that they
should be investigated further by undertaking a new round of simulations. The
scientist instructs the SA to arrange for particular simulations to be arranged. The SA
enters a marketplace where providers of processing capabilities offer their resources
(via the offerSimulationResource service). The SA will arrange for the appropriate
amount of resource to be made available at the desired time such that the simulations
can be run. Once these contracts have been established, the SA will invoke the
18

The Semantic Grid
simulation (via the utiliseSimulationResource service). During the course of
these simulations, the scientist will make use of the Problem Solving Environment
Agent (PSEA) to assist in the tasks of determining what simulation tools to exploit
(via the whatSimulationTools service), setting the simulation parameters
appropriately for these tools (via the simulationSettingInfo service), and
analysing the results (via the analyseResults service).

This then characterises our scenario as an active marketplace of agents offering and
consuming services. As already indicated, we do not expect that this complete set of
interactions will be dealt with seamlessly by computational agents in the near future.
However, it provides a level of abstraction and defines capabilities that we claim it is
important to aspire to if the full potential of the Semantic Grid is to be realised.

4. The Knowledge Layer
The aim of the knowledge layer is to act as an infrastructure to support the
management and application of scientific knowledge to achieve particular types of
goal and objective. In order to achieve this, it builds upon the services offered by the
data/computation and information layers (see [DeRoure02b] for more details of the
services and technologies at these layers).

The first thing to reiterate with respect to this layer is the problem of the sheer scale of
content we are dealing with. We recognise that the amount of data that the data grid is
managing is likely to be huge. By the time that data is equipped with meaning and
turned into information we can expect order of magnitude reductions in the amount.
However what remains will certainly be enough to present us with the problem of
infosmog – the condition of having too much information to be able to take effective
action or apply it in an appropriate fashion to a specific problem. Once information is
delivered that is destined for a particular purpose, we are in the realm of the
knowledge grid. Thus at this level we are fundamentally concerned with abstracted
and annotated content and with the management of scientific knowledge.

We can see this process of scientific knowledge management in terms of a life cycle
of knowledge-oriented activity that ranges over knowledge acquisition and modelling,
knowledge retrieval and reuse, knowledge publishing and knowledge maintenance
(section 4.1). Next we discuss the fundamental role that ontologies will play in
providing the underpinning semantics for the knowledge layer (section 4.2). Section
4.3 then considers the knowledge services aspects of our scenario. Finally, we review
the research issues associated with our requirements for a knowledge grid (section
4.4).

4.1 The Knowledge Lifecycle

The knowledge lifecycle can be regarded as a set of challenges as well as a sequence
of stages. Each stage has variously been seen as a bottleneck. The effort of acquiring
knowledge was one bottleneck recognised early [Hayes-Roth83]. But so too are;
modelling, retrieval, reuse, publication and maintenance. In this section we examine
19

The Semantic Grid
the nature of the challenges at each stage in the knowledge lifecycle and review the
various methods and techniques at our disposal.

Although we often suffer from a deluge of data and too much information, all too
often what we have is still insufficient or too poorly specified to address our
problems, goals and objectives. In short, we have insufficient knowledge. Knowledge
acquisition sets the challenge of getting hold of the information that is around, and
turning it into knowledge by making it usable. This might involve, for instance,
making tacit knowledge explicit, identifying gaps in the knowledge already held,
acquiring and integrating knowledge from multiple sources (e.g. different experts, or
distributed sources on the Web), or acquiring knowledge from unstructured media
(e.g. natural language or diagrams).

A range of techniques and methods has been developed over the years to facilitate
knowledge acquisition. Much of this work has been carried out in the context of
attempts to build knowledge-based or expert systems. Techniques include varieties of
interview, different forms of observation of expert problem solving, methods of
building conceptual maps with experts, various forms of document and text analysis,
and a range of machine learning methods [Shadbolt95]. Each of these techniques has
found to be suited to the elicitation of different forms of knowledge and to have
different consequences in terms of the effort required to capture and model the
knowledge [Hoffman95; Shadbolt99]. Specific software tools have also been
developed to support these various techniques [Milton99] and increasingly these are
now web enabled [Shaw98].

However, the process of explicit knowledge acquisition from human experts remains
a costly and resource intensive exercise. Hence, the increasing interest in methods that
can (semi-) automatically elicit and acquire knowledge that is often implicit or else
distributed on the web [Crow01]. A variety of information extraction tools and
methods are being applied to the huge body of textual documents that are now
available [Ciravegna01]. Examples include programs to extract information about
protein function from various scientific papers, abstracts and databases that are
increasingly available on-line. Another style of automated acquisition consists of
systems that observe user behaviour and infer knowledge from that behaviour.
Examples include recommender systems that might look at the papers downloaded by
a researcher and then detect themes by analysing the papers using methods such as
term frequency analysis [Middleton01]. The recommender system then searches other
literature sources and suggests papers that might be relevant or else of interest to the
user.

Methods that can engage in the sort of background knowledge acquisition described
above are still in their infancy but with the proven success of pattern directed methods
in areas such as data mining they are likely to assume a greater prominence in our
attempts to overcome the knowledge acquisition bottleneck.

Knowledge modelling bridges the gap between the acquisition of knowledge and its
use. Knowledge models must be able both to act as straightforward placeholders for
the acquired knowledge, and to represent the knowledge so that it can be used for
problem-solving. Knowledge representation technologies have a long history in
Artificial Intelligence. There a numerous languages and approaches that cater for
20

The Semantic Grid
different knowledge types; structural forms of knowledge, procedurally oriented
representations, rule based characterisations and methods to model uncertainty, and
probabilistic representations [Brachman83].

Most large applications require a range of knowledge representation formats.
CommonKADS [Schreiber00], one of the most comprehensive methodologies for the
development of knowledge intensive systems, uses a range of modelling methods and
notations- including logic and structured objects. It also factors out knowledge into
various types and identifies recurrent patterns of inference and knowledge type that
denote characteristic problem solvers. These patterns are similar to design patterns in
software engineering [Gamma95] and attempt to propose a set of components out of
which problem solving architectures can be composed. One of the major constituents
of the models built in CommonKADS are domain ontologies which we discuss in the
next section.

Recently with the explosion of content on the web there has arisen the recognition of
the importance of metadata. Any kind of content can be “enriched” by the addition of
annotations about what the content is about [Motta00]. Such semantic metadata is an
important additional element in our modelling activity. It may indicate the origin of
content, its provenance, value or longevity. It may associate other resources with the
content such as the rationale as to why the content is in the form it is and so on.

Certainly given the sheer amount of content available in a grid context it is crucial to
have some technical support for metadata “enrichment”. To this end a number of
systems are now under development that aim to take given metadata structures and
help annotate, tag or associate content with that metadata [Motta02, Handschuh02].

In any modelling exercise it is important to recognise that the modelling reflects a set
of interests and perspectives. These may be made more or less explicit but they are
always present. It is also important to recognise that models may be more or less
formal and aspire to various degrees of precision and accuracy. The model is, of
course, not the object or process, rather it is an artefact built with a particular set of
goals and intentions in mind.

Once knowledge has been acquired and modelled, it needs to be stored or hosted
somewhere so that it can be retrieved efficiently. In this context, there are two related
problems to do with knowledge retrieval. First, there is the issue of finding knowledge
again once it has been stored. And second, there is the problem of retrieving the
subset of content that is relevant to a particular problem. This will set particular
problems for a knowledge retrieval system where content alters rapidly and regularly.

Technologies for information retrieval exist in many forms [Sparck-Jones97]. They
include methods that attempt to encode structural representations about the content to
be retrieved such as explicit attributes and values. Varieties of matching algorithm can
be applied to retrieve cases that are similar to an example or else a partial set of
attributes presented to the system. Such explicit Case Based Reasoning [Lenz98] and
Query engines have been widely adopted. They suffer from the problem of content
encoding – the ease with which new content and examples can be represented in the
required structural format. There are also perennial issues about the best measures of
similarity to use in these systems.
21

The Semantic Grid

Other retrieval methods are based on statistical encoding of the objects to be retrieved.
These might be as vectors representing the frequency of the terms in a document or
other piece of content. Retrieval is a matter of matching a query of an example piece
of content against these stored representations and generating closest matches
[Croft00].

Search engines such as Google that are manifestly capable of scaling and also
demonstrate good retrieval performance rely on concepts such as relevance ranking.
Here given any set of terms to search Google looks at the interconnected nature of
content and the frequency of its being accessed to help determine in part the rank of
how good a match to the material sought it is likely to be.

In the general field of content retrieval there is no one dominant paradigm - it can
occur at the fine grained level at which point it is a form of information extraction, or
else at the level of complete documents or even work flows or data logs that might
encode entire experimental configurations and subsequent runs.

One of the most serious impediments to the cost-effective use of knowledge is that too
often knowledge components have to be constructed afresh. There is little knowledge
reuse. This arises partly because knowledge tends to require different representations
depending on the problem-solving that it is intended to do. We need to understand
how to find patterns in knowledge, to allow for its storage so that it can be reused
when circumstances permit. This would save a good deal of effort in reacquiring and
restructuring the knowledge that had already been used in a different context.

We have already alluded to the form of reuse embodied in methodologies such as
CommonKADS. Here a problem-solving template for monitoring might be used in
one domain and its general structure reused elsewhere. The actual ontology of
components or processes might be another candidate for reuse. Complete problem
solving runs or other results might offer the chance to reuse previously solved
problems in areas that are similar. Workflows themselves might be reused. Technical
support in the area of reuse tends to be focused on the type of product being reused.
At one end of the spectrum we have reuse of ontologies in tools such as Protégé
[Schreiber00b], at the other there are tools to facilitate the reuse of complete problem
solving architectures [Motta99, Fensel99, Crubézy02]. Obstacles to reuse include the
very real possibility that it is sometimes easier to reconstruct the knowledge fragment
than hunt for it. Even when it is found it is often necessary to modify it to suit the
current context. Some knowledge is so difficult to model in a reusable fashion that an
explicit decision is made to reacquire when needed.

Having acquired knowledge, modelled and stored it, the issue then arises as to how to
get that knowledge to the people who subsequently need it. The challenge of
knowledge publishing or disseminating can be described as getting the right
knowledge, in the right form, to the right person or system, at the right time. Different
users and systems will require knowledge to be presented and visualised in different
ways. The quality of such presentation is not merely a matter of preference. It may
radically affect the utility of the knowledge. Getting presentation right involves
understanding the different perspectives of people with different agendas and systems
with different requirements. An understanding of knowledge content will help to
22

The Semantic Grid
ensure that important related pieces of knowledge get published at the appropriate
time.

Technologies to help publish content in fast and flexible are now starting to appear.
One such is the distributed link service (DLS). This is a method for associating
hyperlinks with content in such a way that the link is held separate from the content
and not represented in the content itself. This means that different link structures or
link bases can be associated with the same content. This allows very different
hypertext structures to be associated with the same content and supports very different
styles of publishing and subsequently navigating content [Carr98]. More recently DLS
systems have been built that generate links that can be switched in and out depending
on the ontology or conceptualisation in play at the time [Carr01]. Ontologies can also
act as filters on portals. By looking at the metadata associated with content the portal
can elect to show various content in different ways to different users
[http://www.ontoportal.org.uk/].

Given accumulated fragments of knowledge, methods now exist to thread this
information together and generate connected text to explain or present the fragments
[Bontcheva01, Bontcheva01b]. Some publication models seek to generate extended
narratives from harvested web content [Sanghee02]. Publishing services extend to
concepts such as the Open Archives Initiative [Harnad01] and ePrints [Hitchcock00].
In these models individuals deposit their papers with associated metadata. The ePrints
system for example can then offer the basis for additional services running on a
significant publication base. For example, it currently runs on the Los Alamos Physics
Archive consisting of some 100,000 documents and offers citation and automatic
cross-indexing services [http://opcit.eprints.org/].

Problems with publication include the fact that it has to be timely and it should not
overwhelm the recipient with detail nor content that is not of interest. Related to these
last two issues we find technologies under development to carry out summarisation
[Knight00] of texts and subject content identification [Landauer97, Landauer98].

Finally, having acquired and modelled the knowledge, and having managed to retrieve
and disseminate it appropriately, the last challenge is to keep the knowledge content
current – knowledge maintenance. This may involve the regular updating of content as
knowledge changes. Some content has considerable longevity, while other knowledge
dates quickly. If knowledge is to remain useful over a period of time, it is essential to
know which parts of the knowledge base must be updated or else discarded and when.
Other problems involved in maintenance include verifying and validating the content,
and certifying its safety.

Historically, the difficulty and expense of maintaining large software systems has
been underestimated. Where that information and knowledge content is to be
maintained in a distributed fashion the problem would appear to be even more acute.
Whether it is a repository full of documents or databases full of experimental data the
problem of curration needs to be addressed early in the system design process.
Moreover, it needs to be tackled early in the knowledge life cycle. When content is
acquired and modelled metadata regarding its provenance, quality and value ought to
be captured too. Otherwise one has little evidence about what it is important to
maintain and what are the likely consequences if it is changed or removed.
23

The Semantic Grid

Technologies have been developed to look at the effects of refining and maintaining
knowledge bases [Carbonara99]. These attempt to implement a range of checking
algorithms to see if altering the knowledge base leads to cyclic reasoning behaviour or
else disables or enables new classes of inference or behaviours. A different type of
maintenance relates to the domain descriptions or conceptualisations themselves.
Again it is important that at the point at which the ontology is designed careful
thought is given to those parts of the conceptualisation that are likely to remain stable
as opposed to areas where it is recognised that change and modification is likely to
happen. Once built an ontology is typically populated with instances to produce the
knowledge bases over which processing occurs. Populating ontologies with instances
is a constant process of maintenance and whenever it is carried out there can be much
post processing to eliminate for example duplicate instances from the knowledge base
[Alani02].

As with so many aspects of the knowledge life cycle, effective maintenance will also
depend on socio-technical issues having to do with whether there are clear owners and
stakeholders whose primary function is content and knowledge management.

We have already indicated that if the knowledge intensive activities described above
are to be delivered effectively in the Semantic Grid context then a crucial step is to
establish a basic level of semantic interoperation [section 3.2.2). This requires the
development of a shared vocabulary, description or conceptualisation for the
particular domain of interest. It is to this ontological engineering that we now turn.

4.2 Ontologies and the Knowledge Layer
The concept of an ontology is necessary to capture the expressive power that is
needed for modelling and reasoning with knowledge. Generally speaking, an ontology
determines the extension of terms and the relationships between them. However, in
the context of knowledge and web engineering, an ontology is simply a published,
more or less agreed, conceptualization of an area of content. The ontology may
describe objects, processes, resources, capabilities or whatever.

Recently a number of languages have appeared that attempt to take concepts from the
knowledge representation languages of AI and extend the expressive capability of
those of the Web (e.g., RDF and RDF Schema). Examples include SHOE [Luke00],
DAML [Hendler00], and OIL [vanHarmelen00]. Most recently there has been an
attempt to integrate the best features of these languages in a hybrid called
DAML+OIL. As well as incorporating constructs to help model ontologies
DAML+OIL is being equipped with a logical language to express rule-based
generalizations.

However the development of the Semantic Grid is not simply about producing
machine-readable languages to facilitate the interchange and integration of
heterogeneous information. It is also about the elaboration, enrichment and annotation
of that content. To this end, the list below is indicative of how rich annotation can
become. Moreover it is important to recognize that enrichment or meta-tagging can be
applied at any conceptual level in the three tier grid of figure 1. This yields the idea of
meta-data, meta-information and meta-knowledge.
24

The Semantic Grid

ƒ Domain ontologies: Conceptualisations of the important objects, properties
and relations between those objects. Examples would include an agreed set of
annotations for medical images, an agreed set of annotations for climate
information, and a controlled set of vocabulary for describing significant
features of engineering design.

ƒ Task ontologies: Conceptualisations of tasks and processes, their
interrelationships and properties. Examples would include an agreed set of
descriptors for the stages of a synthetic chemistry process, an agreed protocol
for describing the dependencies between optimisation methods, and a set of
descriptions for characterizing the enrichment or annotation process when
describing a complex medical image.

ƒ Quality ontologies: Conceptualisations of the attributes that knowledge assets
possess and their interrelationships. Examples would include annotations that
would relate to the expected error rates in a piece of medical imaging, the
extent to which the quality of a result from a field geologist depended on their
experience and qualifications, and whether results from particular scientific
instruments were likely to be superseded by more accurate devices.

ƒ Value ontologies: Conceptualisations of those attributes that are relevant to
establishing the value of content. Examples would include the cost of
obtaining particular physics data, the scarcity of a piece of data from the fossil
record, and how widely known a particular metabolic pathway was.

ƒ Personalisation ontologies: Conceptualisations of features that are important to
establishing a user model or perspective. Examples would include a
description of the prior familiarity that a scientist had with particular
information resources, the amount of detail that the user was interested in, and
the extent to which the user’s current e-Science activities might suggest other
content of interest.

ƒ Argumentation ontologies – A wide range of annotations can relate to the
reasons why content was acquired, why it was modelled in the way it was, and
who supports or dissents from it. This is particularly powerful when extended
to the concept of associating discussion threads with content. Examples are the
integration of authoring and reviewing processes in on-line documents. Such
environments allow structured discussions of the evolution and development
of an idea, paper or concept. The structured discussion is another annotation
that can be held in perpetuity. This means that the reason for a position in a
paper or a design choice is linked to the object of discussion itself.

The benefits of an ontology include improving communication between systems
whether machines, users or organizations. They aim to establish an agreed and
perhaps normative model. They endeavour to be consistent and unambiguous, and to
integrate a range of perspectives. Another benefit that arises from adopting an
ontology is inter-operability and this is why they figure large in the vision for the
Semantic Web [BernersLee01]. An ontology can act as an interlingua, it can promote
25

The Semantic Grid
reuse of content, ensure a clear specification of what content or a service is about, and
increase the chance that content and services can be successfully integrated.

A number of ontologies are emerging as a consequence of commercial imperatives
where vertical marketplaces need to share common descriptions. Examples include
the Common Business Library (CBL), Commerce XML (cXML), ecl@ss, the Open
Applications Group Integration Specification (OAGIS), Open Catalog Format (OCF),
the Open Financial Exchange (OFX), Real Estate Transaction Markup Language
(RETML), RosettaNet, UN/SPSC (see www.diffuse.org), and UCEC.

We can see examples of ontologies built and deployed in a range of traditional
knowledge intensive applications ranging from chemical processing [Lopez99]
through to engineering plant construction [Mizoguchi00]. Moreover, there are a
number of large-scale ontology initiatives underway in specific scientific
communities. One such is in the area of genetics where a great deal of effort has been
invested in producing common terminology and definitions to allow scientists to
manage their knowledge [http://www.geneontology.org/]. This effort provides a
glimpse of how ontologies will play a critical role in sustaining the e-Scientist.

This work can also be exploited to facilitate the sharing, reuse, composition, mapping,
and succinct characterizations of (web) services. In this vein, [McIlraith01] exploit a
web service markup that provides an agent-independent declarative API that is aimed
at capturing the data and metadata associated with a service together with
specifications of its properties and capabilities, the interface for its execution, and the
prerequisites and consequences of its use. A key ingredient of this work is that the
markup of web content exploits ontologies. They have used DAML for semantic
markup of Web Services. This provides a means for agents to populate their local
knowledge bases so that they can reason about web services to perform automatic web
service discovery, execution, composition and interoperation.

It can be seen that ontologies clearly provide a basis for the communication,
integration and sharing of content. But they can also offer other benefits. An ontology
can be used for improving search accuracy by removing ambiguities and spotting
related terms, or by associating the information retrieved from a page with other
information. They can act as the backbone for accessing information from a
community web portal [Staab00]. Moreover Internet reasoning systems are beginning
to emerge that exploit ontologies to extract and generate annotations from the existing
web [Decker99].

Given the developments outlined in this section, a general process that might drive the
emergence of the knowledge grid would comprise:

• The development, construction and maintenance of application (specific and
more general areas of science and engineering) and community (sets of
collaborating scientists) based ontologies.
• The large scale annotation and enrichment of scientific data, information and
knowledge in terms of these ontologies
• The exploitation of this enriched content by knowledge technologies.

26

The Semantic Grid
There is a great deal of activity in the whole area of ontological engineering at the
moment. In particular, the World Wide Web Consortium (W3C) has a working group
developing a language to describe ontologies on the web; this Web Ontology
language, which is known as OWL, is based on DAML+OIL. The development and
deployment of ontologies is a major topic in the web services world and is set to
assume an important role in grid computing.

4.3 Knowledge Layer Aspects of the Scenario
Let us now consider our scenario in terms of the opportunities it offers for knowledge
services (see table 5). We will describe the knowledge layer aspects in terms of the
agent-based service oriented analysis developed in section 3.3. Important components
of this conceptualization were the software proxies for human agents such as the
scientist agent (SA) and the technician agent (TA). These software agents will
interact with their human counterparts to elicit preferences, priorities and objectives.
The software proxies will then realise these elicited items on the Grid. This calls for
knowledge acquisition services. As indicated in section 4.1, a range of methods could
be used. Structured interview methods invoke templates of expected and anticipated
information. Scaling and sorting methods enable humans to rank their preferences
according to relevant attributes that can either be explicitly elicited or pre-enumerated.
The laddering method enables users to construct or select from ontologies.
Knowledge capture methods need not be explicit – a range of pattern detection and
induction methods exist that can construct, for example, preferences from past usage.

One of the most pervasive knowledge services in our scenario is the partial or fully
automated annotation of scientific data. Before it can be used as knowledge, we need
to equip the data with meaning. Thus agents require capabilities that can take data
streaming from instruments and annotate it with meaning and context. Example
annotations include the experimental context of the data (where, when, what, why,
which, how). Annotation may include links to other previously gathered information
or its contribution and relevance to upcoming and planned work. Such knowledge
services will certainly be one of the main functions required by the Analyser Agent
and Analyser Database Agent (ADA). In the case of the High Resolution Analyser
Agent (HRAA) we have the additional requirement to enrich a range of media types
with annotations. In the original scenario this included video of the actual
experimental runs.

These acquisition and annotation services along with many others will be underpinned
by ontology services that maintain agreed vocabularies and conceptualizations of the
scientific domain. These are the names and relations that hold between the objects and
processes of interest to us. Ontology services will also manage the mapping between
ontologies that will be required by agents with differing interests and perspectives.


Agent Requirements Knowledge Technology Services
Scientist Agent
(SA)
Knowledge Acquisition of Scientist Profile
Ontology Service
Technician Agent
(TA)
Knowledge Acquisition of Technician Profile
Ontology Service
Knowledge Based Scheduling Service to book analyser
Analyser Agent Annotation and enrichment of instrument streams
27

The Semantic Grid
(AA) Ontology Service
Analyser Database Agent
(ADA)
Annotation and enrichment of databases
Ontology Service
High Resolution
Analyser Agent
(HRAA)
Annotation and enrichment of media
Ontology Service
Language Generation Services
Internet Reasoning Services
Interest Notification Agent
(INA)
Knowledge Publication Services
Language Generation Services
Knowledge Personalisation Services
Ontology Service
Experimental Results Agent
(ERA)
Language Generation Services
Result Clustering and Taxonomy Formation
Knowledge and Data Mining Service
Ontology Service
Research Meeting
Convener Agent (RMCA)
Constraint Based Scheduling Service
Knowledge Personalisation Service
Ontology Service
International Sample
Database Agent (ISDA)
Result Clustering and Taxonomy Formation
Knowledge and Data Mining Services
Ontology Service
Paper Repository Agent
(PRA)
Annotation and enrichment of papers
Ontology Service
Dynamic Link Service
Discussion and Argumentation Service
Problem Solving
Environment Agent (PSEA)
Knowledge Based Configuration of PSE Components
Knowledge Based Parameter Setting and Input Selection
Ontology Service

Table 5: Example knowledge services in the scenario

Personalisation services will also be invoked by a number of the agents in the
scenario. These might interact with the annotation and ontology services already
described so as to customize the generic annotations with personal markup – the fact
that certain types of data are of special interest to a particular individual. Personal
annotations might reflect genuine differences of terminology or perspective –
particular signal types often have local vocabulary to describe them. Ensuring that
certain types of content are noted as being of particular interest to particular
individuals brings us on to services that notify and push content in the direction of
interested parties. The Interest Notification Agent (INA) and the Research Meeting
Convener Agent (RMCA) could both be involved in the publication of content either
customized to individual or group interests. Portal technology can support the
construction of dynamic content to assist the presentation of experimental results.

Agents such as the High Resolution Analyser (HRAA) and Experimental Results
Analyser (ERA) have interests in classifying or grouping certain information and
annotation types together. Examples might include all signals collected in a particular
context, or sets of signals collected and sampled across contexts. This in turn provides
a basis for knowledge discovery and the mining of patterns in the content. Should
such patterns arise these might be further classified against existing pattern types held
in international databases – in our scenario this is managed in marketplaces by agents
such as the International Sample Database Agent (ISDA).

28

The Semantic Grid
At this point agents are invoked whose job it is to locate other systems or agents that
might have an interest in the results. Negotiating the conditions under which the
results can be released, determining the quality of results, might all be undertaken by
agents that are engaged to provide result brokering and result update services.

Raw results are unlikely to be especially interesting so that the generation of natural
language summaries of results will be important for many of the agents in our
scenario. Results that are published this way will also want to be linked and threaded
to existing papers in the field and made available in ways that discussion groups can
usefully comment on. Link services are one sort of knowledge technology that will be
ubiquitous here – this is the dynamic linking of content in documents in such a way
that multiple markups and hyperlink annotations can be simultaneously maintained.
Issue tracking and design rationale methods allow multiple discussion threads to be
constructed and followed through documents. In our scenario the Paper Respository
Agent (PRA) will not only retrieve relevant papers but mark them up and thread them
in ways that reflect the personal interests and conceptualizations (ontologies) of
individuals or research groups.

The use of Problem Solving Environment Agents (PSEAs) in our simulation of
experimentally derived results presents us with classic opportunities for knowledge
intensive configuration and processing. Once again these results may be released to
communities of varying size with their own interests and viewpoints.

Ultimately it will be up to application designers to determine if the knowledge
services described in this scenario are invoked separately or else as part of the
inherent competences of the agents described earlier. Whatever the design decisions,
it is clear that knowledge services will play a fundamental role in realizing the
potential of the Semantic Grid for the e-Scientist.
4.4 Research Issues
The following is a list of the key research issues that remain for exploiting knowledge
services in the Semantic Grid. In many cases there are already small-scale exemplars
for most of these services; consequently many of the issues relate to the problems of
scale and distribution

ƒ Languages and infrastructures are needed to describe, advertise and locate
knowledge services. We need the means to invoke and communicate the
results of such services. This is the sort of work that is currently underway in
the Semantic Web effort of DAML-S [ref]. However, it is far from clear how
this work will interface with that of the agent based computing, web services
and grid communities.
ƒ Methods are required to build large-scale ontologies and tools deployed to
provide a range of ontology services.
ƒ Annotation services are required that will run over large corpora of local and
distributed data. In some cases, for example, the annotation and cleaning of
physics data, this process will be iterative and will need to be near real time as
well as supporting fully automatic and mixed initiative modes. These
annotation tools are required to work with mixed media.
ƒ Knowledge capture tools are needed that can be added as plugins to a wide
variety of applications and which draw down on ontology services. This will
29

The Semantic Grid
include a clearer understanding of profiling individual and group e-Science
perspectives and interests.
ƒ Dynamic linking, visualization, navigation and browsing of content from
many perspectives over large content sets
ƒ Retrieval methods based on explicit annotations.
ƒ Construction of repositories of solution cases with sufficient annotation to
promote reuse as opposed to discovering the solution again because the cost of
finding the reusable solution is too high.
ƒ Deployment of routine natural language processing as Internet services.
Capabilities urgently required include: tagging and markup of documents,
discovering different linguistic forms of ontological elements, and providing
language generation and summarization methods for routine scientific
reporting
ƒ Deployment of Internet based reasoning services – whether as particular
domain PSEs or more generic problem solvers such as scheduling and
planning systems.
ƒ Provision of knowledge discovery services with standard input/output APIs to
ontologically mapped data
ƒ Understanding how to embed knowledge services in ubiquitous and pervasive
devices

5. Conclusions
This paper has outlined our vision of the Semantic Grid as a future e-Science
infrastructure in which there is a high degree of easy-to-use and seamless automation
and in which there are flexible collaborations and computations on a global scale. We
have argued that this infrastructure should be conceptualised and implemented as a
service-oriented architecture in which agents interact with one another in various
types of information marketplace. Moreover, we have highlighted the importance of
knowledge services in this vision and have outlined the key research challenges that
need to be addressed at this level.

In order to make the Semantic Grid a reality, a number of research challenges need to
be addressed. These include (in no particular order):

• Smart Laboratories. We believe that for e-Science to be successful and for the
Grid to be effectively exploited much more attention needs to focused on how
laboratories need to be instrumented and augmented. For example, infrastructure
that allows a range of equipment to advertise its presence, be linked together,
annotate and markup content it is receiving or producing.
• Service-Oriented Architectures. Research the provision and implementation of
grid facilities in terms of service oriented architectures. Also research into service
description languages as a way of describing and integrating the Grid’s problem
solving elements.
• Agent Based Approaches. Research the use of agent based architectures and
interaction languages to enable e-Science marketplaces to be developed, enacted
and maintained.
• Trust and Provenance. Further research is needed to understand the processes,
methods and techniques for establishing computational trust and determining the
30

The Semantic Grid
provenance and quality of content in Grid systems. This extends to the issue of
digital rights management in making content available.
• Metadata and Annotation. Whilst the basic metadata infrastructure already exists
in the shape of RDF, metadata issues have not been fully addressed in current grid
deployments. It is relatively straightforward to deploy some of the technology in
this area, and this should be promoted. RDF, for example, is already encoding
metadata and annotations as shared vocabularies or ontologies. However, there is
still a need for extensive work in the area of tools and methods to support the
design and deployment of e-Science ontologies. Annotation tools and methods
need to be developed so that emerging metadata and ontologies can be applied to
the large amount of content that will be present in Grid applications.
• Knowledge Technologies. In addition to the requirement for the research in
metadata and annotation, there is a need for a range of other knowledge
technologies to be developed and customised for use in e-Science contexts. These
include knowledge capture tools and methods, dynamic content linking,
annotation based search, annotated reuse repositories, natural language processing
methods (for content tagging, mark-up, generation and summarisation), data
mining, machine learning and internet reasoning services. These technologies will
need shared ontologies and service description languages if they are to be
integrated into the e-Science workflow. These technologies will also need to be
incorporated into the pervasive devices and smart laboratory contexts that will
emerge in e-Science.
• Integrated Media. Research into incorporating a wide range of media into the e-
Science infrastructure. This will include video, audio, and a wide range of imaging
methods. Research is also needed into the association of metadata and annotation
with these various media forms.
• Content Presentation. Research is required into methods and techniques that allow
content to be visualised in ways consistent with the e-Science collaborative effort.
This will also involve customising content in ways that reflect localised context
and should allow for personalisation and adaptation.
• e-Science Workflow and Collaboration. Much more needs to be done to
understand the workflow of current and future e-Science collaborations. Users
should be able to form, maintain and disband communities of practice with
restricted membership criteria and rules of operation. Currently most studies focus
on the e-Science infrastructure behind the socket on the wall. However this
infrastructure will not be used unless it fits in with the working environment of the
e-Scientists. This process has not been studied explicitly and there is a pressing
need to gather and understand these requirements. There is a need to collect real
requirements from users, to collect use cases and to engage in some evaluative and
comparative work. There is also a need to more fully understand the process of
collaboration in e-Science.
• Pervasive e-Science. Currently most references and discussions about grids imply
that their primary task is to enable global access to huge amounts of computational
power. Generically, however, we believe grids should be thought of as the means
of providing seamless and transparent access from and to a diverse set of
networked resources. These resources can range from PDAs to supercomputers
and from sensor’s and smart laboratories to satellite feeds.
• e-Anything. Many of the issues, technologies and solutions developed in the
context of e-Science can be exploited in other domains where groups of diverse
stakeholders need to come together electronically and interact in flexible ways.
31

The Semantic Grid
Thus it is important that relationships are established and exploitation routes are
explored with domains such as e-Business, e-Commerce, e-Education, and e-
Entertainment.

7 References


[Alani02] Alani, H., Dasmahapatra, S., Gibbins, N., Glaser, H., Harris, S., Kalfoglou,
Y., O'Hara, K., and Shadbolt, N. Managing Reference: Ensuring Referential
Integrity of Ontologies for the Semantic Web. 14th International Conference
on Knowledge Engineering and Knowledge Management, Spain, October,
2002.
[BernersLee01] Berners-Lee,T., Hendler,J. and Lassila, O. “The Semantic Web”,
Scientific American, May 2001.
[BernersLee99] Berners-Lee, T. with Fischetti, M. “Weaving the Web:The Original
Design and Ultimate Destiny of the World Wide Web by its Inventor”, Harper,
San Francisco, 1999.
[Brachman83] Brachman, R.J. & Levesque, H.J. (1983). Readings in Knowledge
Representation. San Mateo, Califomia: Morgan Kaufmann Publishers.
[Bontcheva01] Bontcheva, K. Tailoring the Content of Dynamically Generated
Explanations. M. Bauer, P.J. Gmytrasiewicz, J. Vassileva (eds). User
Modelling 2001: 8th International Conference, UM2001, Lecture Notes in
Artificial Intelligence 2109, Springer Verlag, 2001.
[Bontcheva01b] Bontcheva, K., Wilks, Y. Dealing with Dependencies between
Content Planning and Surface Realisation in a Pipeline Generation
Architecture. In Proceedings of International Joint Conference in Artificial
Intelligence (IJCAI'01), August 7-10. Seattle, 2001.
[Carr98] Carr, L., De Roure, D., Davis, H., Hall, W., Implementing an Open Link
Service for the World Wide Web. World Wide Web Journal. 1(2), 61-71.
Baltzer. 1998
[Carr01] Carr, L., Hall, W., Bechhofer, S. and Goble, C. (2001) Conceptual Linking:
Oncology-based Open Hypermedia. Proceedings of the Tenth International
World Wide Web Conference, Hong Kong, May 1-5 p.334-342.
[Cerf93] Cerf, V. G., et al., “National Collaboratories: Applying Information
Technologies for Scientific Research”, National Academy Press: Washington,
D.C., 1993.
32

The Semantic Grid
[Carbonara99] Carbonara, L. and Sleeman, D. (1999) Effective and Efficient
Knowledge Base Refinement. Machine Learning, Vol 37, pp143-181.
[Ciravegna01] Ciravegna, F.: "Adaptive Information Extraction from Text by Rule
Induction and Generalisation" in Proceedings of 17th International Joint
Conference on Artificial Intelligence (IJCAI 2001), Seattle, August 2001."
[Croft00] Croft, W. B. Information Retrieval Based on Statistical Language Models.
ISMIS 2000: 1-11
[Crow01] Crow, L. and Shadbolt, N. R. (2001) Extracting Focused Knowledge from
the Semantic Web, International Journal of Human Computer Studies, Vol. 54
(1) pp 155-184.
[Crubézy02] Crubézy, M., Lu, W., Motta, E. and Musen, M. A. . Configuring Online
Problem-Solving Resources with the Internet Reasoning Service. Conference
on Intelligent Information Processing (IIP 2002) of the International
Federation for Information Processing World Computer Congress (WCC
2002), Montreal, Canada. Kluwer. 2002.
[DAML02] DAML Services Coalition (alphabetically Anupriya Ankolenkar, Mark
Burstein, Jerry R. Hobbs, Ora Lassila, David L. Martin, Drew McDermott,
Sheila A. McIlraith, Srini Narayanan, Massimo Paolucci, Terry R. Payne and
Katia Sycara), "DAML-S: Web Service Description for the Semantic Web", in
The First International Semantic Web Conference (ISWC), June, 2002, pp
348-363.
[Decker99] Decker, S., Erdmann, M., Fensel, D. and Studer, R. “Ontobroker:
ontology-based access to distributed and semi-structured information” in R.
Meersman (ed.) Semantic Issues in Multimedia Systems: Proceedings of DS-8,
Kluwer Academic, Boston, 1999, 351-369.
[DeRoure01] De Roure, D., Jennings, N. R. and Shadbolt, N. R., “Research Agenda
for the Semantic Grid: A Future e-Science Infrastructure” Technical Report of
the National e-Science Centre, UKeS-2002-02, 2001.
[DeRoure02] De Roure, D., Baker, M., Jennings, N. R., and Shadbolt, N. R., “The
Evolution of the Grid” in this volume, 2002.
[Faratin99] Faratin, P., Sierra, C., and Jennings, N.R., “Negotiation decision functions
for autonomous agents” Int. J. of Robotics and Autonomous Systems 24 (3-4),
1999, 159-182.
[Fensel99] Fensel, D., Benjamins, V. R., Motta, E. and Wielinga, B. (1999). UPML:
A Framework for knowledge system reuse. In Proceedings of the International
Joint Conference on AI (IJCAI-99), Stockholm, Sweden, July 31 - August 5,
1999.
33

The Semantic Grid
[Foster98] Foster, I., and Kesselman, C., (eds), “The Grid: Blueprint for a New
Computing Infrastructure”, Morgan Kaufmann, July 1998.
[Foster01] Foster, I., Kesselman, C., and Tuecke, S., “The Anatomy of the Grid:
Enabling Scalable Virtual Organizations”, Int. Journal of Supercomputer
Applications and High Performance Computing, 2001.
[Foster02] Foster, I., Kesselman, C., Nick, J. and Tuecke, S., The Physiology of the
Grid: Open Grid Services Architecture for Distributed Systems Integration,
presented at GGF4, Feb. 2002 http://www.globus.og/research/papers/ogsa.pdf.
[Gamma95] Gamma, E., Helm, R., Johnson, R., and Vlissides, J. (1995). Design
Patterns: Elements of Reusable Object-Oriented Software. Reading, MA,
Addison-Wesley.
[Guttman98] Guttman, R.H., Moukas, A. G., and Maes, P., “Agent-mediated
electronic commerce: a survey” The Knowledge Engineering Review 13 (2)
1998 147-159.
[Handschuh02] Handschuh, S., Staab, S. and Ciravegna, F. S-CREAM --- Semi-
automatic CREAtion of Metadata. 13th International Conference on
Knowledge Engineering and Knowledge Management (EKAW'2002)
Sigüenza, Spain.
[Harnad01] Harnad, Stevan (2001) The Self-Archiving Initiative. Nature, 410 p.1024-
1025.
[Hayes-Roth83] Hayes-Roth, F. Waterman, D. A. and Lenat, D. B. (1983) Building
Expert Systems. Reading, Mass.: Addison-Wesley.
[Hendler00] Hendler, J., and McGuinness, D., “The DARPA Agent Markup
Language,” IEEE Intelligent Systems 15 (6), 2000, 72–73.
[Hitchcock00] Hitchcock, S., Carr, L., Jiao, Z., Bergmark, D., Hall, W., Lagoze, C.
and Harnad, Stevan (2000) Developing services for open eprint archives:
globalisation, integration and the impact of links. Proceedings of the 5th ACM
Conference on Digital Libraries, San Antonio, Texas, June 2000. p.143-151.
[Hoffman95] Hoffman, R., Shadbolt, N.R., Burton, A.M. and Klein,G. (1995)
“Eliciting Knowledge from Experts: A Methodological Analysis”
Organizational Behavior and Decision Processes, 62 (2) 1995, 129-158.
Academic Press.
[Jennings00] Jennings, N.R.,“On agent-based software engineering”, Artificial
Intelligence 117, 2000, 277-296.
34

The Semantic Grid
[Jennings01] Jennings, N.R., Faratin, P., Lomuscio, A.R., Parsons, S., Sierra, C., and
Wooldridge, M. “Automated Negotiation: Prospects, Methods and Challenges”
Int Journal of Group Decision and Negotiation 10(2) 2001, 199-215.
[Knight00] Knight, K. and Marcu, D. Statistics-Based Summarization --- Step One:
Sentence Compression, Proceedings of National Conference on Artificial
Intelligence (AAAI), 2000.
[Kraus01] S. Kraus “Strategic Negotiation in Multi-agent Environments” MIT Press.
2001.
[Landauer97] Landauer, T. K., & Dumais, S. T. (1997). A solution to Plato's problem:
The Latent Semantic Analysis theory of the acquisition, induction, and
representation of knowledge. Psychological Review, 104, 211-240.
[Landauer98] Landauer, T. K., Foltz, P. W., & Laham, D. (1998) Introduction to
Latent Semantic Analysis. Discourse Processes, 25, 259-284.
[Lenz98] Lenz, M., Bartsch-Spörl, B., Burkhard, H., and Wess, S. (Eds.) Case-Based
Reasoning Technology - From Foundations to Applications.
Lecture Notes in Artificial Intelligence 1400, Springer Verlag, 1998
[Lopez 99] Lopez, M.F., Gomez-Perex, A. et al., Building a chemical ontology using
Methontology and the ontology design environment, IEEE Intelligent
Systems, Vol.14, No.1, pp.37-46, 1999.
[Luke00] S. Luke and J. Heflin, “SHOE 1.01. Proposed Specification”,
www.cs.umd.edu/projects/plus/SHOE/spec1.01.html, 2000 (current 20 Mar.
2001).
[McIlraith01] McIlraith, S. A., Son, T. C., and Zeng, H., “Semantic Web Services”
IEEE Intelligent Systems, 16 (2) 2001, 46-53.
[Middleton01] Middleton, S. E., De Roure, D and Shadbolt, N. R. (2001) Capturing
knowledge of user preferences: ontologies in recommender systems.
Proceedings of the First International Conference on Knowledge Capture, K-
CAP2001, ACM Press.
[Milton99] Milton, N., Shadbolt, N., Cottam, H. and Hammersley, M. (1999).
Towards a Knowledge Technology for Knowledge Management. International
Journal of Human-Computer Studies, 51(3), 615-64.
[Motta99] Motta, E., Fensel, D., Gaspari, M. and Benjamins, R. (1999). Specifications
of Knowledge Components for Reuse. Eleventh International Conference on
Software Engineering and Knowledge Engineering (SEKE '99). June 1999
[Motta00] Motta, E., Buckingham Shum, S. and Domingue, J. (2001) Ontology-
Driven Document Enrichment: Principles, Tools and Applications.
35

The Semantic Grid
International Journal of Human Computer Studies, 52(5), pp. 1071-1109,
2000.
[Motta02] Motta, E., Vargas-Vera, M., Domingue, J., Lanzoni, M., Stutt, A., and
Ciravegna, F. MnM: Ontology Driven Semi-Automatic and Automatic
Support for Semantic Markup. 13th International Conference on Knowledge
Engineering and Knowledge Management (EKAW'2002) Sigüenza, Spain.
[Mizoguchi00] Mizoguchi, R., Kozaki, K., Sano, T., and Kitamura, Y.: Construction
and Deployment of a Plant Ontology, 12th International Conference on
Knowledge Engineering and Knowledge Management, Juan-les-Pins, French
Riviera, October, 2000.
[Newell82] Newell, A., “The Knowledge Level” Artificial Intelligence 18 1982, 87-
127.
[Sandholm00] Sandholm, T., “Agents in Electronic Commerce: Component
Technologies for Automated Negotiation and Coalition Formation”
Autonomous Agents and Multi-Agent Systems 3(1) 2000, 73-96.
[Sanghee02] Sanghee, K., Alani, H., Hall, W., Lewis, P., Millard, D., Shadbolt, N.,
Weal, M. Artequakt: Generating Tailored Biographies with Automatically
Annotated Fragments from the Web. In Proceedings Semantic Authoring,
Annotation and Knowledge Markup Workshop in the 15th European
Conference on Artificial Intelligence, Lyon, France. 2002.
[Schreiber00] Schreiber G., Akkermans, H., Anjewierden, A., de Hoog, R., Shadbolt,
N.R, Van de Velde, W. and Wielinga, B. (2000) Knowledge Engineering and
Management. MIT Press.
[Schreiber00b] Schreiber, G., Crubezy, M. & Musen, M. A. A Case Study in Using
Protege-2000 as a Tool for CommonKADS. 12th International Conference on
Knowledge Engineering and Knowledge Management (EKAW'2000), Juan-
les-Pins, France, 33-48. 2000. Springer LNAI.
[Shadbolt99] Shadbolt, N.R., O'Hara, K., Crow, L. (1999) The Experimental
Evaluation of Knowledge Acquisition Techniques and Methods: History,
Problems and New Directions International Journal of Human-Computer
Studies, 51(4), 729-755.
[Shadbolt95] Shadbolt, N.R. and Burton, M. (1995) Knowledge elicitation: a
systematic approach, in Evaluation of human work: A practical ergonomics
methodology edited by J. R. Wilson and E. N. Corlett, Taylor and Francis,
London, England, 1995. pp.406-440. ISBN-07484-0084-2.
[Shaw98] Shaw, M.L.G. & Gaines, B.R. (1998). WebGrid-II: developing hierarchical
knowledge structures from flat grids. In Proceedings of the 11th Knowledge
36

The Semantic Grid
37
Acquisition Workshop (KAW' 98). Banff, Canada, April 18-23, 1998.
Available at http://repgrid.com/reports/KBS/WG/.
[Shehory98] Shehory, O., and S. Kraus, S., “Methods for task allocation via agent
coalition formation” Artificial Intelligence, 101 (1-2) 1998, 165-200.
[Sparck-Jones97] Sparck-Jones, K., & Willett, P., Readings in information retrieval,
Mountain View: Morgan Kaufmann, 1997.
[Staab00] Staab, S., Angele, J., Decker, S., Erdmann, M., Hotho, A., Maedche, A.,
Schnurr, H.-P., Studer, R. and Sure, Y. “Semantic community Web portals”
Proc. of WWW-9, Amsterdam, 2000.
[vanHarmelen00] van Harmelen, F., and Horrocks, I., “FAQs on OIL: The Ontology
Inference Layer,” IEEE Intelligent Systems 15 (6), 2000, 69–72.
[WebServices01] Proceedings of W3C Web Services Workshop, April 11-12, 2001.
http://www.w3.org/2001/03/wsws-program
[Wooldridge97] Wooldridge, M., “Agent-based software engineering”. IEE Proc on
Software Engineering 144 (1) 1997, 26-37.