The Computing of Digital Ecosystems: Gerard Briscoe Philippe de Wilde

Download as pdf or txt
Download as pdf or txt
You are on page 1of 18

The Computing of Digital Ecosystems

Gerard Briscoe Philippe De Wilde


Media and Communications Department of Computer Science
London School of Economics and Political Science Heriot Watt University
United Kingdom United Kingdom
Abstract
A primary motivation for our research in digital ecosystems is the desire to exploit the
self-organising properties of biological ecosystems. Ecosystems are thought to be robust,
scalable architectures that can automatically solve complex, dynamic problems. However,
the computing technologies that contribute to these properties have not been made explicit
in digital ecosystems research. Here, we discuss how dierent computing technologies can
contribute to providing the necessary self-organising features, including Multi-Agent Systems
(MASs), Service-Oriented Architectures (SOAs), and distributed evolutionary computing
(DEC). The potential for exploiting these properties in digital ecosystems is considered,
suggesting how several key features of biological ecosystems can be exploited in Digital
Ecosystems, and discussing how mimicking these features may assist in developing robust,
scalable self-organising architectures. An example architecture, the Digital Ecosystem, is
considered in detail. The Digital Ecosystem is then measured experimentally through
simulations, considering the self-organised diversity of its evolving agent populations relative
to the user request behaviour.
1 Introduction
Digital Ecosystems are distributed adaptive open socio-technical systems, with properties
of self-organisation, scalability and sustainability, inspired by natural ecosystems [4, 8, 5],
and are emerging as a novel approach to the catalysis of sustainable regional development
driven by Small and Medium sized Enterprises (SMEs). Digital Ecosystems aim to help
local economic actors become active players in globalisation, valorising their local culture and
vocations, and enabling them to interact and create value networks at the global level [14, 43].
We have previously considered the biological inspiration for the technical component of
Digital Ecosystems [10, 9], and we will now consider the relevant computing technologies.
Based on our understanding of biological ecosystems in the context of Digital Ecosystems [6],
we will now introduce elds from the domain of computer science relevant to the creation
of Digital Ecosystems. As we are interested in the digital counterparts for the behaviour
and constructs of biological ecosystems, instead of simulating or emulating such behaviour or
constructs, we will consider what parallels can be drawn. We will start by considering MASs
to explore the references to agents and migration [10, 9]; followed by evolutionary computing
and SOAs for the references to evolution and self-organisation [10, 9].
The value of creating parallels between biological and computer systems varies substan-
tially depending on the behaviours or constructs being compared, and sometimes cannot
1
a
r
X
i
v
:
1
1
0
1
.
5
4
2
8
v
1


[
c
s
.
D
C
]


2
8

J
a
n

2
0
1
1
be done so convincingly. For example, both have mechanisms to ensure data integrity. In
computer systems, that integrity is absolute, data replication which introduces even the
most minor change is considered to have failed, and is supported by mechanisms such
as the Message-Digest algorithm 5 [39]. While in biological systems, the genetic code is
transcribed with a remarkable degree of delity; there is, approximately, only one unforced
error per one hundred bases copied [32]. There are also elaborate proof-reading and correction
systems, which in evolutionary terms are highly conserved [32]. In this example establishing
a parallel is infeasible, despite the relative similarity in function, because the operational
control mechanisms in biological and computing systems are radically dierent, as are the
aims and purposes. This is a reminder that considerable nesse is required when determining
parallels, or when using existing ones.
2 Multi-Agent Systems
A software agent is a piece of software that acts, for a user in a relationship of agency,
autonomously in an environment to meet its designed objectives. A MAS is a system
composed of several software agents, collectively capable of reaching goals that are dicult to
achieve by an individual agent or monolithic system. Conceptually, there is a strong parallel
between the software agents of a MAS and the agent-based models of a biological ecosystem
[15], despite the lack of evolution and migration in a MAS. There is an even stronger parallel
to a variant of MASs, called mobile agent systems, in which the mobility also mirrors the
migration in biological ecosystems [38].
Agent
Station
Figure 1: Mobile Agent System: Visualisation that shows mobile agents
as programmes that can migrate from one host to another in a network of
heterogeneous computer systems and perform a task specied by its owner.
The term mobile agent contains two separate and distinct concepts: mobility and agency
[40]. Hence, mobile agents are software agents capable of movement within a network
[38]. The mobile agent paradigm proposes to treat a network as multiple agent-friendly
environments and the agents as programmatic entities that move from location to location,
performing tasks for users. So, on each host they visit mobile agents need software which is
responsible for their execution, providing a safe execution environment [38].
Generally, there are three types of design for mobile agent systems [38]: (1) using
a specialised operating system, (2) as operating system services or extensions, or (3)
2
as application software. The third approach builds mobile agent systems as specialised
application software that runs on top of an operating system, to provide for the mobile agent
functionality, with such software being called an agent station [31]. In this last approach, each
agent station hides the vendor-specic aspects of its host platform, and oers standardised
services to visiting agents. Services include access to local resources and applications; for
example, web servers or web services, the local exchange of information between agents via
message passing, basic security services, and the creation of new agents [31]. Also, the third
approach is the most platform-agnostic, and is visualised in Figure 1.
3 Service-Oriented Architectures
To evolve high-level software components in Digital Ecosystems, we propose taking advantage
of the native method of software advancement, human developers, and the use of evolutionary
computing for combinatorial optimisation of the available software services. This involves
treating developer-produced software services as the functional building blocks, as the
base unit in a genetic-algorithms-based process. Our approach to evolving high-level
software applications requires a modular reusable paradigm to software development, such
as SOAs. SOAs are the current state-of-the-art approach, being the current iteration
of interface/component-based design from the 1990s, which was itself an iteration of
event-oriented design from the 1980s, and before then modular programming from the
1970s [22]. Service-oriented computing promotes assembling application components into
a loosely coupled network of services, to create exible, dynamic business processes and agile
applications that span organisations and computing platforms [37]. This is achieved through
a SOA, an architectural style that guides all aspects of creating and using business processes
throughout their life-cycle, packaged as services. This includes dening and provisioning
infrastructure that allows dierent applications to exchange data and participate in business
processes, loosely coupled from the operating systems and programming languages underlying
the applications. Hence, a SOA represents a model in which functionality is decomposed into
distinct units (services), which can be distributed over a network, and can be combined and
reused to create business applications [37].
A SOA depends upon service-orientation as its fundamental design principle. In a SOA
environment, independent services can be accessed without knowledge of their underlying
platform implementation. Services reect a service-oriented approach to programming that
is based on composing applications by discovering and invoking network-available services to
accomplish some task. This approach is independent of specic programming languages or
operating systems, because the services communicate with each other by passing data from
one service to another, or by co-ordinating an activity between two or more services [37]. So,
the concepts of SOAs are often seen as built upon, and the development of, the concepts of
modular programming and distributed computing [22].
SOAs allow for an information system architecture that enables the creation of applications
that are built by combining loosely coupled and interoperable services. They typically
implement functionality most people would recognise as a service, such as lling out an online
application for an account, or viewing an online bank statement [22]. Services are intrinsically
unassociated units of functionality, without calls to each other embedded in them. Instead
of services embedding calls to each other in their source code, protocols are dened which
describe how services can talk to each other, in a process known as orchestration, to meet
new or existing business system requirements. This is allowing an increasing number of
third-party software companies to oer software services, such that SOA systems will come
3
Figure 2: Service-Oriented Architectures: Abstract visualisations, with the
loosely joined services as cuboids, and the service orchestration as a polyhedron.
to consist of such third-party services combined with others created in-house, which has the
potential to spread costs over many users and uses, and promote standardisation both in and
across industries. For example, the travel industry now has a well-dened, and documented,
set of both services and data, sucient to allow any competent software engineer to create
travel agency software using entirely o-the-shelf software services [13]. Other industries,
such as the nance industry, are also making signicant progress in this direction [48].
The vision of SOAs assembling application components from a loosely coupled network
of services, that can create dynamic business processes and agile applications that span
organisations and computing platforms, is visualised in Figure 2. It will be made possible
by creating compound solutions that use internal organisational software assets, including
enterprise information and legacy systems, and combining these solutions with external
components residing in remote networks. The great promise of SOAs is that the marginal cost
of creating the n-th application is virtually zero, as all the software required already exists
to satisfy the requirements of other applications. Only their combination and orchestration
are required to produce a new application [33]. The key is that the interactions between the
chunks are not specied within the chunks themselves. Instead, the interaction of services
(all of whom are hosted by unassociated peers) is specied by users in an ad-hoc way, with
the intent driven by newly emergent business requirements [27].
The pinnacle of SOA interoperability, is the exposing of services on the internet as web
services. A web service is a specic type of service that is identied by a Uniform Resource
Identier (URI), whose service description and transport utilise open Internet standards.
Interactions between web services typically occur as Simple Object Access Protocol (SOAP)
calls carrying eXtensible Markup Language (XML) data content. The interface descriptions
of web services are expressed using the Web Services Denition Language (WSDL) [36].
The Universal Description Discovery and Integration (UDDI) standard denes a protocol for
4
directory services that contain web service descriptions. UDDI enables web service clients
to locate candidate services and discover their details. Service clients and service providers
utilise these standards to perform the basic operations of SOAs [36]. Service aggregators
can then use the Business Process Execution Language (BPEL) to create new web services
by dening corresponding compositions of the interfaces and internal processes of existing
services [36].
SOA services inter-operate based on a formal denition (or contract, e.g. WSDL) that
is independent of the underlying platform and programming language. Service descriptions
are used to advertise the service capabilities, interface, behaviour, and quality [36]. The
publication of such information about available services provides the necessary means for
discovery, selection, binding, and composition of services [36]. The (expected) behaviour
of a service during its execution is described by its behavioural description (for example,
as a workow process). Also, included is a quality of service (QoS) description, which
publishes important functional and non-functional service quality attributes, such as service
metering and cost, performance metrics (response time, for instance), security attributes,
integrity (transactional), reliability, scalability, and availability [36]. Service clients (end-user
organisations that use some service) and service aggregators (organisations that consolidate
multiple services into a new, single service oering) utilise service descriptions to achieve their
objectives [36]. One of the most important and continuing developments in SOAs is Semantic
Web Services (SWS), which make use of semantic descriptions for service discovery, so that
a client can discover the services semantically [11].
There are multiple standards available and still being developed for SOAs, most notably of
recent being REpresentational State Transfer (REST) [42]. The software industry now widely
implements a thin SOAP/WSDL/UDDI veneer atop existing applications or components that
implement the web services paradigm, but the choice of technologies will change with time.
Therefore, the fundamentals of SOAs and its services are best dened generically, because
SOAs are technology agnostic and need not be tied to a specic technology [37]. Within the
current and future scope of the fundamentals of SOAs, there is clearly potential to evolve
complex high-level software applications from the modular services of SOAs, instead of the
instruction level evolution currently prevalent in genetic programming [21].
4 Distributed Evolutionary Computing
Having previously suggested evolutionary computing [10, 9], and the possibility of it occurring
within a distributed environment, not unlike those found in mobile agent systems, leads us to
consider a specialised form known as DEC. The fact that evolutionary computing manipulates
a population of independent solutions actually makes it well suited for parallel and distributed
computation architectures [12]. The motivation for using parallel or distributed evolutionary
algorithms is twofold: rst, improving the speed of evolutionary processes by conducting
concurrent evaluations of individuals in a population; second, improving the problem-
solving process by overcoming diculties that face traditional evolutionary algorithms,
such as maintaining diversity to avoid premature convergence [44]. There are several
variants of distributed evolutionary computing, leading some to propose a taxonomy for
their classication, with there being two main forms [12, 44]:
multiple-population/coarse-grained migration/island
single-population/ne-grained diusion/neighbourhood
5
P
31
P
13
P
21
P
12
P
51
P
15
P
41
P
14
P
23
P
32
P
24
P
42
P
25
P
52
P
43
P
34
P
53
P
35
P
45
P
54
Island
3
Island
4
Island
5
Island
1
Island
2
Figure 3: Island-Model of Distributed Evolutionary Computing: There are
dierent probabilities of going from island 1 to island 2, as there is of going
from island 2 to island 1.
In the coarse-grained island models [28, 12], evolution occurs in multiple parallel sub-
populations (islands), each running a local evolutionary algorithm, evolving independently
with occasional migrations of highly t individuals among sub-populations. The core
parameters for the evolutionary algorithm of the island-models are as follows [28]:
number of the sub-populations: 2, 3, 4, more
sub-population homogeneity
size, crossover rate, mutation rate
topology of connectivity: ring, star, fully-connected
static or dynamic connectivity
migration mechanisms:
isolated/synchronous/asynchronous
how often migrations occur
which individuals migrate
Fine-grained diusion models [44] assign one individual per processor. A local neigh-
bourhood topology is assumed, and individuals are allowed to mate only within their
neighbourhood, called a deme. The demes overlap by an amount that depends on their
shape and size, and in this way create an implicit migration mechanism. Each processor
runs an identical evolutionary algorithm which selects parents from the local neighbourhood,
produces an ospring, and decides whether to replace the current individual with an ospring.
However, even with the advent of multi-processor computers, and more recently multi-core
processors, which provide the ability to execute multiple threads simultaneously, this approach
would still prove impractical in supporting the number of agents necessary to create a Digital
Ecosystem. Therefore, we shall further consider the island models.
An example island-model [28, 12] is visualised in Figure 3, in which there are dierent
probabilities of going from island 1 to island 2, as there is of going from island 2 to island
6
1. This allows maximum exibility for the migration process, and mirrors the naturally
inspired quality that although two populations have the same physical separation, it may be
easier to migrate in one direction than the other, i.e. sh migration is easier downstream
than upstream. The migration of the island models is like the notion of migration in nature,
being similar to the metapopulation models of theoretical ecology [26]. This model has also
been used successfully in the determination of investment strategies in the commercial sector,
in a product known as the Galapagos toolkit [45]. However, all the islands in this approach
work on exactly the same problem, which makes it less analogous to biological ecosystems in
which dierent locations can be environmentally dierent [3]. We will take advantage of this
property later when dening the Ecosystem-Oriented Architecture of Digital Ecosystems.
5 The Digital Ecosystem
We will now dene the architectural principles of Digital Ecosystems. We will use our
understanding of theoretical biology [6, 4], mimicking the processes and structures of
life, evolution, and ecology of biological ecosystems. We will achieve this by combining
elements from mobile agents systems, distributed evolutionary computing, and Service-
Oriented Architectures, to create a hybrid architecture which is the digital counterpart of
biological ecosystems. Combing these technologies, based on the biological inspiration, the
technical component of Digital Ecosystems provide a two-level optimisation scheme inspired
by natural ecosystems, in which a decentralised peer-to-peer network forms an underlying
tier of distributed agents. These agents then feed a second optimisation level based on
an evolutionary algorithm that operates locally on single habitats (peers), aiming to nd
solutions that satisfy locally relevant constraints. The local search is sped up through this
twofold process, providing better local optima as the distributed optimisation provides prior
sampling of the search space by making use of computations already performed in other
peers with similar constraints. So, the Digital Ecosystem supports the automatic combining
of numerous agents (which represent services), by their interaction in evolving populations to
meet user requests for applications, in a scalable architecture of distributed interconnected
habitats. The sharing of agents between habitats ensures the system is scalable, while
maintaining a high evolutionary specialisation for each user. The network of interconnected
habitats is equivalent to the abiotic environment of biological ecosystems [3]; combined with
the agents, the populations, the agent migration for distributed evolutionary computing, and
the environmental selection pressures provided by the user base, then the union of the habitats
creates the Digital Ecosystem, which is summarised in Figure 4. The continuous and varying
user requests for applications provide a dynamic evolutionary pressure on the applications
(agent aggregations), which have to evolve to better full those user requests, and without
which there would be no driving force to the evolutionary self-organisation of the Digital
Ecosystem.
If we consider an example user base for the Digital Ecosystem, the use of SOAs in its
denition means that business-to-business (B2B) interaction scenarios [22] lend themselves
to being a potential user base for Digital Ecosystems. So, we can consider a business ecosystem
of Small and Medium sized Enterprise (SME) networks [34], as a specic class of examples
for B2B interaction scenarios; and in which the SME users are requesting and providing
software services, represented as agents in the Digital Ecosystem, to full the needs of their
business processes, creating a Digital Business Ecosystem as shown in Figure 5. SOAs
promise to provide potentially huge numbers of services that programmers can combine,
via the standardised interfaces, to create increasingly more sophisticated and distributed
7
Genetic
Algorithm
Agent
Station
H
H
H
H
H
H H
H
H
H
H
H
H
H
H
H
H
H
Figure 4: Digital Ecosystem: Optimisation architecture in which agents
(representing services) travel along the P2P connections; in every node
(habitat) local optimisation is performed through an evolutionary algorithm,
where the search space is determined by the agents present at the node.
applications [36]. The Digital Ecosystem extends this concept with the automatic combining
of available and applicable services, represented by agents, in a scalable architecture, to meet
user requests for applications. These agents will recombine and evolve over time, constantly
seeking to improve their eectiveness for the user base. From the SME users point of view
the Digital Ecosystem provides a network infrastructure where connected enterprises can
advertise and search for services (real-world or software only), putting a particular emphasis
on the composability of loosely coupled services and their optimisation to local and regional,
needs and conditions. To support these SME users the Digital Ecosystem is satisfying the
companies business requirements by nding the most suitable services or combination of
services (applications) available in the network. An application (composition of services) is
dened be an agent aggregation (collection) in the habitat network that can move from one
peer (company) to another, being hosted only in those where it is most useful in satisfying
the SME users business needs.
The agents consist of an executable component and an ontological description. So, the
Digital Ecosystem can be considered a MAS which uses distributed evolutionary computing
[12, 44] to combine suitable agents in order to meet user requests for applications.
The landscape, in energy-centric biological ecosystems, denes the connectivity between
habitats [3]. Connectivity of nodes in the digital world is generally not dened by geography
or spatial proximity, but by information or semantic proximity. For example, connectivity
in a peer-to-peer network is based primarily on bandwidth and information content, and
not geography. The island-models of distributed evolutionary computing use an information-
centric model for the connectivity of nodes (islands) [28]. However, because it is generally
dened for one-time use (to evolve a solution to one problem and then stop) it usually
has a xed connectivity between the nodes, and therefore a xed topology [12]. So,
supporting evolution in the Digital Ecosystem, with a multi-objective selection pressure
(tness landscape with many peaks), requires a re-congurable network topology, such that
habitat connectivity can be dynamically adapted based on the observed migration paths
of the agents between the users within the habitat network. Based on the island-models
of distributed evolutionary computing [28], each connection between the habitats is bi-
directional and there is a probability associated with moving in either direction across the
8
Business
Ecosystem
Digital
Ecosystem
SME
Habitat
Figure 5: Digital Business Ecosystem: Business ecosystem, network of SMEs,
using the Digital Ecosystem. The habitat clustering will therefore be parallel
to the business sector communities.
connection, with the connection probabilities aecting the rate of migration of the agents.
However, additionally, the connection probabilities will be updated by the success or failure of
agent migration using the concept of Hebbian learning: the habitats which do not successfully
exchange agents will become less strongly connected, and the habitats which do successfully
exchange agents will achieve stronger connections. This leads to a topology that adapts over
time, resulting in a network that supports and resembles the connectivity of the user base. If
we consider a business ecosystem, network of SMEs, as an example user base; such business
networks are typically small-world networks [47]. They have many strongly connected clusters
(communities), called sub-networks (quasi-complete graphs), with a few connections between
these clusters (communities). Graphs with this topology have a very high clustering coecient
and small characteristic path lengths [46]. So, the Digital Ecosystem will take on a topology
similar to that of the user base, as shown in Figure 5.
The novelty of our approach comes from the evolving populations being created in response
to similar requests. So whereas in the island-models of distributed evolutionary computing
there are multiple evolving populations in response to one request [28], here there are multiple
evolving populations in response to similar requests. In our Digital Ecosystems dierent
requests are evaluated on separate islands (populations), and so adaptation is accelerated by
the sharing of solutions between evolving populations (islands), because they are working to
solve similar requests (problems).
The users will formulate queries to the Digital Ecosystem by creating a request as a
semantic description, like those being used and developed in SOAs, specifying an application
they desire and submitting it to their local peer (habitat). This description denes a metric
for evaluating the tness of a composition of agents, as a distance function between the
semantic description of the request and the agents ontological descriptions. A population
is then instantiated in the users habitat in response to the users request, seeded from the
agents available at their habitat. This allows the evolutionary optimisation to be accelerated
in the following three ways: rst, the habitat network provides a subset of the agents
available globally, which is localised to the specic user it represents; second, making use of
applications (agent aggregations) previously evolved in response to the users earlier requests;
and third, taking advantage of relevant applications evolved elsewhere in response to similar
requests by other users. The population then proceeds to evolve the optimal application
(agent aggregation) that fulls the user request, and as the agents are the base unit for
9
evolution, it searches the available agent combination space. For an evolved agent aggregation
(application) that is executed by the user, it then migrates to other peers (habitats) becoming
hosted where it is useful, to combine with other agents in other populations to assist in
responding to other user requests for applications.
6 Simulation and Results
We have previously compared the performance of Digital Ecosystems to traditional SOA
based systems, and found that it outperforms at greater scales [6], and so here we focus
on the self-organising behaviour which makes this possible. Self-organisation is perhaps one
of the most desirable features in the systems that we design, and a primary motivation
for our research in Digital Ecosystems is the desire to exploit the self-organising properties
of biological ecosystems [25], which are thought to be robust, scalable architectures that
can automatically solve complex, dynamic problems. Over time a biological ecosystem
becomes increasingly self-organised through the process of ecological succession [3], driven
by the evolutionary self-organisation of the populations within the ecosystem. Analogously,
a Digital Ecosystems increasing self-organisation comes from the agent populations within
being evolved to meet the dynamic selection pressures created by the requests from the
user base. The self-organisation of biological ecosystems is often dened in terms of the
complexity, stability, and diversity [19], which we will also apply to our Digital Ecosystems,
here considering the diversity.
Self-organisation has been around since the late 1940s [1], but has escaped general
formalisation despite many attempts [35, 20]. There have instead been many notions and
denitions of self-organisation, useful within their dierent contexts [17]. They have come
from cybernetics [1, 2, 18], thermodynamics [35], mathematics [24], information theory [41],
synergetics [16], and other domains [23]. The term self-organising is widely used, but there is
no generally accepted meaning, as the abundance of denitions would suggest. Therefore, the
philosophy of self-organisation is complicated, because organisation has dierent meanings
to dierent people. So, we would argue that any denition of self-organisation is context
dependent, in the same way that a choice of statistical measure is dependent on the data
being analysed.
While we could measure the self-organised diversity of individual evolving agent pop-
ulations, or even take a random sampling, it will be more informative to consider their
collective self-organised diversity. Also, given that the Digital Ecosystem is required to
support a range of user behaviour, we can consider the collective self-organised diversity of
the evolving agent populations relative to the global user request behaviour. So, when varying
a behavioural property of the user requests according to some distribution, we would expect
the corresponding property of the evolving agent populations to follow the same distribution.
We are not intending to prescribe the expected user behaviour of the Digital Ecosystem, but
investigate whether the Digital Ecosystem can adapt to a range of user behaviour. So, we
will consider Uniform, Gaussian (Normal) and Power Law distributions for the parameters
of the user request behaviour. The Uniform distribution will provide a control, while the
Normal (Gaussian) distribution will provide a reasonable assumption for the behaviour of
a large group of users, and the Power Law distribution will provide a relatively extreme
variation in user behaviour. So, we simulated the Digital Ecosystem, varying aspects of
the user behaviour according to dierent distributions, and measuring the related aspects
of the evolving agent populations. This consisted of a mechanism to vary the user request
properties of length and modularity (number of attributes per atomic service), according to
10
Uniform, Gaussian (normal) and Power Law distributions, and a mechanism to measure the
corresponding application (agent aggregation) properties of size and number of attributes
per agent. For statistical signicance each scenario (experiment) will be averaged from ten
thousand simulation runs. We expect it will be obvious whether the observed behaviour of the
Digital Ecosystem matches the expected behaviour from the user base. Nevertheless, we will
also implement a chi-squared (
2
) test to determine if the observed behaviour (distribution)
of the agent aggregation properties matches the expected behaviour (distribution) from the
user request properties.
6.1 User Request Length
We started by varying the user request length according to the available distributions,
expecting the size of the corresponding applications (agent aggregations) to be distributed
according to the length of the user requests, i.e. the longer the user request, the larger the
agent aggregation needed to full it.
We rst applied the Uniform distribution as a control, and graphed the results in Figure
6. The observed frequencies of the application (agent aggregation) size mostly matched the
expected frequencies, which was conrmed by a
2
test; with a null hypothesis of no signicant
dierence and sixteen degrees of freedom, the
2
value was 2.588, below the critical 0.95
2
value of 7.962.
We then applied the Gaussian distribution as a reasonable assumption for the behaviour
of a large group of users, and graphed the results in Figure 7. The observed frequencies of
the application (agent aggregation) size matched the expected frequencies with only minor
variations, which was conrmed by a
2
test; with a null hypothesis of no signicant dierence
and sixteen degrees of freedom, the
2
value was 2.102, below the critical 0.95
2
value of
7.962.
Finally, we applied the Power Law distribution to represent a relatively extreme variation
in user behaviour, and graphed the results in Figure 8. The observed frequencies of the
application (agent aggregation) size matched the expected frequencies with some variation,
which was conrmed by a
2
test; with a null hypothesis of no signicant dierence and
sixteen degrees of freedom, the
2
value was 5.048, below the critical 0.95
2
value of 7.962.
There were a couple of minor discrepancies, similar to all the experiments. First, there
were a small number of individual agents at the thousandth time step, caused by the
typical user behaviour of continuously creating new services (agents). Second, while the
chi-squared tests conrmed that there was no signicant dierence between the observed
and expected frequencies of the application (agent aggregation) size, there was still a bias to
larger applications (solutions). Evident visually in the graphs of the experiments, and evident
numerically in the chi-squared test of the Power Law distribution experiment as it favoured
shorter agent-sequences. The cause of this bias was most likely some aspect of bloat not fully
controlled.
11
0
5
10
15
20
25
30
35
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
F
r
e
q
u
e
n
c
y
Application (Agent Aggregation) Size
Expected Observed
Figure 6: Graph of Uniformly Distributed Agent-Sequence Length Frequencies:
The observed frequencies of the application (agent aggregation) size mostly
matched the expected frequencies.
0
10
20
30
40
50
60
70
80
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
F
r
e
q
u
e
n
c
y
Application (Agent Aggregation) Size
Expected
Observed
Figure 7: Graph of Gaussian Distributed Agent-Sequence Length Frequencies:
The observed frequencies of the application (agent aggregation) size matched
the expected frequencies with only minor variations.
0
50
100
150
200
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
F
r
e
q
u
e
n
c
y
Application (Agent Aggregation) Size
Expected
Observed
Figure 8: Graph of Power Law Distributed Agent-Sequence Length Frequencies:
The observed frequencies of the application (agent aggregation) size matched
the expected frequencies with some variation.
12
6.2 User Request Modularity
Next, we varied the user request modularity (number of attributes per atomic service)
according to the available distributions, expecting the sophistication of the agents to be
distributed according to the modularity of the user requests, i.e. the more complicated (in
terms of modular non-reducible tasks) the user request, the more sophisticated (in terms of
the number of attributes) the agents needed to full it.
We rst applied the Uniform distribution as a control, and graphed the results in Figure
9. The observed frequencies for the number of agent attributes mostly matched the expected
frequencies, which was conrmed by a
2
test; with a null hypothesis of no signicant
dierence and ten degrees of freedom, the
2
value was 1.049, below the critical 0.95
2
value of 3.940.
We then applied the Gaussian distribution as a reasonable assumption for the behaviour
of a large group of users, and graphed the results in Figure 10. The observed frequencies
for the number of agent attributes again followed the expected frequencies, but there was
variation which led to a failed
2
test; with a null hypothesis of no signicant dierence and
ten degrees of freedom, the
2
value was 50.623, not below the critical 0.95
2
value of 3.940.
Finally, we applied the Power Law distribution to represent a relatively extreme variation
in user behaviour, and graphed the results in Figure 11. The observed frequencies for the
number of agent attributes also followed the expected frequencies, but there was variation
which led to a failed
2
test; with a null hypothesis of no signicant dierence and ten
degrees of freedom, the
2
value was 61.876, not below the critical 0.95
2
value of 3.940.
In all the experiments the observed frequencies of the number of agent attributes followed
the expected frequencies, with some variation in two of the experiments. This could only
be conrmed statistically, by a
2
test, for the Uniform distribution experiment. In the
Gaussian and Power distribution experiments the
2
tests failed by considerable margins,
most likely because the evolving agent populations were still self-organising to match the
user behaviour, shown by the observed frequencies approaching the expected frequencies, but
not yet suciently to meet
2
tests, because by the thousandth time step (user request event)
each user had placed an average of only ten requests.
Collectively, the experimental results conrm that the self-organised diversity of the
evolving agent populations is relative to the selection pressures of the user base, which
was conrmed statistically for most of the experiments. So, we have determined an
eective understanding and quantication for the self-organised diversity of the evolving
agent populations of our Digital Ecosystem. While the minor experimental failures, in
which the Digital Ecosystem responded more slowly than in the other experiments, have
shown that there is potential to optimise the Digital Ecosystem, because the evolutionary
self-organisation of an ecosystem is a slow process [3], even the accelerated form present in
our Digital Ecosystem.
13
0
100
200
300
400
500
600
5 6 7 8 9 10 11 12 13 14 15
F
r
e
q
u
e
n
c
y
Number of Attributes per Agent
Expected Observed
Figure 9: Graph of Uniformly Distributed Agent Attribute Frequencies: The
observed frequencies for the number of agent attributes mostly matched the
expected frequencies.
0
250
500
750
1000
1250
1500
5 6 7 8 9 10 11 12 13 14 15
F
r
e
q
u
e
n
c
y
Number of Attributes per Agent
Expected
Observed
Figure 10: Graph of Gaussian Distributed Agent Attribute Frequencies: The
observed frequencies for the number of agent attributes again followed the
expected frequencies, but there was variation.
0
500
1000
1500
2000
5 6 7 8 9 10 11 12 13 14 15
F
r
e
q
u
e
n
c
y
Number of Attributes per Agent
Expected
Observed
Figure 11: Graph of Power Law Distributed Agent Attribute Frequencies:
The observed frequencies for the number of agent attributes also followed the
expected frequencies, but there was variation.
14
7 Conclusions
We have conrmed the fundamentals for a new class of system, Digital Ecosystems,
created through combining understanding from theoretical ecology, evolutionary theory,
MASs, distributed evolutionary computing, and SOAs. Digital Ecosystems, where the word
ecosystem is more than just a metaphor, being the digital counterpart of biological ecosystems,
and therefore having their desirable properties, such as scalability and self-organisation. It is
a complex system that shows emergent behaviour, being more than the sum of its constituent
parts.
With the core infrastructure of Digital Ecosystems dened, some intelligence-based
augmentations have been considered for its acceleration and optimisation [7, 4]. In which
the agents were endowed with memory, intelligence and the ability to share information, with
the aim of optimising the ecological dynamics, through the emergent optimisation of the agent
migration patterns. While not yet as sophisticated as swarm intelligence, we would argue it
is a step in that direction, and we would like to investigate further in the future.
The ever-increasing challenge of software complexity in creating progressively more
sophisticated and distributed applications, makes the design and maintenance of ecient and
exible systems a growing challenge [30], for which current software development techniques
have hit a complexity wall [29]. In response we have created Digital Ecosystems, the
digital counterparts of biological ecosystems, possessing their properties of self-organisation,
scalability and sustainability [25]; Ecosystem-Oriented Architectures that overcome the
challenge by automating the search for new algorithms in a scalable architecture, through
the evolution of software services in a distributed network.
Acknowledgments
We would like to thank for helpful discussions Paolo Dini of the London School of Economics
and Political Science. This work was supported by the EU-funded Open Philosophies
for Associative Autopoietic Digital Ecosystems (OPAALS) Network of Excellence (NoE),
Contract No. FP6/IST-034824.
References
[1] W Ashby. Principles of the self-organizing dynamic system. Journal of General
Psychology, 37:125128, 1947.
[2] S. Beer. Decision and Control: The Meaning of Operational Research and Management
Cybernetics. Wiley, 1966.
[3] M Begon, J Harper, and C Townsend. Ecology: Individuals, Populations and
Communities. Blackwell Publishing, 1996.
[4] G Briscoe. Digital Ecosystems. PhD thesis, Imperial College London, 2009.
[5] G Briscoe. Complex adaptive digital ecosystems. In ACM Management of Emergent
Digital Ecosystems Conference, 2010.
[6] G Briscoe and P De Wilde. Digital Ecosystems: Evolving service-oriented architectures.
In IEEE Bio Inspired Models of Network, Information and Computing Systems
Conference, 2006.
15
[7] G Briscoe and P De Wilde. Digital Ecosystems: Optimisation by a distributed
intelligence. In IEEE Digital Ecosystems and Technologies Conference, 2008. (Awarded
Best Track Paper).
[8] G Briscoe and P De Wilde. Computing of applied digital ecosystems. In ACM
Management of Emergent Digital Ecosystems Conference, 2009.
[9] G Briscoe and S Sadedin. Natural science paradigms. In Digital Business Ecosystems,
pages 4855. European Commission, 2007.
[10] G Briscoe, S Sadedin, and G Paperin. Biology of applied digital ecosystems. In IEEE
Digital Ecosystems and Technologies Conference, pages 458463, 2007.
[11] L. Cabral, J. Domingue, E. Motta, T. Payne, and F. Hakimpour. Approaches to semantic
web services: an overview and comparisons. In The Semantic Web: Research and
Applications, pages 225239. Springer, 2004.
[12] E. Cantu-Paz. A survey of parallel genetic algorithms. Reseaux et syst`emes repartis,
Calculateurs Parall`eles, 10:141171, 1998.
[13] Jorge Cardoso and Amit Sheth. Introduction to semantic web services and web process
composition. In Jorge Cardoso and Amit Sheth, editors, Semantic Web Services and
Web Process Composition, pages 113. Springer, 2004.
[14] P. Dini, G. Lombardo, R. Mansell, A.R. Razavi, S. Moschoyiannis, P. Krause, A. Nicolai,
and L Rivera Leon. Beyond interoperability to digital ecosystems: regional innovation
and socio-economic development led by SMEs. International Journal of Technological
Learning, Innovation and Development, 1:410426, 2008.
[15] D Green, N Klomp, G Rimmington, and S Sadedin. Complexity in Landscape Ecology.
Springer, 2006.
[16] H. Haken. Synergetics: an introduction: nonequilibrium phase transitions and self-
organization in physics, chemistry, and biology. Springer, 1977.
[17] F. Heylighen. The science of self-organization and adaptivity. The Encyclopedia of Life
Support Systems, pages 253280, 2002.
[18] F. Heylighen and C. Joslyn. Cybernetics and second order cybernetics. Encyclopedia of
Physical Science & Technology, 4:155170, 2001.
[19] A. King and S. Pimm. Complexity, diversity, and stability: A reconciliation of theoretical
and empirical results. American Naturalist, 122:229239, 1983.
[20] T. Kohonen. Self-organization and associative memory. Springer, 1989.
[21] J Koza. Overview of genetic programming. In Genetic Programming: On the
Programming of Computers by Means of Natural Selection, pages 7378. MIT, 1992.
[22] D. Krafzig, K. Banke, and D. Slama. Enterprise SOA: Service-Oriented Architecture
Best Practices. Prentice Hall, 2004.
[23] J.M. Lehn. Perspectives in supramolecular chemistryfrom molecular recognition
towards molecular information processing and self-organization. Angewandte Chemie
International Edition in English, 29:13041319, 1990.
[24] G. Lendaris. On the denition of self-organizing systems. In Wolfgang Banzhaf, editor,
European Conference on Advances in Articial Life, pages 324325. IEEE Press, 1964.
[25] S Levin. Ecosystems and the biosphere as complex adaptive systems. Ecosystems, 1:431
436, 1998.
16
[26] R. Levins. Some demographic and genetic consequences of environmental heterogeneity
for biological control. Bulletin of the Entomological Society of America, 15:237240,
1969.
[27] F. Leymann, D. Roller, and M. Schmidt. Web services and business process management.
IBM Systems Journal, 41:198211, 2002.
[28] S. Lin, W Punch III, and E Goodman. Coarse-grain parallel genetic algorithms:
categorization and new approach. In Symposium on Parallel and Distributed Processing,
pages 2837. IEEE Press, 1994.
[29] K. Lyytinen and Y. Yoo. The next wave of nomadic computing: A research agenda for
information systems research. Sprouts: Working Papers on Information Systems, 1:120,
2001.
[30] J Marko. Faster chips are leaving programmers in their dust. Technical report, New
York Times, 2007.
[31] F McCabe and K Clark. April-agent process interaction language. In Michael
Wooldridge and Nick Jennings, editors, Intelligent Agents: Workshop on Agent Theories,
Architectures, and Languages, pages 324340. Springer, 1994.
[32] S McCulloch, R Kokoska, O Chilkova, C Welch, E Johansson, P Burgers, and T Kunkel.
Enzymatic switching for ecient and accurate translesion DNA replication. Nucleic
Acids Research, 32:46654675, 2004.
[33] Gurpreet Modi. Service oriented architecture & web 2.0. Technical report, Guru Tegh
Bahadur Institute of Technology, 2007.
[34] J Moore. The Death of Competition: Leadership and Strategy in the Age of Business
Ecosystems. Harvard Business School Press, 1996.
[35] G Nicolis and I Prigogine. Self-organization in Nonequilibrium Systems: From Dissipative
Structures to Order Through Fluctuations. Wiley, 1977.
[36] M Papazoglou and D Georgakopoulos. Service-oriented computing. Communications of
the ACM, 46:2528, 2003.
[37] MP Papazoglou. Service-oriented computing: concepts, characteristics and directions.
In Tiziana Catarci, Massimo Mecella, John Mylopoulos, and Maria Orlowska, editors,
International Conference on Web Information Systems Engineering, pages 312. IEEE
Press, 2003.
[38] V Pham and A Karmouch. Mobile software agents: an overview. IEEE Communications
Magazine, 36:2637, 1998.
[39] R. Rivest. The MD5 message-digest algorithm. Technical report, MIT, 1992.
[40] K. Rothermel and F. Hohl. Mobile agents. In A. Kent and J. Williams, editors,
Encyclopedia for Computer Science and Technology, volume 40, pages 155176. CRC
Press, 1999.
[41] C Shalizi. Causal Architecture, Complexity and Self-Organization in Time Series and
Cellular Automata. PhD thesis, University of Wisconsin-Madison, 2001.
[42] M. Singh and M. Huhns. Service-Oriented Computing: Semantics, Processes, Agents.
Wiley, 2005.
[43] J Stanley and G Briscoe. The ABC of digital business ecosystems. Communications Law
- Journal of Computer, Media and Telecommunications Law, 15(1), 2010.
17
[44] J. Stender. Parallel Genetic Algorithms: Theory and Applications. IOS Press, 1993.
[45] M Ward. Life oers lessons for business, 2004.
[46] D Watts and S Strogatz. Collective dynamics of small-world networks. Nature, 393:440
442, 1998.
[47] D White and M. Houseman. The navigability of strong ties: Small worlds, tie strength,
and network topology. Complexity, 8:7281, 2002.
[48] O. Zimmermann, S. Milinski, M. Craes, and F. Oellermann. Second generation web
services-oriented architecture in production in the nance industry. In John Vlissides
and Douglas Schmidt, editors, Conference on Object-Oriented Programming, Systems,
Languages, and Applications, pages 283289. ACM Press, 2004.
18

You might also like