Ubiquitous Computing and Distributed Agent-Based Simulation

Abstract

As much as ubiquitous computing systems are already claimed to exist in the real world, further development of these systems still pose challenges to computer science that are still quite beyond the state of the art. Two challenges stand out in particular: the complexity of next-generation ubiquitous computing systems, and their inherent scalability issues. This paper aims to establish that agent-based modelling provides a powerful tool in tackling these issues. As an example of a practical solution, readily available, this paper highlights the distributed agent-based simulation infrastructure PDES-MAS as particularly suited for the task. Using the PDES-MAS infrastructure, designers, developers, and builders of next-generation ubiquitous computing systems can, through an iterative agent-based simulation process, gain the required knowledge and information about these systems, without hav-ing precede to deployment of the system itself.

Full text

Page 1

Ubiquitous Computing and Distributed Agent-based Simulation
B.G.W. Craenen, G.K. Theodoropoulos
School of Computing
University of Birmingham
Birmingham, United Kingdom
b.g.w.craenen@cs.bham.ac.uk
g.k.theodoropoulos@cs.bham.ac.uk
Abstract—As much as ubiquitous computing systems are
already claimed to exist in the real world, further development
of these systems still pose challenges to computer science that
are still quite beyond the state of the art. Two challenges stand
out in particular: the complexity of next-generation ubiquitous
computing systems, and their inherent scalability issues. This
paper aims to establish that agent-based modelling provides
a powerful tool in tackling these issues. As an example of
a practical solution, readily available, this paper highlights
the distributed agent-based simulation infrastructure PDES-
MAS as particularly suited for the task. Using the PDES-
MAS infrastructure, designers, developers, and builders of
next-generation ubiquitous computing systems can, through
an iterative agent-based simulation process, gain the required
knowledge and information about these systems, without hav-
ing precede to deployment of the system itself.
Keywords-distributed simulation; agent-based systems; ubi-
quitous computing;
I. INTRODUCTION
While computing devices have become ever more per-
vasive throughout every day life in modern society, the
size and capabilities of the computing devices has changed
dramatically as well. Computing devices, or computers,
have become smaller, more capable and complex, and have
become more and more interconnected. Computers have
moved from large desktop models through smaller comput-
ers that can be carried onto a person’s lap to mobile (phone)
devices and computing pads that can be carried in a person’s
pocket. And the miniaturisation of computing devices is
still ongoing with already promising work is being done
on computing devices of a cubic centimeter and smaller.
In the post-desktop model of human-computer interaction,
ubiquitous computing allows for information processing that
is thoroughly integrated into everyday objects and activities.
Someone using ubiquitous computing in the course of ordi-
nary activities commonly engages many computing devices
and systems simultaneously, and may not necessarily even be
aware that they are doing so. A considerable advancement
from the desktop paradigm already, ubiquitous computing
is more formally defined as “machines that fit the human
environment instead of forcing humans to enter theirs” [1].
The ubiquitous computing paradigm is also described
as pervasive computing or ambient intelligence [2], where
each term emphasizes slightly different aspects. Rather than
propose a single definition for ubiquitous computing and
these related terms, a taxonomy of properties for ubiquitous
computing has been proposed, from which different kinds
or flavours of ubiquitous systems and applications can be
described [3].
At the core of all models of ubiquitous computing is
a shared vision of small, inexpensive, robust networked
processing devices, distributed at all scales throughout ev-
eryday life, turned generally to distinct common-place ends.
Because this vision is still quite beyond the current state
of the art, ubiquitous computing presents challenges across
computer science: in systems design and engineering, in sys-
tems modelling, and in user interface design. Contemporary
human-computer interaction models remain inappropriate
and inadequate to the ubiquitous case and suggests that the
interaction paradigm appropriate to a fully robust ubiquitous
computing framework has yet to emerge.
This even though there is some recognition that in many
ways humanity already lives in a ubiquitous computing
world. In [4], Manuel Castells suggests that there is an
ongoing shift from already-decentralised, stand-alone mi-
crocomputers and mainframes towards entirely pervasive
computing. In his model, the example of the Internet is
used as the start of a pervasive computing system with
the logical progression from that paradigm, a system where
networking logic becomes applicable in every realm of daily
activity, in every location, and in every context. Castells
envisages a system where billions of miniature, ubiquitous
inter-communication devices will be spread worldwide, “like
pigment in the wall paint”, the latter echoing developments
in speckled computing and smart dusts.
But that still leaves computer scientists the task of solving
the problems in systems design and engineering, systems
modelling, and user interface design. Leaving the latter
for the human-computer interaction specialists, this paper
will focus on the former two and focusses on discussing
how agent-based modelling and simulation, and in particular
distributed agent-based modelling and simulation, can play
an important part.
And while the use of distributed agent-based modelling
and simulation in ubiquitous computing in and of itself

Page 2

is not new, the application of these techniques is often
topical and ad-hoc. More often than not it is applied to
support solving the particular problem itself. And while this
is obviously understandable, it ignores what the combination
could contribute towards understand ubiquitous computing
itself. In that respect, advances in the distributed multi-agent
simulation field are there to be leveraged as well. And as an
example, this paper will focus on the PDES-MAS system
as a distributed multi-agent simulation system particularly
well-suited for support ubiquitous computing in general.
The rest of the paper is then organised as follows. In
section II we give a description of agent-based modelling
and simulation with section III describing how it can be used
to further ubiquitous computing in general. Section IV then
describes the need for distribution with section V providing
a means to do so with a description of the PDES-MAS
platform. Finally in section VI we offer some conclusions
and further discussion.
II. AGENT-BASED MODELLING
An agent-based model (ABM) is a class of computational
models for simulating the actions and interactions of auto-
nomous agents in order to assess the effects on a system
as a whole. Instead of ABM, sometimes the term multi-
agent system (MAS), multi-agent simulation (also MAS),
or individual-based models is used. The autonomous agents
can describe both individual or collective entities such as or-
ganisations or groups within the system. ABMs simulate the
simultaneous operations and interactions of multiple agents
in an attempt to re-create and predict the appearance of
complex phenomena. For this is a process of emergence from
the micro to the macro level of the system. A key notion here
is that simple behavioural rules generate complex behaviour,
a principle extensively adopted in the modeling community.
Individual agents are typically characterised as boundedly
rational, presumed to be acting in what they perceive as
their own interests.
Most computational modeling research describes systems
in equilibrium, or as moving between equilibria. ABMs,
however, uses simple rules to generate complex behaviour,
usually resulting in far more complex and dynamic be-
haviours to examine. The three ideas central to ABM are
social agents as objects, emergence, and complexity. Be-
cause ABMs consist of dynamically interacting rule-based
agents, the systems within they interact, or of which they
are part, can create real world-like complexity if the agents
are: intelligent and purposeful as well as situated in space
and time. The modeling process itself is best described as
inductive, with the modeler making assumptions relevant to
the situation and then watching phenomena emerge from the
agents’ interactions, where ABMs are generally seen to be
complementing traditional analytic methods.
Where analytic methods focus on characterising the equi-
libria of systems, ABMs allow the possibility of generating
those equilibria. It is this generative contribution to the
analysis that may be the most mainstream of the poten-
tial benefits of ABMs. The ABMs’ ability to explain the
emergence of higher order patterns; their ability to identify
those moments in time in which interventions have extreme
consequences; and their ability to distinguish among types
of path dependency; all make ABMs valuable analytic tools
to use. Rather than focusing on stable states and equilibria,
the ABM focuses on the system’s robustness, i.e., the ways
that complex systems adapt to internal and external pres-
sures so as to maintain their functionalities. Harnessing that
complexity requires consideration of the agents themselves:
their diversity; connectedness; and their level of interactions
are all worthy of consideration.
ABMs have been used since the mid-1990s to solve
a variety of business and technology problems. Examples
include supply chain optimisation and logistics, modelling
of consumer behaviour, social network effect, distributed
computing, and traffic congestion. In these and other ap-
plications, the system of interest is simulated by capturing
the behaviour of individual agents and their interconnections.
ABM tools can be used to test how changes in individual be-
haviours will affect the system’s emerging overall behaviour.
In addition, ABMs have been used to simulate information
delivery in ambient assisted environments [5].
III. ABM AND UBIQUITOUS COMPUTING
Ubiquitous computing systems lend themselves well to
be modelled by ABMs. The system itself consists of many
computing devices, each with a clearly defined purpose, act-
ing independently, and interacting with the other computing
devices. The correlation between the computing devices that
make up the ubiquitous computing system and autonomous
and intelligent agents in an ABM is as such relatively
straightforward. Using the ubiquitous computing system are
humans, and their interactions with the system are also
clearly defined and predominately independent from each
other. They too can be encapsulated as agents in an ABM
[6].
Modelling a ubiquitous computing system and it’s users is,
as such, relatively straightforward. But what is the benefit of
doing so? The answer is that, the more complex a ubiquitous
computing system becomes, the harder it becomes to design,
develop and build. A ubiquitous computing system is almost
by definition a dynamic and changing environment, and one
where an equilibrium of provided services and functionali-
ties with the requirements placed on the system is something
to be aimed for in general. Of relevance here is that the
system itself should be able to handle the dynamically
changing circumstances as well, i.e., be robust.
Moreover, it is important to also capture and predict the
emergence of those requirements, as they happen. To do so,
modelling and simulation provides an obvious benefit with
ABMs particularly suited. By encapsulating the ubiquitous

Page 3

computing system itself into agents, and having it interact
with users, also encapsulated in agents, allows designers, and
developers of the system to predict the use, requirements,
and robustness of the system beforehand. In addition, the
ABM allows these interested parties to analyse how these
aspects of the ubiquitous computing systems came to be.
Not only the emergence of the system itself can be studied,
but emergence of behaviours when the system is altered or
extended can be studied as well. If ubiquitous computing is
the emergence of an embedded computing system, simulat-
ing the system through ABMs, and analysing the emergence
behaviour, even before actually putting it into practice, can
provide both information and behavioural patterns invaluable
to the design and development of the system itself.
ABMs can provide further information on three aspects
of ubiquitous computing valuable for ubiquitous computing:
the interactions between the users and the system, as well as
the system internally; the requirements placed on the system
by the users; and the required functionality constraints of the
individual parts of the system. As in an ABM both the users
and the system itself are modelled as collections of agents,
this boils down to investigating the interactions between the
agents, and the behavioural rules of the agents themselves.
This is, in essence, divide and conquer system development
in action, with the ABM breaking down the complexity of
the whole system and its users into relatively constituent
parts governed by simple rules acting and interacting on and
with each other.
The ABM then serves as a paradigm for reducing the
complexity of the system as a whole into the complexity of
its constituent parts and the complexity of the rules by which
the controllers of these parts interact. The process, following
the ABM paradigm, is in essence inductive, with a set of
what-if scenarios providing information of the system with-
out physical implementation or deployment being required.
Although physical factors do certainly inform the ABM,
concentrating primarily on the behaviour of the system
allows for a level of abstraction, inductively diminished
over time with increasingly complex models, that makes
experimental prediction possible and relatively cheap. An
ABM of a ubiquitous computing system allows the designer,
developer, and builder of the system to predict and define:
the interactions within the system; the requirements of each
constituting part of the system; and the functionality, or rule-
set, of (the controller of) each constituting part (for examples
see [7], [8], for an alternate architecture using ABM for
ubiquitous computing see [9]).
The idea of using ABM for ubiquitous computing is
not entirely new however, and several research projects
have made use of it. The FireGrid project (e.g. [10]) for
example uses distributed and grid computing for emergency
response applications relying on ABM techniques. But there
is no unified approach to develop these sort of applications,
all seem to be ad-hoc solutions for particular problems.
And in general, advances in the distributed ABM field are
available for application. In this respect we here like to
focus on PDES-MAS, as a distributed agent-based modelling
infrastructure well-suited to be used in support of agent-
based modelling of ubiquitous computing systems. The use
of a DSM systems is well suited to support the fluid nature
of ubiquitous computing systems with respect to the location
of the computing devices that form such an important part
of it. The optimistic synchronisation technique complements
the dynamic nature of ubiquitous computing systems equally
well. With user interaction with those computing devices,
and the interaction between the ubiquitous computing sys-
tem devices so unpredictable, the PDES-MAS infrastructure
and system provides a viable alternative for handling such
circumstances.
IV. DISTRIBUTED SIMULATION
As much as ABM can provide a means for tackling the in-
herent complexity of ubiquitous computing systems, part of
this complexity is a problem shared with ABM: scalability.
Larger scale systems, be they ubiquitous computing systems,
or ABMs in general, become more complex the larger they
become. Ubiquitous computing systems, if one accepts the
notion that they are already existent in society, are becoming
larger, and the complexity of its constituent parts as well as
its interactions has increased likewise. Without introducing
increasing levels of abstraction into the model, ABMs of
such ubiquitous computing systems will have to follow suit.
Part of this problem can be handled by using the induc-
tive approach central to agent-based modelling. Modelling
ubiquitous computing systems would start with a relatively
small, highly abstract, low complexity ubiquitous computing
system, to gradually, over various iterations, remove more
and more layers of abstraction, introducing more and more
constituent parts of the system. The process in itself will
be valuable, and with proper validation at each step of the
process, a close approximation of the entire system will be
arrived at.
This leaves, however, technical limitations with the si-
mulation itself. Building and refining an ABM is not only
an inductive process, but a technical one as well. ABMs
depend on an experimental simulation process, and the
analysis of the results thereof. Validation of the model is then
matching the behaviour of the ABM with the observed or
expected behaviour of the system at each step in the process.
Running those simulations however is possible only when
enough computing resources are available to do so. With
increasingly larger and complex ABMs and simulations, the
computational resources available on one computing system
will inevitably run out. The solution to this is to combine
and effectively utilise the computational resources of several
computing systems; in other words: Distributed Agent-based
Simulation.

Page 4

Distributed agent-based simulation, or Distributed Multi-
Agent Systems (DMAS), focuses on distributing a ABM,
or MAS, over multiple computer platforms. Various ap-
proaches, resulting in various simulation-engines, have been
proposed [11]. Some simulation-engines focus on partition-
ing the simulation topology into several semi-autonomous
regions, each distributed over the computing resources avail-
able. Others partition the shared state of the simulation using
a Distributed Shared Memory (DSM) system, and distribut-
ing this over the computing resources available. These, and
other approaches, have specific limitations and advantages,
leading to a differentiation of the various approaches to the
systems they are used to support. In this paper we use the
PDES-MAS framework [12] as an example of a DSM of the
later approach.
The shared state partitioning approach, using a DSM sys-
tem, can be shortly described as follows. An ABM consists
of a number of agents interacting with themselves or other
agents. The environment in an ABM can then be described
as either unchanging (static), or be encapsulated in yet other
agents. The agents can be said to go through a sense-think-
act cycle, where each agent first senses its environment and
other agents, considers (thinks about) this input, and then
acts on it. Most often ABMs of this type are event-based,
where each action or interaction consists of a time-stamped
event. Individual agent behaviour is determined during the
think sub-cycle, usually according to a rule-set of varying
complexity.
During the course of the simulation, called an exper-
iment, each agent maintains a state. That is, a variable-
set that includes all the information encapsulated in the
agent. Determining the behaviour can require the agent to
sense information encapsulated in other agents. State can
be internal (private) and not shared or not visible to other
agents, or external (public) and shared or visible to other
agents. As such, agents can only sense the shared state of
other agents.
The shared state partitioning approach focusing on dis-
tributing the shared state of agents. The internal state of
an agent can not be sensed and need not be distributed.
By distributing the agents themselves over the available
computing resources, exceeding local computing resources
is avoided. The DSM then makes available to all computing
resources all shared state of all agents over the computing
network. Distributing an ABM using the DSM approach then
consists of distributing the agents (and the static part of
the simulation) over the available computing resources such
that the local limitations are not exceeded, while making
available the shared state of all agents to all other agents
on all the available computing resources. The focus of the
DSM is then to provide a scalable method of getting access
to the collective shared state of the agents in the simulation.
Scalability is important because while the simulation can
still only be run on with sufficient computing resources, a
scalable system can grow when more computing resources
are available and required.
V. PDES-MAS
In this paper we will use the PDES-MAS framework as
an implementation of a DSM system for ABMs [12]. It
implements a DSM structure where the shared state of the
agents is represented by Shared State Variables (SSV). SSVs
are data-structures that store the time-stamped history of
values of a particular variable over time [13]. The collective
state of the shared variables, encapsulated in SSV, can then
be seen as similar to the ABM simulation’s space-time
memory [14], [15]. The shared variables stored in SSVs
offer a natural representation of the simulation context when
interactions between agents are described as interactions on
these variables. Consequently we assume that agents alter
the state of the ABM simulation by interacting with these
SSVs in an event-based fashion.
The PDES-MAS framework then acts to make the space-
time memory available across the different distributed com-
puter platforms without exposing the exact way in which
the memory is accessed or organised internally. Access to
the SSVs is provided through read and write operations
performed as events. These events allow the reading of a
value stored in the SSV (ID-query), and writing a value in a
certain SSV (ID-write) [16]. Reading of values over a range
of SSVs satisfying a condition (associative memory) is an
extension of the ID-query event, and is called a range-query.
Range-queries are used to retrieve aggregate information
during the sense-cycle, e.g., retrieving the location of all
agents within viewing range [17].
Following the parallel distributed event simulation (PDES)
paradigm, agents in the ABM are assigned to Logical
Processes (LP), known as either Agent Logical Processes
(ALP) that handle the agents themselves, or Communication
Logical Processes (CLP) where SSVs are assigned. An ALP
potentially models more than one agent, with multiple ALPs
allowed concurrent access to the set of SSVs associated to
the agents by connecting the ALPs to a tree-like network
of CLPs. Conceptually, the ALP provides an access-node
for the ABM simulation to the PDES-MAS framework.
SSVs are then distributed among the CLPs in a scalable
and balanced manner. Figure 1 shows a depiction of the
PDES-MAS framework with 4 ALPs, 3 CLPs, and 4 SSVs.
Agents in a ABM are assumed to be intelligent with
agents interacting with each other and their environment
(partly also captured within agents). Because agents act inde-
pendently and intelligently, based on sensed information, it
is often hard to predict these interactions in advance, indeed
discover them at all. As such, a defining characteristic of
agents is their autonomy [18]. Moreover, how agents interact
is often the primary goal of the simulation, particularly
when considering ubiquitous computing. The sense-think-
act cycle is then used to determine the behaviour translated

Page 5

VV
V
V
12
CLP0
ALP
CLP CLP21
ALPALP
124
3
4
ALP3
Figure 1. The PDES-MAS framework.
into actions with agents, over the life-time of the simulation,
going through many of the cycles. Within the sense-cycle
the PDES-MAS framework provides ID-query and Range-
query events, while during the act-cycle it provides ID-write
events. ALPs link to the leaf (parent) CLPs in the tree-
like structure with ALPs, issuing requests to access SSVs
through their parent CLPs. If the SSV is not assigned to the
parent CLP, the access requests are passed along the tree
until the SSV is located. The relevant data is the returned
along the same route in reverse, back to the parent CLP,
and from there to the ALP. Control messages internal to the
PDES-MAS framework are conveyed through the tree in a
similar way.
The PDES-MAS framework is used to provide simul-
taneous distributed access to the shared data of the ABM
simulation and in this context maintaining data consistency
is an important part of the functionality of the framework.
Two main synchronisation mechanisms for maintaining data
consistency can generally be recognised: conservative syn-
chronisation, and optimistic synchronisation. Conservative
synchronisation disallows conflicting access to the data
by predicting when conflicting access will occur to then
apply strict access-rules (pre-emptive locking). Optimistic
synchronisation allows free access to the data at all times,
but repairs data inconsistency afterwards through a roll-back
mechanism [13].
Data consistency in PDES-MAS, and as such synchronis-
ing the events received from the ALP, is the responsibility
of the CLP. The PDES-MAS framework uses optimistic
synchronisation (see [19], [20], [21], [13], [22], [23], [24]).
Each SSV is associated with a list of Write Periods (WP),
each representing the values of the SSV at different times
throughout the simulation. When a WP is invalidated by
a straggler write, any agents associated to the ALPs that
have read that WP with a subsequent time-stamp as the
straggler write will be asked to roll-back to the time-stamp
preceding the one of the straggler write. A rolled-back
agent then resumes its time-progressing from before the data
was invalidated, i.e., any inconsistenties in the data will be
ignored and the data consistency itself will be repaired for
that agent.
The CLP tree structure in the PDES-MAS framework is
reconfigured dynamically and automatically, reflecting the
interaction patterns of the agents in the overlaying ABM
simulation based on the access patterns of the SSVs. SSVs
accessed more frequently by an ALP are moved closer
in the CLP tree structure to that ALP, i.e., the SSV is
moved towards the parent leaf CLP of the ALP. The aim
is to concurrently minimise the average number of hops
required to access the SSV, as well as to reduce the load
imbalance between the CLPs. In principle reconfiguration
of the CLP tree structure can be achieved by creating or
deleting CLPs; moving ALPs to different parent CLPs, or by
migrating SSVs between CLPs. The PDES-MAS framework
described in this paper implements a fixed binary tree-
structure with leaf CLPs hosting a fixed and constant number
of ALPs. Only SSVs are migrated through the tree to achieve
redistribution [25].
VI. CONCLUSION
This paper discusses how distributed agent-based mod-
elling and simulation can be a useful tool for designing,
developing, and building ubiquitous computing. The spread
and availability of computing resources and devices has
increased dramatically in resources. The way these resources
are made available, how these devices connect with each
other, and how they interact with their human users has
become ever more complex. Agent-based modelling and
simulation can provide an invaluable tool to understanding
and analysing this complexity, as well as providing a means
for designing, developing, and building future (extensions
to) ubiquitous computing systems.
But as ubiquitous computing becomes ever more per-
vasive and as ubiquitous computing systems increase in
scale, so will the agent-based models of these systems if
they are to provide analysis and information of sufficient
fidelity. Increasingly large-scale agent-based models will
inevitably out-grow the computing resources available on
single computer platforms. For this problem, distributed
multi-agent systems provide an answer. Considering the