1939-1374 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSC.2016.2533348, IEEE
Transactions on Services Computing
1939-1374 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSC.2016.2533348, IEEE
Transactions on Services Computing
1
Goal-Driven Service Composition in Mobile and
Pervasive Computing
Nanxi Chen, Nicol
´
as Cardozo and Siobh
´
an Clarke
Abstract—Mobile, pervasive computing environments respond to users’ requirements by providing access to and composition of
various services over networked devices. In such an environment, service composition needs to satisfy a request’s goal, and be
mobile-aware even throughout service discovery and service execution. A composite service also needs to be adaptable to cope with
the environment’s dynamic network topology. Existing composition solutions employ goal-oriented planning to provide flexible
composition, and assign service providers at runtime, to avoid composition failure. However, these solutions have limited support for
complex service flows and composite service adaptation.
This paper proposes a self-organizing, goal-driven service model for task resolution and execution in mobile pervasive environments.
In particular, it proposes a decentralized heuristic planning algorithm based on backward-chaining to support flexible service discovery.
Further, we introduce an adaptation architecture that allows execution paths to dynamically adapt, which reduces failures, and lessens
re-execution effort for failure recovery. Simulation results show the suitability of the proposed mechanism in pervasive computing
environments where providers are mobile, and it is uncertain what services are available. Our evaluation additionally reveals the
model’s limits with regard to network dynamism and resource constraints.
Index Terms—Services Composition, Requirements Driven Service Discovery, Pervasive computing, Mobile Computing.
F
1 INTRODUCTION
Pervasive computing environments respond to human
users’ requirements by providing access to various resources
over networked systems. Such environments have evolved
from closed (special purpose) and static to open and dy-
namic, embracing a large number of third-party mobile en-
tities (e.g., wearable technologies, smart phones, etc.). Such
open, mobile pervasive computing environments are likely
to incorporate abundant functionalities and heterogeneous
smart devices that have the potential to collaborate.
Service-oriented computing’s (SOC) packaging of het-
erogeneous resources as services that are discoverable,
accessible, and reusable has emerged as an important
paradigm in pervasive computing environments [1] [2].
SOC provides unifying interfaces for services to ease users’
access via communication networks. To address a partic-
ular requirement, a combination of multiple services may
be required, and so a fully-functional service composition
process will tackle potentially complex and mutable user
requests [3] with flexible composition of value-added ser-
vices, rather than simply providing information query or
file transmission services.
Given the environment’s openness and dynamism, such
a composition process faces significant challenges (see the
dash rectangle in Fig.1 ):
• Flexible service composition- services of interest are
independently deployed and maintained by different
service hosts that form a wireless network in ad hoc
ways. An individual service host has only a local
• N. Chen, N. Cardozo and S. Clarke are with the Department of Computer
Science, Trinity College, Dublin, Dublin 2, Dublin, Ireland.
E-mail: nchen@tcd.ie, cardozon@scss.tcd.ie, Siobhan.Clarke@scss.tcd.ie
system view because of its limited communication
ranges and resources. It is a challenge for service
hosts to collaborate to find possibly more potential
service providers to reduce composition failure.
• Adaptable composites- service hosts’ mobility leads
to changes to the network topology as well as the
established communication channel between com-
posite participants (service links). Such changes may
result in service link loss during service execution,
which further causes service execution path loss.
Efficient interactions between composite participants
are required to reduce the dependency on such
an error-prone communication channel. In addition,
composite services must be adaptable to increase
the chance that results can be delivered even when
a communication channel between their composite
participants drops.
This paper focuses on flexible service composition
and adaptable composites, assuming trustworthy service
providers. The remaining issues in Fig.1, service heterogene-
ity and third-party providers, are out of this paper’s scope
and are discussed further in Section 5.
As a motivating scenario (Fig.2), Anne is in a shopping
mall’s car park with her 1-year-old son. She would like to
get a step-free route to a shop that sells nappies and from
there to the nearest baby-changing facility in the mall, using
her smart watch. The mall offers an official website and a
mobile application for information browsing. But Anne’s
smart watch is incompatible with the application and its
screen is too small to display the route properly. However,
there is a pervasive computing environment in the mall
including various embedded devices owned by customers,
shop clerks, taxi drivers, information desks, or house keep-
1939-1374 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSC.2016.2533348, IEEE
Transactions on Services Computing
1939-1374 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSC.2016.2533348, IEEE
Transactions on Services Computing
2
Fig. 1. Open issues in service composition in pervasive computing
environment (a) service deployment device, (b) service composition
problems, (c) service composition requirements; the red dash line rep-
resents the challenges addressed in this paper.
Fig. 2. Services and a service composite in a shopping mall’s pervasive
computing environment
ers. These devices can package their capabilities, like GPS,
navigation, facility routing, taxi booking or indoor map to be
accessible via network connections. Anne’s smart watch has
been configured to incorporate surrounding communication
networks to make use of available resources [4]. Thus, her
smart watch may be able to directly get an audio route
stream via an ad hoc network that forms from different
devices in the pervasive computing environment, such as a
personal-shopper’s phone, a nearby car’s satellite navigator,
and the mall’s information kiosk. The pervasive computing
environment helps Anne avoid browsing the website, in-
creasing the likelihood of matching all her hardware and
software capabilities to get the routing result she needs.
Open issues with current research on service composi-
tion (e.g., goal-oriented planning [5], open workflows [6]
[7] [8], adaptable composition [9] [10] [11] and AI planning
[12] [13] ) are that i) existing service composition models are
inflexible; and ii) the composition models have limited sup-
port for handling composite service adaptation. Specifically,
workflow-driven solutions [6] [7] [8] model user require-
ments as a complex task request in the form of workflows
(i.e., abstract composite services). However, dynamic service
availability is unpredictable. Such an exactly-defined task
request removes the possibility of using services that may
contribute to the user’s request, but are not outlined in the
predefined workflow. For example, with a workflow ”get
a shop location→ get ramp access points → get a baby-
changing facility’s location → get an audio route”, Anne’s
request cannot be solved by the environment shown in Fig.2.
Goal-driven solutions [13] [14] [15] [16] are more flexible,
and model service composition problems by dynamically
composing multiple services when an individual one for
a service request is unavailable. However, existing goal-
oriented approaches either handle only sequential service
flows for such a composite service [15] [16], or employ AI
planning algorithms that support various service flows but
requires planning infrastructures [14]. In addition, existing
approaches towards composite service adaptation and re-
planning [12] [10] [11] [17] [18] rely on structured net-
works, central controllers or service repositories. Keeping
such network topology or central controllers up to date in
highly dynamic environments requires periodic interactions
over error-prone channels to monitor whether a composite
participant is absent.
This paper proposes a self-organizing, Goal-driven ser-
vices Composition model in Mobile and pervasive comput-
ing environments, called GoCoMo. In particular, it leverages
a decentralized planning algorithm based on backward-
chaining [19] to support flexible service querying. This
algorithm supports complex service flows such as parallel
service flows and hybrid service flows. Further, GoCoMo
introduces an on-demand adaptation overlay that allows ex-
ecution paths to dynamically adapt, which reduces failures.
This model uses, with some extensions, an opportunistic re-
source allocation scheme [9] to lock a participant provider’s
resource for a composite only when this provider’s service
is about to execute.
This work extends our existing algorithm [20] with
three contributions: First, the number of potential services
available to the system may expand as providers enter the
system’s scope, automatically merging with the existing
service-flows to resolve user requirements. Second, dynamic
candidate execution path adaptation and selection, based on
path reliability, are introduced to accomplish service execu-
tion with less failure. Third, a novel heuristic service request
transition model is employed that prevents service requests
flooding the network, which trades off traffic overhead with
service discovery scope in service planning.
The evaluation compares this work with a distributed
goal-driven service composition approach [13] [21]. Simula-
tion results show the suitability of the proposed mechanism
in a pervasive computing environment where providers are
mobile and it is uncertain what services are available.
The reminder of this paper is organized as follows.
Section 2 defines the problem and introduces the service
1939-1374 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSC.2016.2533348, IEEE
Transactions on Services Computing
1939-1374 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSC.2016.2533348, IEEE
Transactions on Services Computing
3
model that supports service composition. Section 3 de-
scribes the service composition algorithm. Section 4 presents
the evaluation and result. Section 5 discusses the validity of
this approach and remaining challenges in detail. Section 6
outlines related work. Section 7 summarizes this work and
discusses the future work.
2 SERVICE MODEL
The target network for this work is open and so may
be unstructured or semi-structured with some autonomic
service hosts. In other words, service hosts do not follow
any authority in the network requiring them to schedule
services for service requesters. Service hosts of interests
will be cooperative and be prepared to process composite
requests.
A service composite for complex user tasks can be mod-
elled as a restrictive data transition in which the system data
changes from initial data (i.e., a user’s input parameters) to
goal data (the requested output data), while satisfying all
the requested functionalities or constraints. A participating
service for the composite, packaging its resources (e.g., data,
functionality), can support all or a part of (based on its
resource provision’s granularity) the data transition. This
paper assumes services’ functions and I/O parameters are
semantically annotated using globally understood seman-
tics and language, and able to match a service request
with semantic matchmakers [20]. We assume such semantic
service annotations are kept in local service hosts (devices)
and can be advertised using probe messages. A service’s
invocation must be based on all the specified input data,
and local devices can form an ad hoc network, cooperating
with each other to resolve a complex user task.
A service is described as S = hS
f
, IN, OU T, QoS
time
i,
where S
f
represents the semantic description of service
S’s functionality. IN = {hIN
S
, IN
D
i} and OUT =
{hOUT
S
, OUT
D
i} describe the service’s input and output
parameters as well as their data types respectively. For
this work, execution time QoS
time
is the most important
quality of service (QoS) criterion as delay in composition
and execution can cause failures [9]. A service composition
model should select services with short execution time to
reduce delay in execution.
Complex user tasks can be modeled as a service request
R = hR
id
, I, O, F, Ci, where R
id
is a unique id for a request.
The set F represents all the functional requirements, which
consists of a set of essential while unordered functions. The
composition constraints set C are execution time constraints.
A composition process fails if C expires and the client
receives no result during service execution. A service com-
posite request also includes a set of initial parameters (input)
I = {hI
S
, I
D
i} and a set of goal parameters (output)
O = {hO
S
, O
D
i}.
Given the mobile nature of devices of interest, this
paper discusses service provision in networks with high
dynamism where the network topology is likely to change
faster than it takes to entirely complete service composition
or execution. A client is assumed to have limited resource
and communication range, which are insufficient to ag-
gregate enough service providers for a task. The solution
proposed in this paper addresses mobility by providing
Fig. 3. General distributed backward-chaining model for service compo-
sition
a fully decentralized self-organizing/adapting composition
model.
3 SERVICE COMPOSITION
A service composition process can be modelled as classic AI
backward-chaining [12]. A backward-chaining process, also
known as goal-driven reasoning [13], starts by searching
for the knowledge that can infer a request’s goals (conse-
quences), and then the request is resolved backward from
the goal to the request’s antecedents, by converting the
goal into subgoals, resolving back through these subgoals
(e.g., Goal a → c in Fig.3). The process finds a solution
when all the antecedents are reached. A general goal-driven
reasoning process for service composition is shown in Fig.3.
The initiator issues a service request a → d to start the
process, which relies on distributed knowledge bases stored
in local service hosts (planners). In each step of the service
discovery, a part of the request’s goal can be solved (e.g.,
Goal : a → d can be partially solved on Service 1 that
provides data transition of c → d), and the remaining
request (Goal : a → c) is forwarded to the next hop
service providers (i.e., Service
2). In such a process, it is
the request’s goal that determines which services will be
selected and used. This process produces more flexible plan-
ning results than that of workflow-driven approaches, as it
considers service discovery as an open-ended problem and
dynamically generates composite services according to run-
time service availability. For example, in Anne’s scenario,
she can specify her requirements as a goal: an audio step-free
route to a store and then to a baby-changing facility which will
be resolved hop-by-hop and eventually supported by a com-
posite service. As illustrated in Fig.2, the planning resolves
first to AudioNavigator, back to RampAccessChecker, then
to StoreRouter as well as FacilityRouter, to StoreLocator, to
IndoorMap, and finally to StoreQuery.
Modelling service composition as such a process has
been explored in infrastructure-rich networks where compo-
sition planning is based on infrastructures like repositories
[22] or pre-existing overlay networks [7] . However, this
kind of infrastructure, as mentioned in Section 1, is not suit-
able for our target environment, and neither are the existing
goal-driven reasoning processes. To allow mobile pervasive
computing environments to benefit from the flexibility that
such a goal-driven service composition brings, this paper
1939-1374 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSC.2016.2533348, IEEE
Transactions on Services Computing
1939-1374 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSC.2016.2533348, IEEE
Transactions on Services Computing
4
Fig. 4. Decentralized service composition model
proposes a service composition model that handles the
following issues:
• Goal-Driven Service Discovery- is handled via a
composite participant cooperation mechanism to co-
ordinate distributed knowledge bases and indepen-
dent planners to support the generation and the
maintenance of various service flows. (Section 3.1)
• Opportunistic Service Execution- is handled via on-
line adaptable reasoning to create awareness of and
compose potentially better services that may appear
during service execution. (Section 3.2)
• Heuristic Discovery Checking- is handled via a
distributed heuristic discovery mechanism based on
QoS attributes to increase the likelihood of time-
efficient services being selected during execution and
prevent composite requests flooding the network.
(Section 3.3)
The model captures a goal-driven reasoning algorithm to
support a fully decentralized service composition process,
which is modelled as a state-transition diagram in Fig.4.
A transition between these illustrated states is triggered
by message communication events (e.g., msgIn, msgOut)
or local conditions (e.g., participate, usable). The model in-
cludes i) a goal-driven service discovery protocol to discover
and link all the reachable and usable services, and ii) an
opportunistic service composition protocol that is based on
the discovery result, to select, compose and execute services
hop-by-hop on demand.
In particular, the global service discovery (listening − a
state) starts when a client sends out a composite request
to look for composite participants. Composite participants
in this model are service hosts who are capable of reacting
to and reasoning about a composite request. They are also
responsible for invoking their subsequent services during
execution. From a composite participant’s perspective, the
local discovering loop starts in the listening state and ends
when the composite participant is invoked (the invoking
state), while a local composing process, starts in the invoking
state and ends in a composition-handover state.
3.1 Goal-Driven Service Discovery
Global service discovery is a process to group composite
participants. Each composite request can be resolved par-
tially (or completely), and the remaining request is for-
warded to the composite participant’s neighbours to con-
tinue the discovering process. In this composition protocol,
any remaining request is enclosed in a discovery message
that is forwarded between composite participants.
1939-1374 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSC.2016.2533348, IEEE
Transactions on Services Computing
1939-1374 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSC.2016.2533348, IEEE
Transactions on Services Computing
5
Fig. 5. Dynamic composition overlay
Definition 1. A discovery message including a request’s re-
maining part R
0
, is represented as DscvMsg = hR
0
, cache, hi,
where cache stores the progress of resolving split-join controls for
parallel service flows (see Section 3.1.2), and h is a criterion value
for request forwarding and service allocation (see Section 3.2 and
3.3).
The global service discovery process establishes a tempo-
rary overlay network called a dynamic composition overlay
(Fig.5), which contains all the reached usable composite
participants. Such a network only lasts for the duration of a
composition. It is managed by a set of execution guideposts,
which are control elements for the discovered service flow.
As shown in Fig.5, each guidepost is maintained by a
composite participant, linking the corresponding service to
who sent the discovery message. An execution guidepost is
adaptable and a byproduct of the composite participant’s
local discovery process.
Definition 2. An execution guidepost G = hR
id
, Di includes
a set of execution directions D and the id of its corresponding
composite request. For each d
j
∈ D, d
j
= hd
id
j
, S
post
, ω, Qi,
where d
id
j
is a unique id for d
j
, and the set S
post
stores the
participant’s post-condition services that can be chosen for next-
hop execution. The set ω represents possible waypoints on the
direction to indicate execution branches’ join-nodes when the
participant is engaged in parallel data flows. The set Q reflects
the execution path’s reliability of this direction, e.g., the estimated
execution path strength and the execution time (Section 3.2).
The local discovering process for a composite participant
is described in Algorithm 1. Note that the state transiting
conditions like ¬end, cost and usable are not shown in this
algorithm since they have been illustrated in Fig.4. This
process reacts when a composite request or a discovery
message is received, and generates an execution guidepost
as well as new discovery messages that enclose the part
of the composite request which cannot be solved on this
composite participant.
Data : Sender Y sends DscvM sg. Receiver X hosts
S = hS
f
, IN, OU T, QoS
time
i.
Result: an execution guidepost, A DscvMsg containing
the remaining request R
1 /* Listening */ ;
2 while receive message DscvMsg do
3 /* Configuring */;
4 GoalMatch(S, R);
5 /* Planning (when usable)*/;
6 if @DscvMsg
log
then
7 New D;
8 DscvMsg
log
← DscvMsg;
9 if partUsable then
10 Event← addJoin
11 else
12 Event← add
13 end
14 else
15 if DscvMsg
log
and DscvMsg have matched
cache value then
16 Update cache; Event← addSplit
17 end
18 if Progress(DscvMsg
log
) < Progress(DscvMsg)
then
19 Event← adapt;
20 end
21 if Progress(DscvMsg
log
) ==
Progress(DscvM sg) then
22 Event← add ;
23 end
24 end
25 switch Event do
26 case addSplit:
27 foreach d
i
∈ D do S
post
← S
post
+ Y
28 endsw
29 case adapt:
30 Clean D; d
y
← hR
id
, Y i;
31 endsw
32 case add||addJoin: d
y
← hR
id
, Y i;
33 endsw
34 //When a branch’s resolving is finished
35 Initiate cpltMsg
0
= hR
0
, cache, hi;
36 Send cpltcvMsg
0
;
37 /* Handover */;
38 if (Event!=add)&&(RemainReq) then
39 Initiate DscvMsg
0
= hR
0
, cache
0
, h
0
i,
R
0
= hI
0
, O
0
, F
0
, C
0
i;
40 O
0
← IN;
41 F
0
← F − S
f
matched
;
42 C
0
← C − QoS
time
;
43 Calculate h;
44 if GoalMatch(S, R) = partial then
45 Initiate cache
i
= hS
id
, m, ci;
46 S
id
← P
rec
, m ← OUT ∩ O;
47 c ← h num(m)/ num(O)i;
48 cache
0
← cache + cache
i
;
49 end
50 Send DscvMsg
0
;
51 end
52 end
Algorithm 1: Local Service Discovering Algorithm
