Automatic Definition of MOISE Organizations for Adaptive Workflows
Massimo Cossentino
a
, Salvatore Lopes
b
and Luca Sabatucci
c
National Research Council of Italy (CNR), ICAR, Palermo, Italy
t
Keywords:
Multi-Agent Systems, Business Process, BPMN Adaptation.
Abstract:
The enactment of dynamic workflows may take advantage of the multi-agent system paradigm. The approach
presented in this paper allows exploiting a high-level BPMN process definition to generate an agent organisa-
tion that can enact the workflow using different strategies. These are implemented as organisational schemes
representing alternative goal decomposition trees. The availability of several equivalent solutions enables the
optimisation and adaptation features of the approach. The mapping of the initial workflow to organisations
starts with the automatic generation of goals from the BPMN, and it exploits a metamodeling approach to
generate MOISE organisation definition.
1 INTRODUCTION
In the last two decades, web services enabled busi-
ness process running over the Internet (Sawhney and
Zabin, 2002). Business enterprises yield to redesign
their information and process management systems to
implement advanced features such as adaptation abil-
ity (Laukkanen and Helin, 2004; Gottschalk et al.,
2008).
However, the objective to increase agility and flex-
ibility of business processes conflicts with the current
trend of over-specifying workflow details. Moreover,
the activity-diagram style makes hard to re-arrange
the order of service.
A promising direction is to use multi-agent sys-
tems. This is an alliance, very frequent in litera-
ture (Buhler and Vidal, 2005; Singh and Huhns, 1999;
Ceri et al., 1997), in which the enactment of work-
flows takes great benefits from distinctive agent fea-
tures like distribution, adaptation and smartness.
It remains the problem of coordinating heteroge-
neous, autonomous agents, whose internal designs
could not be fully known a-priori.
Based on some recent results (Sabatucci and
Cossentino, 2019), this paper presents a novel ap-
proach for automatically generating an agent-based
adaptive workflow that can enact high-level business
process. This paper is an extension of a previous work
that has been presented in (Cossentino et al., 2020a).
a
https://orcid.org/0000-0003-1258-9767
b
https://orcid.org/0000-0003-1488-9755
c
https://orcid.org/0000-0003-2852-9355
The proposed approach is based on three pillars:
1. The use of BPMN (Chinosi and Trombetta, 2012)
for specifying the workflow; BPMN is a high-
level language that focuses on analysis activity,
differently from other execution languages such
as BPEL4WS (Andrews et al., 2003) (Business
Process Execution Language for Web Services)
that proposes a design-time bounding between
tasks and services.
2. Goal-Driven enactment. A BPMN may be au-
tomatically translated into goals (Sabatucci and
Cossentino, 2019). The advantage is that goals
allow for breaking the rules: whereas sequence
flows specify the precise order in which services
are invoked, goals can relax strict constraints,
widening the space for adaptation (Sabatucci and
Cossentino, 2017).
3. Workflow is a cooperative problem-solving ap-
plication. Multi-agent systems and services are
a synergic alliance for achieving the generated
goals.
The novelty of the paper is to link the goal generation
of (Sabatucci and Cossentino, 2019) to an automatic
definition of social organisations for agents. We show
that, given a set of goals and a set of available ser-
vices, it is possible to generate a social organisation
for enacting the workflow. The organization is com-
posed of several alternate solutions (goal decomposi-
tion trees) in form of schemes.
To support this claim, we selected MOISE (Han-
noun et al., 2000), an organisational model for agents
Cossentino, M., Lopes, S. and Sabatucci, L.
Automatic Definition of MOISE Organizations for Adaptive Workflows.
DOI: 10.5220/0010319201250136
In Proceedings of the 13th International Conference on Agents and Artificial Intelligence (ICAART 2021) - Volume 1, pages 125-136
ISBN: 978-989-758-484-8
Copyright
c
2021 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
125
for demonstrating the automatic generation.
Therefore, we map concepts from (Sabatucci
and Cossentino, 2019) and from (Sabatucci and
Cossentino, 2017) into the MOISE metamodel.
An automated tool derived from the approach, that
has been discussed in (Cossentino et al., 2020b).
The paper is structured as follows: Section 2
presents the problem and an overview of the ap-
proach. Section 3 summarizes the theoretical back-
ground at the basis of this work. The main contribu-
tion is discussed in Section 4, where Section 5 pro-
vides a justification scenario.Finally, conclusions and
future works are sketched in Section 6.
2 PROPOSED APPROACH
This section provides motivations for converting
BPMN into MOISE organizations. Furthermore, it
collocates the contribution of the paper into an ex-
tended framework for enacting dynamic workflows.
Despite discussing the whole framework is out of the
scope of this paper, we provide an overview for the
reader to get the idea behind goal-driven agent-based
adaptive workflows.
A BPMN diagram defines a set of activities, doc-
uments and messages and, above all, a set of proce-
dural rules that specify the order in which activities
must be executed to achieve a specific business goal.
Since the beginning, intelligent agents have been of-
ten associated to workflow management to simplify
service provisioning, execution and exception (Judge
et al., 1998; Singh and Huhns, 1999). An agent is suit-
able for these operations because it is an autonomous,
co-operative and intelligent entity able to collaborate
with other agents to process the tasks (Wooldridge,
1999; Ferber and Gutknecht, 1998). One of the moti-
vations because agents have been used in workflows is
their stateful and non deterministic nature. Whereas a
service is similar to a stateless method invocation, an
agent observes its environment, maintains an internal
state, learns from the experience if necessary, and its
response may change along time (Buhler and Vidal,
2005).
Moreover, agents promise to solve one of the
most interesting open challenge in this research ar-
eas: managing changes (van der Aalst and Jablonski,
2000) when changes of operating conditions cannot
be avoided and cannot be anticipated.
The use of agents carries out the advantage of pro-
viding workflow management systems with the ability
to face dynamic changes in resource levels and task
availability, as well as redistributing workload when
required (Judge et al., 1998). So far, agent organiza-
tions play a central role in workflow management and
one of the open challenge (Maia and Sichman, 2020)
is to identify interaction mechanisms enabling work-
flow adaptation.
The approach we propose uses agents, organi-
zations and goals. Agents encapsulate workflow
tasks, providing the advantage of autonomous deci-
sion making, easiness in distribution, and, optionally,
the ability to maintain a state and learn from pre-
vious cases. Organizations represent the interaction
mechanism for composing the workflow as sum of
services with the additional advantage of acting un-
der a customizable normative background. Goals are
the instrument for binding organizational rules (for
formation and adaptation) to the initial BPMN def-
inition. These goals are automatically derived from
the BPMN diagram and capture the nature of depen-
dencies between states of the system. Providing the
workflow as goals (rather than BPMN diagrams) al-
lows relaxing some of the strong relationships im-
posed by sequence-flows, supporting the workflow
management engine with a higher degree of freedom
in defining (at run-time) the flow of activities.
2.1 A Motivating Example
This paper uses the email-voting business process,
available in (Object Management Group (OMG),
2010). It is a high-level description of the process
for mediating and coordinating voting members in re-
solving issues. For reasons of readability, a compact
version of the workflow is shown in Figure 1: a man-
ager prepares the issue list for a discussion; all partic-
ipants propose solutions via email. After one week,
the discussion is closed with a voting session and the
manager communicates the results. If the issue has
not been solved, then a new discussion starts.
Let us suppose that, given high-level specification
of this process, developers create a set of web services
for enacting the workflow.
According to the available services, there could
be several workflows because: 1) they use different
services, 2) their control flow structure is different to
overcome different set of input/output parameters.
In any case, the workflow will present sev-
eral potential points of failure. A web-service
could be temporarily unavailable, or it may produce
wrong/incomplete results. Moreover, given the pres-
ence of humans in the loop, the process may fail be-
cause a human’s delay/failure. To cope with situations
like these, the most promising approach is changing
behaviour at run-time.
The use of a dynamic workflow management sys-
tem allows run-time modification of the original busi-
ICAART 2021 - 13th International Conference on Agents and Artificial Intelligence
126
Determine
Issues
Announce
Issues
for Vote
Announce
Results
vote
announcement
vote
results
issue list[in vote]
issue votes
not majority(issue votes)
Evaluate
Discussion
Progress
Check
Calendar
for Conference
issue list
[initial]
Announce
Issues
for Discussion
issue list
[in discussion]
Moderate E-Mail
Discussion
Moderate
Conference
scheduled
(conference)
not scheduled(conference)
issue list
[ready]
Announce
Discussion
Deadline
deadline
warning
Collect
Votes
vote
issue votes
[final]
issue
announcement
issue list
[in discussion]
+ +
x
x
x
x
issue
voting list
conference
time
6 days
7 days
Activity
Artifact
Message
Timer
x
+
Exclusive Gateway
Parallel Gateway
user task decorator
KEY
Figure 1: A compact version of the voting-by-email process, ri-elaborated from (Object Management Group (OMG), 2010).
ness process. However, despite the vast literature in
dynamic worklfow, modifying at run-time the work-
flow structure still poses many challenges.
In our vision, if an agent organization encapsu-
lates the workflow, then the use of agents in this con-
text could provide significant benefits:
1. agent organizations are flexible structures: agents
can enter/exit the organization at run-time; this
allows repairing the portion of a malfunctioning
workflow (local failure) by merely replacing the
agent who is responsible of a failing service (or a
block of services);
2. more complex changing strategies may be applied
by exploiting agent interaction mechanisms: a
distributed decision making could apply structural
changes, i.e. modifying the order of activities for
overcoming global problems;
3. agents may exploit learning activities for antic-
ipating the need of adaptation, for example by
associating an environmental condition with the
probability that a failure happen.
This paper focuses on presenting the approach for au-
tomatically generating various organizational config-
urations for handling the same BPMN. Many aspects
of self-adaptation are out the scope of this paper: 1)
the algorithm for automatically extracting goals from
BPMN is summarized in Section 3, but more de-
tails are in (Sabatucci and Cossentino, 2019); 2) the
procedures for binding available services to the ex-
tracted goals are in (Sabatucci and Cossentino, 2015;
Sabatucci et al., 2017).
2.2 Overview of the Proposed Approach
This work relies on the use of BPMN (or similar no-
tations) for specifying the standard flow of tasks, at a
high level of abstraction. It is suggested to model di-
agram as an information-centric workflow (Kumaran
et al., 2008) i.e. a process with specific attention to the
life cycles of information entities (documents, busi-
ness objects, and artifacts). The workflow is abstract
because tasks are not directly bounded to service.
This condition is necessary for enabling the algo-
rithm to properly generate significant milestones and
constraints to rule process execution. In general, this
formal way to model workflows, including ontologies
for specifying input and output data objects, is of-
ten encountered in literature (Francescomarino et al.,
2011).
The overall framework is shown in Figure 2. Af-
ter the workflow definition, an offline pre-processing
phase precedes a run-time phase of self-adaptive
workflow enactment. Two specific tools support the
offline phase:
• the BPMN2Goal is a tool for automatically gener-
ating goals from BPMN workflow (Sabatucci and
Cossentino, 2019);
• the Proactive Means-End Reasoning
(PMR) (Sabatucci and Cossentino, 2015) is
Automatic Definition of MOISE Organizations for Adaptive Workflows
127
Execution
Monitoring
Commitment
Organization
Selection
Automatic
Goal
Extraction
Abstract
Workflow
Definition
BPMN
Automatic
Organization
Definition
Goal
MOISE
Organization(s)
self-adaptive workflow
pre-processing
Service
Repository
Agent
Deployment
Figure 2: Overview of the proposed approach.
a service composition planner able of composing
workflows by combining non-deterministic
actions (i.e. actions with many possible results)
to address linear-temporal logic goals (Sabatucci
et al., 2017).
Here, the novel contribution is the Automatic Orga-
nization Definition, which elaborates goals and avail-
able services for producing one or more executable
MOISE specification(s).
The self-adaptive workflow compartment of Fig-
ure 2 is implemented as a multi-agent system respon-
sible for forming the organisation (by making each
agent to play a role in it).
When an agent is not able to execute its service, or
when a goal is violated, the current organization dis-
miss and another organization schema is selected, if
available. In practice, self-adaptation ability is based
on the online switching between an agent organisa-
tion configuration to another one. The Organization
Selection phase acts as an organization manager: it
evaluates all the available organisational options (pre-
viously generated and sorted according to some met-
ric), and it is responsible for selecting the alternative
one.
3 PRELIMINARY CONCEPTS
The first part of this section briefly introduces the
main concepts used in this paper: BPMN, implicit
goals and states of the workflow. The second part of
this section provides an intuition of the BPMN2Goal
procedure.
3.1 Background
The Business Process Model and Notation
(BPMN) (Chinosi and Trombetta, 2012; Object
Management Group (OMG), 2010) is a very ex-
pressive graphical notation for representing business
processes of diverse natures. A graphical notation is
the de-facto standard choice to express a representa-
tion of a process. The BPMN notation is very rich
and allows several modelling perspectives (Chinosi
and Trombetta, 2012; Object Management Group
(OMG), 2010). This paper focuses on collaboration
diagrams (similar to activity diagrams), in which a
process is described as a collection of participants
(each in a Swimlane) exchanging messages via
message flows. Therefore these diagrams are com-
posed of five categories of objects: activities, events,
messages, data objects, and many kinds of gateways.
Figure 1 reports the perspective of the Issue Manager
role in the voting-by-email process.
Typically a designer encodes individual business
goals to be accomplished in the model, after that the
system is responsible to achieve these goals. The re-
lationships between BPMN and goals is greatly stud-
ied (De la Vara Gonz
´
alez and Diaz, 2007; De la
Vara Gonz
´
alez and Diaz, 2007; Adamo et al., 2018).
Indeed, the OMG also offers alternative approaches
for representing a BPMN in terms of goals, rather
than activities. An example is the Case Management
Model and Notation (CMMN) (Marin et al., 2012)in
which designers describe what is allowed and disal-
lowed in the process rather than how to actualize it.
In (Adamo et al., 2018), authors present an investiga-
tion and a categorisation of the notion of goal in the
context of business process.
This work focuses on Implicit Goals, declarative
entities embedded in the structure of a workflow that
define functional aspects of the process. They are
operative, i.e. they aim at explaining why a work-
flow can evolve in a given way. Given this nature,
implicit goals are necessarily formally expressed ac-
cording to a precise formalism. Consequently, the
definition of implicit goal, inspired by Zambonelli et
al. (Abeywickrama et al., 2012) and aligned with most
of the literature about goal models (Yu, 2011; Bres-
ciani et al., 2004; Morandini et al., 2008) is:
Definition 3.1 (Implicit Goal). An implicit goal is a
pair: htc, f si where tc and fs are state-conditions. Re-
spectively, the trigger condition (tc) describes when
the goal becomes active, and it may be pursued; the
final state (fs) represents the desired state to be ad-
dressed.
We describe the workflow by observing its evolu-
ICAART 2021 - 13th International Conference on Agents and Artificial Intelligence
128
tion from the initial to the final state. The logic of
possible worlds allows studying state dynamics along
with a temporal line as a finite sequence of states. In
order to support the analysis of States of a Work-
flow, in (Sabatucci and Cossentino, 2019) we intro-
duced a specific semantic (based on predicate log-
ics) for describing relevant states. For instance, the
availability of a given artifact is expressed through a
predicate like available(hDatai), whereas the predi-
cate hstatei(hDatai) specifies the artifact assumes a
given state. Similarly, received, sent, caught, thrown,
at, a f ter, done, error are examples of predicates for
other relevant states of the system.
For instance, a task with issue list[initial] as in-
put and issue list[in discussion] as output and an out-
going issue management message is described as fol-
lows:
data in(e) = initial(issue list)
data out(e) = in discussion(issue votes)
mess in(e) = sent(issue announc, member)
3.2 The BPMN2Goal Procedure
The idea underlying converting BPMN into goals is:
if we were able to decompose the overall state evolu-
tion into a finite set of evolution steps, then each of
these is one of the implicit goals we are looking for.
At a first analysis, activities, events, and gateways
contribute in different ways to the state evolution of
a workflow. In particular activities are directly re-
sponsible for state changes, event elements capture
either external events or produce internal signals, and
gateway combines states according different strate-
gies (exclusive choice, inclusive choice, parallel, and
so on).
The idea, expressed in details in (Sabatucci and
Cossentino, 2019), is that each BPMN element is
ruled by internal conditions, i.e. conditions that holds
when analyzing the element as isolated by the rest of
of workflow (internal factors). However, when an el-
ement is connected to other elements, it also under-
goes to external conditions that are generated by pre-
decessors and successors. We must solve a ‘balance
of states’: the internal factors must balance the exter-
nal factors. By opportunely combining these forces,
it is possible to estimate the state transition, and for-
mulating the corresponding goal.
We report an example of the analysis of
states for the Announce Issues for Discussion.
Figure 3 shows relationships with predecessors
and successors of the Announce Issues for Dis-
cussion. Two predecessors tasks are joined
through some exclusive gateways. Briefly, we
can assert that Announce Issues for Discussion
Determine
Issues
Announce
Results
vote
results
not majority
(issue votes)
issue list
[initial]
Announce
Issues
for Discussion
issue votes
[final]
x
x
Discussion
issue list
[in discussion]
Figure 3: Predecessors and successors of the Announce Is-
sues for Discussion task.
is ready for execution when the following condi-
tion holds: initial(issue list) ⊕ (sent(vote results) ∧
f inal(issue votes) ∧ ¬ma jority(issue votes)) where
the ⊕ derives from the exclusive gateways.
After its execution, to allow the work-
flow proceeds, the state must include
in discussion(issue list). Combining internal
and external factors, we conclude the implicit goal is:
GOAL: Announce Issues for Discussion
WHEN:
(((final(issue_votes) and sent(vote_results))
and not majority(issue_votes))
xor initial(issue_list))
THE SYSTEM SHALL:
in_discussion(issue_list)
and sent(issue_announcement)
The BPMN2Goals tool is currently available as a web
service
1
; it accepts BPMN files in the XMI format,
according to the OMG specification. In the next sec-
tion, two metamodels will be introduced to discuss
how BPMN goals are mapped to MOISE organisa-
tions.
4 MAPPING GOALS TO
ORGANIZATIONS
For realizing the agent organizations, we selected
Jason (Bordini et al., 2007) for implementing BDI
agents (Rao et al., 1995) and MOISE (Hannoun et al.,
2000) for defining functional/structural/deontic as-
pects of the organization. The use of BDI agents
amplifies the agent ability of taking decisions also
considering environmental conditions. MOISE is
the natural choice when working with Jason: it is
supported by a well defined meta-model and by a
strong theory behind. Agents encapsulate services
1
The web service is available, with a front-page, at http:
//aose.pa.icar.cnr.it:8080/BPMN2Goal/
Automatic Definition of MOISE Organizations for Adaptive Workflows
129
and make them available to the organization by reg-
istering them into a shared Yellow Page. A workflow
is ready to be enacted when, selected the MOISE or-
ganization schema, interested agents commit to ev-
ery missions. The architecture also comprises an
internal-monitoring agent that estimates goal viola-
tions (Cossentino et al., 2018), thus raising adaptation
signals when necessary.
In order to generate organizations for the adap-
tative execution of worklows, we need to map some
concepts of the metamodel underlying the approach
as proposed in (Sabatucci and Cossentino, 2019) to
the corresponding MOISE concepts. The metamodel
implicitly adopted in (Sabatucci and Cossentino,
2019) is represented in Fig. 4, the composing ele-
ments are:
• BPMN Workflow: this is the BPMN representa-
tion of the input workflow. Generally speaking,
this is the result of the work of a business ana-
lyst who draws a solution to the problem. This
solution is composed by activities and flow re-
lationships (relationships are used for organizing
the activities control flows). In this work we are
now omitting some elements of a classical BPMN
workflow (sequence flow relationships, data ob-
jects, events, messages, . . . ).
• BPMN Activity: an activity can be a task or a sub-
process. A task is an atomic activity (a portion of
work to be done). A sub-process is a compound
defined by a sub-flow of activities. We are cur-
rently omitting sub-processes in the proposed ap-
proach, although their introduction would not af-
fect that.
• Implicit Goal: a goal is a pair <tc,fs>where tc
and fs are logical conditions combined by classi-
cal logic operators (AND, OR, NOT). The trigger
condition (tc) describes when the goal may be pur-
sued and the final state (fs) describes the desired
state of the world. Roughly speaking, we could
say that goals are extracted from activities, and
therefore a ’Goal Extraction’ procedural relation-
ship is drawn from the BPMN Activity element to
the Goal element.
• Service: it can be a cloud-, web-service or any
other abstraction of an action that can be done in
the agent’s world.
• Capability: it represents the concretization by an
agent of one of the n possible strategies for ful-
filling a goal. Frequently used as a wrapper to
services.
• Concrete Plan: it is a flow of capabilities that
can pursue a goal extracted from BPMN with the
approach proposed in (Sabatucci and Cossentino,
1..n
BPMN
Workflow
BPMN
Activity
Implicit
Goal
Goal Extraction
Capability
Concrete
Plan
1..n
1
Concrete WF
1..n 1..m
Means-End
Analysis
Capability
Relationship
1..n
Agent
registers capability
executes
Yellow Pages
Service
Wraps
1 1
1..n
Figure 4: The metamodel implicitly adopted in extracting
goal-oriented solutions from BPMN workflows.
2019). The plan may be produced by the Proactive
Means End Reasoning algorithm (Sabatucci and
Cossentino, 2015) that composes a set of capabil-
ities for satisfying the goal. A plan is made of a
sequence of capabilities and relationships among
them. Given a goal, many different plans may be
found to satisfy that, if enough capabilities are
available in the Yellow Pages repository. Each
plan will achieve different performances because
of the various behaviours of employed capabili-
ties.
• Yellow Pages: a directory of the capabilities that
can be provided by the agents populating the sys-
tem. It is used to compose plans.
It is significant to note that we can identify three
main zones in this metamodel: first, the input part
(on the left in Fig. 4). It consists of BPMN ele-
ments (BPMN workflow activity and others omitted
for conciseness in this paper). This part of the model
represents the workflow processed to obtain the cen-
tral part of Fig. 4: implicit goal, the key abstraction of
the approach. Implicit goals are obtained by process-
ing BPMN elements, mainly activities, as reported in
Sect. 3. Finally, the rightmost part of Fig. 4 describes
the solutions computed to pursue goals extracted from
BPMN. Each solution consists of a Concrete Work-
flow that is in turn, composed by Plans. Each Plan
pursues one of the goals. Considering that the same
goal may be fulfilled by composing a different set of
capabilities (and therefore by several plans), it is pos-
sible to create several Concrete Workflows that can
pursue the set of (implicit) Goals extracted from the
same BPMN workflow.
As discussed in sect. 2, a suitable agent organiza-
tion is to be defined to enact the concrete workflows
deduced from implicit goals extracted from the input
BPMN workflow and, for discussing that in details,
we will now refer to the metamodel of the MOISE
organizational framework reported in Fig. 5.
MOISE models organizations from three differ-
ent perspectives: structural, functional, and norma-
tive (Hannoun et al., 2000). The metamodel proposed
in Fig. 5 only reports MOISE elements that are sig-
nificant for the proposed approach. Indeed it also
ICAART 2021 - 13th International Conference on Agents and Artificial Intelligence
130
includes two elements (Collective Goal, Individual
Goal) we introduced to better explicate the proposed
approach; such elements are both coded as MOISE
goals.
Below a description of the MOISE metamodel el-
ements based on definitions proposed by Hubner in
(Hannoun et al., 2000):
• Organization: A MOISE organization is a special-
ized Group that is devoted to pursue some goal.
An organization is a collection of roles. Indeed,
defining an organization implies the definition of
its structural, functional and normative perspec-
tives.
• Group: a group is composed by roles. The num-
ber of agents that can play a role is constrained by
a minimum and maximum number. Links among
roles in the group specify the type of relationship
(Link element in the metamodel) between agents
playing two roles. A set of compatibilities be-
tween roles may be specified as well.
• Goal: MOISE goals may belong to two different
types: achieve and maintain. A plan is specified
to pursue the goal and it may describe the de-
composition of the goal into sub-goals by means
of Plan operators (sequence, parallel, choice).
Goal’s specifications also include the time to fulfil
(ttf) and the cardinality (how many agents have to
achieve the goal to consider that as globally sat-
isfied). For facilitating our mapping intents, we
decided to specialize the goal concept into Col-
lective Goal and Individual Goal. Indeed this is
inspired by the MOISE framework itself that ad-
dresses the definition of a Scheme as a specifica-
tion at the collective level and the Mission at in-
dividual level (see MOISE tutorial (Hubner et al.,
2010)).
• Collective Goal: It is a goal at the intermediate or
root level in the goal decomposition tree. It is used
as a label for intermediate parts of the scheme in
the organization functional description. Not part
of MOISE specifications.
• Individual Goal: It is a leaf goal of the goal de-
composition tree. It is the objective of a MOISE
mission and therefore some agent will pursue that
by using some of its capabilities. Not part of
MOISE specifications.
• Plan: a plan is composed of (sub-)goals related by
operators (sequence, choice, parallel). It is used to
create a sort of goal tree for decomposing the root
goal (that is a kind of collective goal) into finer
grained goals.
• Scheme: a Scheme is a global goal decomposition
tree, it is the collective specification of the goals
Organization Goal
pursues
Collective
Goal
Individual
Goal
1..n
Plan
1..n
Group Scheme Mission
Role
is committed to
Norms
Links
Compatibility
1..n
Figure 5: The portion of MOISE metamodel adopted in the
proposed approach for workflow enactment.
to be pursued by an organization. It also includes
Missions allowing to pursue the goals defined in
the scheme.
• Mission: ”is a set of goals for an agent commit-
ment in the context of a scheme execution” (Hub-
ner et al., 2010). It is the specification, at the in-
dividual level, of the functional perspective on or-
ganizations. It lists the goals to be pursued by the
agent(s) committing to the specific mission.
• Link: the type of relationship between roles. Al-
lowed types are: acquaintance, authority, commu-
nication.
• Compatibility: the specification of the compati-
bility in playing one role and another by the same
agent.
• Norms: there are two types of norms: obligation
and permission. While the first obliges the agent
playing a role to perform some mission, the latter
allows for that but it is not compulsory.
In order to define the organization that could enact the
concrete workflows while ensuring some degree of
adaptation as discussed in sect. 2, we need to instan-
tiate the elements of the MOISE metamodel reported
in Fig. 5 starting from an instance of the metamodel
reported in Fig. 4. The process for generating the
MOISE organisation will be introduced in the follow-
ing by referring to two main phases: the first consists
in the definition of a set of solutions by employing the
PMR algorithm as detailed before, the second phase
consists in the instantiation of the MOISE metamodel
elements.
4.1 BPMN Solutions Definition with
PMR Algorithm
Let us suppose the BPMN process reported in Fig. 6
is the input and the extracted implicit goals are G1,
G2, G3. They are all to be achieved in order to reach
the objectives of the BPMN process. Therefore we
could represent this situation with the implicit goal
tree of Fig. 6. For an easier tracing of the details about
goals extraction that are omitted in this paper, we are
Automatic Definition of MOISE Organizations for Adaptive Workflows
131
G
G1 G2 G3
Task1 Task2 Task3
BPMN
Automatic goals extraction from BPMN
Figure 6: An example of implicit goal tree extracted from a
simple BPMN process.
referring to the same example used in (Sabatucci and
Cossentino, 2019).
Now, let us suppose the solutions found using the
Proactive Means-end Reasoning (PMR) algorithm are
w1, w2, w3, w4, w5 as reported in the leftmost part
of Fig. 7. They are represented in the rightmost part
of the same figure using this notation: rectangles rep-
resent services that can be used to achieve the goals.
They are organized in plans using PMR solution con-
trol nodes like sequence (–>) parallel (||) or ExOR
(×). Agents exhibit the behaviour prescribed by these
services employing their capabilities (a capability of-
ten is purely a wrapping of a service). For instance,
solution w1 is the sequence of three services: s1, s2,
s3. Each of these services is connected to the goal it
fulfils by a dashed line ending with a not filled arrow.
The first level plan operator defining the outermost
configuration of the solution (a sequence or parallel
of other nodes) is the PMR solution plan (in black in
Fig. 7). This is identified with the solution itself and
therefore named after that. For the sake of simplicity,
only solutions w1 and w5 are reported in Fig. 7.
Solution w4 represents an interesting situation:
two services (s1”, s1”’) are to be executed in a se-
quence in order to fulfil goal G1, this is represented by
using a PMR Solution Control Node (sequence type)
and it is represented by introducing a blue hexagon
with the sequence symbol (–>). The correct order is
prescribed in numbers at the tail of the aggregation
relationship between the hexagon and the rectangles
representing the services.
Similarly a parallel is prescribed by solution w5
and this is represented by a plan icon with the par-
allel symbol ||. Solution w3 employs a service (s1’)
that can fulfil the two goals G1 and G2 at the same
time. In Fig. 7 this is represented by introducing a
grey coloured G1 2 goal. The same happens for the
already cited parallel of s4’, s4” that fulfils goals G2
and G3.
4.2 MOISE Organisation Definition
The definition of the MOISE organisation starts with
the identification of its goals. One goal is defined
per each service (see Fig. 8). Goal type is achieve-
ment and the pursued final state derives from the exit
condition of the corresponding service. The plan
for achieving each goal is straight since it just im-
plies executing the corresponding service. Goals are
composed into plans inside schemes according to the
PMR solution plan (see later on).
One scheme is defined per each PMR solution.
The plan for the scheme is represented by using a ma-
genta hexagon in Fig. 8, and it is composed of two
kind of elements: (MOISE) goals and (MOISE) plans
(both in magenta). The root level goal is assumed to
be the BPMN root level implicit goal (G in Fig. 8).
Goal decomposition is reconstructed by referring to
the PMR solution plan (black hexagons in Fig. 8)
that have a 1:1 correspondence with the MOISE plan
of each scheme (magenta hexagons). Goals in the
scheme are those related to the services employed in
the corresponding PMR solution.
Therefore, PMR solution w1 generates a MOISE
scheme (label ”Scheme sol1”) whose root goal is the
same as the original BPMN goal tree (G in Fig. 8).
The plan for achieving G is Plan1, and that is a se-
quence of goals g1, g2, g3. One mission is defined
per each goal in the plan. Fig. 9 reports the schemes
and groups defined from solutions w1 and w5.
One group is created per each scheme. Indeed we
are here talking of subgroups of a root group contain-
ing a general role labelled Worker. All roles in the
subgroups are extensions of the Worker role.
Inside each group, one role is defined per each ser-
vice listed in the corresponding PMR solution. Roles
are related by communication links when the plan in
the scheme is a sequence of goals. In this way the
first role may communicate to the second one when
its work is done, and so on. Roles to be executed in a
parallel are related by a communication link to roles
responsible for tasks following the parallel (i.e. tasks
positioned after the Join element in the BPMN) (not
represented in this example).
One mission is defined per each goal in the
scheme. It is worth saying that as prescribed by
MOISE syntax (see (Hubner et al., 2010), pag. 16)
plans inside the first level one are to be replace by
auxiliary goals but this is a syntactic expedient that
does not effect the proposed approach.
Roles are related to their corresponding mission
by means of an Obligation norm.
Summarizing, we define one goal per service, one
role per service (wrapped into a capability in agents),
one mission per goal and we oblige each role to com-
mit to the mission with the goal springing from its
service.
One group is defined per scheme, it collects the
roles involved in all the missions defined within the
scheme.
ICAART 2021 - 13th International Conference on Agents and Artificial Intelligence
132
G
s1
||
G1 G2 G3
s1’ s1’’ s1’’’ s2 s2’ s3 s4’ s4’’
G2_3
—>
1
2
G1_2
Proactive Means-End Reasoning
2
—>
1
2
w1
3
—>
1
w5
KEYS
G
Business Goal
—>
PMR Solution Plan
—>
PMR Decomposition
Node (sequence)
s1
Service
||
PMR Decomposition
Node (parallel)
Figure 7: The solutions found by the PMR algorithm for fulfilling the goal tree.
2
G
s1
||
G1 G2 G3
s1’ s1’’ s1’’’ s2 s2’ s3 s4’ s4’’
G2_3
—>
1
2
G1_2
—>
1
2
3
—>
1
||
2
g1’ g1 g1’’ g1’’’ g2 g2’ g3 g4’ g4’’
Plan1 Plan5
MOISE
goals
and plans
Proactive Means-End Reasoning
(PMR)
—>
1
2
w1
3
—>
1
w5
KEYS
G
Business Goal
—>
PMR Solution Plan
—>
PMR Decomposition
Node (sequence)
s1
Service
g1
MOISE Goal
—>
MOISE Plan
||
PMR Decomposition
Node (parallel)
Figure 8: Goals and Plans for the MOISE organisation.
Figure 9: MOISE organisation schemes and groups.
We are aware this process could be enhanced, for in-
stance, by clustering more goals in one mission like
the alternate solution proposed in Fig. 10. In this so-
lution, one mission is defined per each element in the
second level of the scheme (the first level is repre-
sented by the plan connected to G in Fig. 10). Indeed
such a solution would need some modification in the
MOISE framework. If more than one role is obliged
to commit to the mission, the framework requires that
each role must be able to perform all the goals listed
in the mission. This is against the way we construct
roles (one role per service from the PMR solution).
Building roles that are able to perform all the goals
in a mission raises the need for more complex agents
that have more capabilities (i.e. the ability to perform
all the required services).
In more straightforward approach we propose in
Fig. 9, each agent needs to have only one of the ca-
pabilities required for satisfying the mission’s goals.
This assumption is sufficient for it to participate in the
collaborative effort towards the solution.
Future extensions of this work will explore en-
hancement strategies, also including the adoption of
subgroups.
Fig. 11 reports the mapping between the elements
of the metamodels proposed in Fig. 4 and Fig. 5 and a
list of some elements introduced during the transfor-
mation process detailed in this section.
An automated tool for the generation of MOISE
organisations was presented in (Cossentino et al.,
2020b).
Automatic Definition of MOISE Organizations for Adaptive Workflows
133
Scheme_sol5
goal goal
goal
Scheme_sol1
goal goal
goal
g1 g2 g3
G
—>
1
2
3
Mission_s1
Max=1
Min=1
Mission_s2
Max=1
Min=1
Mission_s3
Max=1
Min=1
Mission_s1
Max=1
Min=1
—>
1
||
2
g1
g4’ g4’
G
Mission_s4’-s4’’
Max=1
Min=1
Group_sol1
Role_s1
Max=1
Min=1
Role_s2
Max=1
Min=1
Role_s3
Max=1
Min=1
Communication
Communication
Obligation
Obligation
Obligation
Group_sol4
Role_s1
Max=1
Min=1
Role_s4’
Max=1
Min=1
Role_s4’’
Max=1
Min=1
Communication
Communication
Obligation
Permission
Permission
Figure 10: MOISE organisation with several goals inside
one mission.
Figure 11: The proposed mapping among PMR solutions
and MOISE organisation metamodel elements.
5 ORGANIZATIONS AT WORK
In this section, we will provide a justification scenario
for the proposed approach.
Suppose a business analyst designs (using BPMN)
a workflow for dealing with some process. Despite
the great attention and effort adopted in the design
task, the analyst is well aware that a fixed process
cannot cope with runtime changes in the execution en-
vironment like: entering new agents or disappearing
existing agents, variations in the performance or cost
of the services offered by agents, exogenous factors
affecting the real world where the system operates.
For these reasons, the analyst prefers to adopt an
adaptive execution environment for her workflow. Us-
ing an online utility (see sect. 3.2) she gets a set of
goals from her BPMN file.The PMR algorithm com-
putes a set of alternative solutions for achieving these
goals. Each solution is composed of a set of capabil-
ities (corresponding to those available in the yellow
pages).
Behind each capability there is a service that can
be invoked by the agent or some action the agent can
perform. For each couple capability-agent, the yellow
pages also list some quality of service (QoS) value.
This latter is domain dependent and for each process
the designer should provide the system with a formula
for calculating the QoS of the overall solution.
Now, for the sake of conciseness, we will refer
to the same example reported in the previous section.
Therefore, let us suppose the solutions are w1...w5
and the previously described organization is adopted
to execute the solutions.
It is now relevant to remind that the organiza-
tion encompasses 5 schemes (one per each solution
w1. . . w5), one different group of roles is defined for
each scheme. The different set of roles in each group
implies that a different set of agents is necessary to
deploy the capabilities that can satisfy the goals in
the scheme. It may be useful to remark that some-
times one agent offers more than one capability and
therefore it may play several roles in the same group
or it may participate in different groups. We label
each set of agents that forms one group and can ex-
ecute the corresponding scheme as grounding config-
uration. Generically speaking for the 5 groups of the
example, we can suppose more than 5 grounding con-
figurations may exist (it depends on the number of
agents and the redundancy in their capabilities). For
each grounding configuration, the system calculates
the (expected) QoS, and the best one defines which
configuration will be instantiated.
Hence, the best grounding configuration from the non
functional point of view (expressed by a QoS) will be
deployed to execute the process.
Process execution may incur in unexpected fail-
ures, for instance, because a remote service is no more
reachable or the code of an agent is bugged. In that
case process adaptation is needed and our approach
offers an advantage. If the selected grounding con-
figuration has not achieved all the goals, others (with
a lower QoS) are promptly available to be deployed
and continue the work. The organization remains the
same, a new set of agents will commit to the missions
of the same scheme (alternative grounding configura-
tion for the same scheme) or to another one.
We are aware that adopting solutions computed
to execute the whole process may not be the optimal
choice when starting from a partially executed one. It
could be anyway a reasonable compromise with the
ICAART 2021 - 13th International Conference on Agents and Artificial Intelligence
134
time and resources needed to compute a new set of
solutions for the current state of the world, extracting
goals, defining the supporting organization and inject-
ing that in the running system (not trivial with the se-
lected JaCaMo framework). If no grounding config-
uration will be able to conclude the execution, a new
iteration will start from the calculation of a new set of
solutions with the PMR algorithm.
Indeed, there is also another optimisation possi-
bility we did not consider in this paper. The PMR al-
gorithm is usually perpetually run if new capabilities
are registered in the yellow pages or new agents enter
the system and could produce new solutions. The new
solutions could be better than the previous ones. Fur-
ther development of the proposed approach will con-
sider them, thus allowing their introduction during the
adaptation phase.
6 CONCLUSIONS AND FUTURE
WORKS
This paper presented an approach for a runtime goal-
driven adaptation of MOISE organizations in the exe-
cution of BPMN processes. The approach is based
on: 1) automatic generation of goals from BPMN
and 2) mapping goals and service-oriented solutions
into different schemes of the agent organization that
can be selected according to performance criteria or
to overcome a failure. Goals may relax BPMN con-
straints, and the proposed approach has the advantage
of automatically defining alternative organizational
solutions (several schemes inside one organization)
for pursuing the goals underlying the input workflow
(manually created by a business analyst). As a mat-
ter of fact, the availability of different organization’s
schemes allows selecting the scheme (goal decompo-
sition tree and set of missions) that provides the best
performance, according to the quality attributes reg-
istered in the yellow pages. It also allows switching
from the current scheme to another one, in case of
agent/service failures (runtime adaptation feature).
Although the proposed approach produces effec-
tive results, it could be improved in many ways. For
instance, agent roles may be optimized in number and
specialization. So far, different roles are defined for
different capabilities even when they have the same
pre- and post-conditions but different name. Future
works may propose some improvement on that.
As part of the future works (and limits of the cur-
rent work), we also would like to note that a few el-
ements of the BPMN metamodel have been omitted
in Fig. 4 for simplicity. While this is not relevant for
most of them, we consider the sub-process a signifi-
cant element that could lead to interesting extensions
to the proposed approach. In fact, dealing with that
as a kind of process in the process (as it is indeed),
the result could bring to the design of organizations
conceived to act within other higher-level (or lower-
level) ones in a kind of hierarchy that may resemble
a holarchy and some methodological issue may arise
from that (Cossentino et al., 2010).
REFERENCES
Abeywickrama, D. B., Bicocchi, N., and Zambonelli, F.
(2012). Sota: Towards a general model for self-
adaptive systems. In Enabling Technologies: In-
frastructure for Collaborative Enterprises (WETICE),
2012 IEEE 21st International Workshop on, pages 48–
53. IEEE.
Adamo, G., Borgo, S., Di Francescomarino, C., Ghidini,
C., and Guarino, N. (2018). On the notion of goal in
business process models. In International Conference
of the Italian Association for Artificial Intelligence,
pages 139–151. Springer.
Andrews, T., Curbera, F., Dholakia, H., Goland, Y., Klein,
J., Leymann, F., Liu, K., Roller, D., Smith, D., Thatte,
S., et al. (2003). Business process execution language
for web services.
Bordini, R. H., H
¨
ubner, J. F., and Wooldridge, M. (2007).
Programming multi-agent systems in AgentSpeak us-
ing Jason, volume 8. John Wiley & Sons.
Bresciani, P., Perini, A., Giorgini, P., Giunchiglia, F., and
Mylopoulos, J. (2004). Tropos: An agent-oriented
software development methodology. Autonomous
Agents and Multi-Agent Systems, 8(3):203–236.
Buhler, P. A. and Vidal, J. M. (2005). Towards adaptive
workflow enactment using multiagent systems. Infor-
mation technology and management, 6(1):61–87.
Ceri, S., Grefen, P., and Sanchez, G. (1997). Wide-a dis-
tributed architecture for workflow management. In
Research Issues in Data Engineering, 1997. Proceed-
ings. Seventh International Workshop on, pages 76–
79. IEEE.
Chinosi, M. and Trombetta, A. (2012). Bpmn: An intro-
duction to the standard. Computer Standards & Inter-
faces, 34(1):124–134.
Cossentino, M., Gaud, N., Hilaire, V., Galland, S.,
and Koukam, A. (2010). Aspecs: an agent-
oriented software process for engineering complex
systems. Autonomous Agents and Multi-Agent Sys-
tems, 20(2):260–304.
Cossentino, M., Lopes, S., and Sabatucci, L. (2020a). Goal-
driven adaptation of moise organizations for workflow
enactment. In Proc. of the 8th International Workshop
on Engineering Multi-Agent Systems (EMAS 2020),
National Research Council of Italy.
Cossentino, M., Lopes, S., and Sabatucci, L. (2020b). A
tool for the automatic generation of moise organisa-
tions from bpmn. In CEUR Workshop Proceedings of
Workshop “From Objects to Agents”, September 14–
16, 2020, Bologna, Italy.
Automatic Definition of MOISE Organizations for Adaptive Workflows
135
Cossentino, M., Sabatucci, L., and Lopes, S. (2018). Partial
and full goal satisfaction in the musa middleware. In
European Conference on Multi-Agent Systems, pages
15–29. Springer.
De la Vara Gonz
´
alez, J. L. and Diaz, J. S. (2007). Busi-
ness process-driven requirements engineering: a goal-
based approach. In Proceedings of the 8th Workshop
on Business Process Modeling. Citeseer.
Ferber, J. and Gutknecht, O. (1998). A meta-model for
the analysis and design of organizations in multi-agent
systems. In Proceedings International Conference on
Multi Agent Systems (Cat. No. 98EX160), pages 128–
135. IEEE.
Francescomarino, C. D., Ghidini, C., Rospocher, M., Ser-
afini, L., and Tonella, P. (2011). A framework for the
collaborative specification of semantically annotated
business processes. Journal of Software Maintenance
and Evolution: Research and Practice, 23(4):261–
295.
Gottschalk, F., Van Der Aalst, W. M., Jansen-Vullers,
M. H., and La Rosa, M. (2008). Configurable work-
flow models. International Journal of Cooperative In-
formation Systems, 17(02):177–221.
Hannoun, M., Boissier, O., Sichman, J. S., and Sayettat, C.
(2000). Moise: An organizational model for multi-
agent systems. In Advances in Artificial Intelligence,
pages 156–165. Springer.
Hubner, J. F., Sichman, J. S., and Boissier, O. (2010). Moise
Tutorial (for Moise 0.7). http://moise.sourceforge.net.
Judge, D., Odgers, B., Shepherdson, J., and Cui, Z. (1998).
Agent-enhanced workflow. BT Technology Journal,
16(3):79–85.
Kumaran, S., Liu, R., and Wu, F. Y. (2008). On the du-
ality of information-centric and activity-centric mod-
els of business processes. In International Conference
on Advanced Information Systems Engineering, pages
32–47. Springer.
Laukkanen, M. and Helin, H. (2004). Composing work-
flows of semantic web services. Extending Web Ser-
vices Technologies, pages 209–228.
Maia, A. V. and Sichman, J. S. (2020). Representing plan-
ning autonomy in agent organizational models. Theo-
retical Computer Science, 805:92–108.
Marin, M., Hull, R., and Vacul
´
ın, R. (2012). Data centric
bpm and the emerging case management standard: A
short survey. In International Conference on Business
Process Management, pages 24–30. Springer.
Morandini, M., Penserini, L., and Perini, A. (2008). To-
wards goal-oriented development of self-adaptive sys-
tems. Proceedings of the 2008 international work-
shop on Software engineering for adaptive and self-
managing systems, pages 9–16.
Object Management Group (OMG) (2010). Business Pro-
cess Model and Notation (BPMN 2.0) by Exam-
ple. Available online at https://www.omg.org/cgi-
bin/doc?dtc/10-06-02.pdf.
Rao, A. S., Georgeff, M. P., et al. (1995). Bdi agents: from
theory to practice. In ICMAS, volume 95, pages 312–
319.
Sabatucci, L. and Cossentino, M. (2015). From means-end
analysis to proactive means-end reasoning. In Pro-
ceedings of the 10th International Symposium on Soft-
ware Engineering for Adaptive and Self-Managing
Systems, pages 2–12. IEEE Press.
Sabatucci, L. and Cossentino, M. (2017). Self-adaptive
smart spaces by proactive means–end reasoning. Jour-
nal of Reliable Intelligent Environments, 3(3):159–
175.
Sabatucci, L. and Cossentino, M. (2019). Supporting dy-
namic workflows with automatic extraction of goals
from bpmn. ACM Transactions on Autonomous and
Adaptive Systems (TAAS), 14(2):1–38.
Sabatucci, L., Lopes, S., and Cossentino, M. (2017). Self-
configuring cloud application mashup with goals and
capabilities. Cluster Computing, pages 1–17.
Sawhney, M. and Zabin, J. (2002). The seven steps to nir-
vana: Strategic insights into ebusiness transforma-
tion. McGraw-Hill, Inc.
Singh, M. P. and Huhns, M. N. (1999). Multiagent systems
for workflow. Intelligent Systems in Accounting, Fi-
nance & Management, 8(2):105–117.
van der Aalst, W. M. and Jablonski, S. (2000). Dealing
with workflow change: identification of issues and so-
lutions. Computer systems science and engineering,
15(5):267–276.
Wooldridge, M. (1999). Intelligent agents. Multiagent sys-
tems, 6.
Yu, E. (2011). Modelling strategic relationships for process
reengineering. Social Modeling for Requirements En-
gineering, 11:2011.
ICAART 2021 - 13th International Conference on Agents and Artificial Intelligence
136