TOWARDS PROBLEM SOLVING METHODS
IN MULTI-AGENT SYSTEMS
Paul Bogg
1
, Ghassan Beydoun
2
and Graham Low
1
1
University of New South Wales, Australia
2
University of Wollongong, Australia
Keywords: Knowledge engineering, Multi-agent system methodologies, Problem-solving methods.
Abstract: Problem Solving Methods (PSM) are abstract structures that describe specific reasoning processes
employed to solve a set of similar problems and have proved very effective at enhancing reuse and
extensibility in developing knowledge-based systems. We envisage that off-the-shelf PSMs can similarly
assist in the development of agent-oriented solutions using Multi-Agent Systems (MAS). A challenge
towards the effective use of PSMs in MAS is that current approaches to formulating PSMs do not
adequately address the complexity of problems to which agent-oriented systems are suited. Towards
addressing this, this paper focuses on providing an approach to guide developers in adequately formulating
PSMs for complex problem-solving where interactions are involved, such as in domains where negotiation
and cooperation are essential for solving a problem.
1 INTRODUCTION
The demand for agent-oriented software motivated
the creation of new development approaches, such
as INGENIAS (Pavon et al., 2005), Tropos
(Bresciani et al., 2004) and MOBMAS (Tran &
Low, 2008). None adequately addressed
extensibility, interoperability and reuse other than
(Beydoun et al., 2007; Tran & Low, 2008) where it
was argued that an ontology-based approach is
needed for a truly domain-independent agent-
oriented development.
Following the reuse paradigm promoted in
knowledge-based systems development (Schreiber et
al., 2001), the work in (Beydoun et al., 2006)
proposes a process that revolves around a domain-
dependant ontology to build individual agents with
problem-solving methods (PSMs). PSMs are high-
level structures describing a reasoning process
employed to solve general types of problems (Fensel
et al., 2002). Continuing the work in (Beydoun, Tran
et al., 2006), we envisage that engineering problem-
solving knowledge as domain-independent
ontology-based PSM structures is beneficial towards
achieving domain-independent agent-oriented
methodologies and systems. A library of these
PSMs would assist the development of agent-
oriented systems in domains where existing
problem-solving knowledge may be reused. A set of
modular, reusable problem-solving components has
the potential to reduce development costs and speed
up the development process. More specifically, this
paper investigates the role that task and problem-
solving knowledge play, arguing that current
approaches to PSMs do not adequately address the
complexity of problems to which agent-oriented
systems are suited. In particular, where problem-
solving software components are dependant on
interactions appropriate PSMs that address
interaction functionality have not been fully
investigated. This paper proposes an extension to
PSMs with an additional interaction dependency
construct through which interaction specific
problem-solving knowledge can be used.
Interaction-specific PSMs describe knowledge about
interactions for problem-solving, and how to design
methods to resolve complex problems where
interactions are necessary. Negotiation is used as
brief example of how PSMs may be used to design
MASs for interaction dependent problem-solving.
2 RELATED WORK
By using a domain ontology and an appropriate
PSM, it was envisaged that knowledge based
308
Bogg P., Beydoun G. and Low G. (2009).
TOWARDS PROBLEM SOLVING METHODS IN MULTI-AGENT SYSTEMS .
In Proceedings of the 4th International Conference on Software and Data Technologies, pages 308-313
DOI: 10.5220/0002252603080313
Copyright
c
SciTePress
systems can be easily developed as new problems
are encountered (Figure 1) (Studer et al., 1998).
Figure 1: As new problems arise, the PSM and the
ontology banks are used to construct suitable KBSs. An
ontology from the ontology bank strengthens a given PSM
from the PSM bank to suit the domain.
Recently, approaches have begun to address the
elicitation of PSMs from common problem-solving
knowledge. CommonKADS (Schreiber, Akkermans
et al., 2001), is a prominent approach which
provides a Task model, which provides a
hierarchical description of tasks, and an Expertise
model, which provides the method for achieving a
task. CommonKADS provides reusable task-specific
PSMs for composing the Expertise model to solve a
variety of pre-determined types of tasks (such as
diagnosis). Knowledge engineering also leverages
ontologies for eliciting and developing domain-
independent and reusable PSMs. One early
approach by (Fensel et al., 1997) tackled reusability
by incorporating ontologies for domain, task, and
PSM-specific knowledge. Another approach,
OntoKADS, extends CommonKADS by way of
introducing ontologies to comprise the expertise
model (Bruaux et al., 2005).
UPML (Fensel, Motta et al., 2002) encapsulated
previous approaches to describing general task and
problem-solving knowledge with general ontology-
based PSM structures. One limitation in UPML is
the absence in consideration given to PSMs for tasks
where multiple software components are required to
interact in order to solve a problem. For instance, e-
commerce problem-solving agents negotiate for
trade. Interaction-dependent problem-solving (such
as negotiation) is prevalent in agent-oriented
systems. To leverage the benefits of PSMs in AOSE,
PSM structures addressing interaction-dependent
problem-solving need to be developed. Recent
approaches to incorporating PSMs into agent-
oriented architectures have not addressed this.
MAS-CommonKADS (Iglesias & Garijo, 2005)
advocates task and problem-solving knowledge use
in its methodology. However, it presumes the
existence of PSM libraries suitable for complex,
interaction-dependent problem-solving. The ORCAS
framework (Gómez & Plaza, 2007) introduces
methods to adapt PSMs to agent capabilities. Their
work addresses cooperation as “agent teams” at the
knowledge level. However, it doesn’t address
interaction dependent problem-solving knowledge
required for negotiation. This is the focus here.
Figure 2: Ontology-based MAS development using PSMs:
(1,2) Domain Ontology produces Goal Analysis is used to
select PSMs from a PSM bank. (3, 4) Knowledge analysis
delineates local agent knowledge. (5, 6).
In (Beydoun, Tran et al., 2006), software
engineering requirements to use PSMs were mapped
out resulting in a methodological model (Figure 2).
This work did not address the issue of how to best
formulate the PSMs for interaction-dependent
problem solving. This paper continues this work by
formulating an appropriate way to construct PSMs
for distributed multi-agent systems (MAS). Much
previous work has gone into integrating ontologies
and PSMs e.g. (Fensel, Motta et al., 1997). It is not
yet clear whether that work needs to be extended for
the integration of domain ontologies with PSMs for
MAS. Invetigating this is left as future work.
3 FORMULATING PSMS FOR
MAS
Three types of knowledge are consistently identified
in formulating a PSM structure (e.g. in (Decker et
al., 1999; Fensel, Motta et al., 2002)): domain
knowledge, task knowledge, and problem-solving
knowledge. In these terms, PSMs are structured
problem-solving knowledge suited to achieving
tasks/goals in particular domains. UPML (Fensel,
Motta et al., 2002) defines a PSM in terms of these
knowledge components. Complex distributed
problems to which MASs are suited to solve may
TOWARDS PROBLEM SOLVING METHODS IN MULTI-AGENT SYSTEMS
309
require interactions between agents to coordinate
solutions. Towards MAS-specific PSMs, this section
extends the UPML PSM definitions. It adds a new
construct notation, interaction dependencies, noting
that when multiple agents are required to solve a
particular problem then further analysis is required
to determine what type of interaction is necessary.
When problem-solving depends on interactions,
further consideration needs to be given towards
understanding how different PSM definitions are
related. This needs to be accounted for in order to
properly formulate PSMs for MAS. For instance
(Fig. 3), in designing two agents required to
coordinate building a house, PSMs for a carpentry
agent may depend on PSMs for a brick layer agent.
Where this type of relationship between PSM
definitions exists, we use the term PSM co-
dependency. Where co-dependencies exist between
PSM definitions for separate agents, we use the term
PSM interaction dependency to specifically mean
that agents may be required to interact with one
another in order to successfully solve problems.
Figure 3: Agent-level PSM composition.
Interaction dependent PSMs bring additional
dynamics to a MAS software development process.
Firstly, interaction dependent PSMs suggest the
presence of additional methods and/or agents during
an analysis phase. Secondly, interaction dependant
PSMs may assist in designing the interaction
structure between agents by suggesting what type of
exchange is required between agents. The type of
exchange required might be as simple as an enquiry,
or as sophisticated as negotiation. Thirdly, since
interaction dependant PSMs are ontology-based,
reuse (as suggested in (Breuker, 1999)) is a natural
feature for future MAS development.
From an individual agent-level perspective, for
distributed problems in which agents are required to
interact, an interaction dependent PSM may be
aimed at achieving agent-level goals. For instance, a
negotiating agent may have a ‘Buy Item PSM’ that
depends on negotiation to satisfy an agent-level goal
to acquire a good. However, a software engineer
may not only be interested in agent-level goals, but
may also be interested in system-level goals.
From a system-level perspective, another type of
relationship may exist between PSMs. As is
illustrated in interaction-dependent problem-solving
literature (such as in negotiation (Jennings et al.,
2001)), sometimes the software engineer is
interested in designing agents whose interactions
produce system-level properties. For instance,
optimal utilitarian agreements can be engineered by
designing negotiating agents to use a correct
combination of strategies under particular
circumstances (e.g. (Fatima et al., 2004)). A PSM
approach may be used to engineer systems where the
selection of Strategy PSM ‘A’ suggests that the
selection of another Strategy PSM ‘B’ facilitates
system-level properties (such as utility optimisation)
in addition to agent-level goals. PSMs with system-
level dependencies may be used to design agent
interactions such that system-level goals are
achieved without resorting to “agent teams” (e.g.
(Gómez & Plaza, 2007)) – coordination and
cooperation are achieved at the agent-level. (Müller,
2002) argues that this may produce agent-oriented
systems more widely applicable to general types of
problems.
Table 1: Examples of interaction dependent PSMs.
PSM PSM Interaction Dependency
Buy Item Sell Item Negotiation for trade
Compensate
for failure of
agent Y
Compensate
for failure of
agent Y
Coordination to continue
system operation during
component failure
Procure
Service for
consumer
Provision
Service to
consumers
Negotiate terms of service
agreement
Examples of interaction dependent PSMs are
provided in Table 1. The first example is PSMs for
commercial activities requiring interactions to
achieve individual agent goals. The type of
interaction required may be a simple retail exchange,
or be a complex multi-issue negotiation. A system-
level goal might be that all agent-level interactions
are optimal according to some criteria (e.g.
utilitarian optimal in (Fatima, Wooldridge et al.,
2004)). The second example may appear in MASs
where robustness is an important system-level
requirement. PSMs may be needed to design
coordinative actions assuring compensation during
component failure (for instance, sensor agents
compensate for the loss of other sensor agents in a
battlefield information system (Deloach et al.,
2008)). The third example may occur where agents
procure service level agreements (for instance, in
acquiring satellite and cable channels for television
viewing, such as in (Cattoni et al., 1999)).
These knowledge engineering-based guidelines
may be used in designing PSM repositories for
interaction dependent problem-solving knowledge
for use in AOSE. During AOSE analysis PSM
House Frame
PSM
Acquire Wood
Frames PSM
Build Brick W all
PSM
Acquire Bricks
PSM
Brick Layer Agent
Interaction
Dependency
Co-dependency
Co-dependency
ICSOFT 2009 - 4th International Conference on Software and Data Technologies
310
repositories may not contain all relevant PSMs, and
may need to be refined or developed – this process is
not described here, and left as future work.
4 INTERACTION-SPECIFIC
PSMS
Applying the insights of the previous section, this
section adds new constructs to UPML to
accommodate complex interactions used to
formulate our new type of PSMs, interaction-
specific PSM. This assists the designers of MASs by
providing a structure to interaction-specific problem-
solving knowledge. This new type is needed
wherever interaction dependent PSMs suggest the
exchange between two agents is sophisticated (such
as negotiation, coordination or cooperation).
Interaction-specific PSMs are intended to be
reusable. Knowledge about interaction-dependent
problem-solving is reusable in different domains,
and for different tasks e.g. similar methods for
negotiation in e-commerce trade might be adopted in
the negotiation of free trade agreements. We use
literature on designing agents for negotiation,
cooperation, and coordination to identify three types
of interaction-specific PSMs (Fig. 4):
Interaction Protocol PSM: defines the rules for
interaction engagement. An interaction protocol
defines an order to engagements between agents
using terms expressed by the communication
protocol.
Model PSMs: structured knowledge about how
to model information that an agent observes.
They directly relate to interactions because
agency requires autonomous assessment of
itself, external agents and the environment. An
interaction protocol may constrain the types of
information an agent may observe.
Strategy PSMs: structured knowledge about
how interactive behaviour is derived from
output from the Model PSMs and the
Interaction Protocol PSM.
The above three types are derived from
classifications of agent design components used for
interaction-dependent problem-solving (such as
described in (Sandholm, 1999; Jennings, Faratin et
al., 2001; Lomuscio et al., 2001)). For example,
(Sandholm, 1999) describes variations of interaction
protocols where particular strategies depend on
models of utility for cooperative distributed
problem-solving. (Lomuscio, Wooldridge et al.,
2001) describes interaction protocols and strategies
as the two basic types of components for agent-
based negotiation. (Jennings, Faratin et al., 2001)
describes areas of negotiation research concerned
with protocols, negotiation objects, and decision
making models.
Figure 4: Knowledge level PSM composition.
Interaction Protocol PSM definitions are refined
by domain and task knowledge to produce specific
Interaction Protocol mappings (Figure 4). Model
PSM definitions are refined by domain knowledge to
produce Model PSM mappings, whereby inputs to
these mappings are provided by the agent. Multiple
Model PSMs may be selected or refined, depending
on agent design. The output from the Interaction
Protocol and Model PSM mappings are then used to
select the Strategy PSM. The strategy PSM is then
refined by task knowledge to produce the Strategy
PSM mapping. The output of the Strategy PSM
mapping is then used by the software engineer to
design the agent’s next interactive move. By
distinguishing between Interaction Protocol PSMs,
Model PSMs, and Strategy PSMs, interaction
specific problem-solving knowledge may be reused
by software engineers to design agent-oriented
solutions to complex problems.
5 APPLICATION OF PSMS FOR
MAS DEVELOPMENT
This section describes an application of interaction
dependent PSMs and interaction-specific PSMs to
designing agents for negotiation. The methodology
follows from Section 2, Figure 2 (from (Beydoun,
Tran et al., 2006)), where ontology-based
development of MASs from PSMs was described.
The scenario is negotiation for e-commerce trade.
Example:
An agent oriented system is required
to automate negotiation in an electronic market
place for buying and selling fish. Autonomous, self-
interested agents act on behalf of people. Agents
TOWARDS PROBLEM SOLVING METHODS IN MULTI-AGENT SYSTEMS
311
determine when and how to negotiate in order to
satisfy the needs of people. Agents are responsible
for collecting relevant information, and negotiating
the best possible utility-based outcome given the
information context (Cuní et al., 2004).
Suppose a software engineer aims to design an
agent that buys fish. At the conclusion of a domain
ontology and goal analysis, the software engineer
establishes a set of goals and task requirements to be
satisfied by an agent. The engineer selects the task
“Buy fish” and identifies “Buy Item PSM” as an
appropriate possible solution. “Buy Item PSM” is
identified as having an interaction dependency with
another PSM, “Sell Item PSM”. Figure 5 illustrates
a PSM approach to designing the agent solution.
Figure 5: Agent-oriented modelling for fish market place
derived from PSMs with an interaction dependency.
The software engineer determines that
negotiation is the interaction type necessary for the
domain. Since negotiation is pervasive in many
domains, the engineer consults libraries for existing
negotiation-specific problem-solving knowledge.
Task and goal analysis revealed that agents are also
required to maximise a utility, where a utility is
defined by the domain ontology. Appropriate
interaction-specific PSMs need to be selected – a
type of Interaction Protocol, Model, and Strategy
PSM. The software engineer attempts to find
interaction-specific PSMs (within the repositories)
oriented towards utility modelling and strategy.
“Utility Modeling PSM” and “Maximise Utility
Strategy PSM” are identified. For defining the
interaction, a “Bargaining Protocol PSM” is suitable.
To complete the development, the software
engineer now needs to design the fish buying agent
for the market place. PSMs are refined by task and
domain knowledge, resulting in mappings that are
task and domain specific methods that can directly
be used to design agent plans. Firstly, the domain
ontology is used to refine the “Utility Model PSMs”
to produce a fish market mapping, and a personal
fish-value mapping (the inputs for these mappings
might come from the person for whom the agent is
acting). Secondly, refinement of the “Maximise
Utility Strategy PSM” is made towards a specific
communication protocol ontology, producing a fish-
buying strategy mapping. The inputs for the fish-
buying strategy mapping are the outputs from the
fish market mapping and personal fish-value
mapping. Thirdly, refinement of the “Bargaining
Protocol PSM” is made towards the domain
ontology to produce a bargaining protocol mapping
which restricts interactions defined in terms of the
communication protocol ontology. At the conclusion
of this design, the software engineer may decide to
engage in a similar process for designing the fish
selling agent, with a view to (possibly) re-using
PSMs and mappings identified for the buying fish
agent. In addition to defining methods for
interaction-dependent problem-solving, interaction-
specific PSMs might also have dependencies with
other PSMs. For instance, suppose the Utility Model
PSM required information from external market
agents – further analysis of interactions (albeit
simple enquiries) may be necessary to design the
agent to acquire this information.
Interaction dependencies between PSMs and
interaction-specific PSMs drive the agent-oriented
development of fish auction agents by using
ontology-based domain, task, and problem-solving
knowledge engineering where re-use and
extensibility are supported.
6 CONCLUSIONS
The use of domain ontologies in AOSE has recently
been investigated e.g. (Iglesias & Garijo, 2005).
That work has been limited to the early phases of
system development. It is our contention that the
potentially knowledge intensive nature of the
analysis involved towards creating the software
components in a Multi Agent System suggests that a
knowledge centred approach throughout the whole
software development cycle is effective. This
approach is underpinned by reusable knowledge
models concomitant with an appropriate set of
reusable domain problem solving processes (aka
methods) that operationalise corresponding chunks
of knowledge as required by the requirements of the
system. (Beydoun, Tran et al., 2006) presented a
methodological model underpinned by the presence
of PSM repositories ‘appropriately’ represented.
This paper bridges the gap between that work and
Buy Item
PSM
Sell Item
PSM
Interaction Dependency
Utility
Model
PSM
Max. Utility
Strategy
PSM
Bargaining
Protocol
PSM
Fish
market
mapping
Personal
fish-value
mapping
Fish-
buying
strategy
mapping
Bargaining
protocol
mapping
refinement refinementrefinement
Fish market
domain
ontology
output
Offer
AGENT
DESIGN
AGENT
ANALYSIS
AGENT
INTERACTION
DESIGN
Protocol
rules
ICSOFT 2009 - 4th International Conference on Software and Data Technologies
312
the representation required to formulate the PSMs
for interaction-dependent problem solving. We
introduce new constructs to model the interaction
dependencies of PSMs, and these are used by
software engineers in the analysis of solutions to
complex problems where interaction is required.
We illustrated these constructs in a simplified
development of a negotiation-based system.
Further work is required to create a formal
underpinning of interaction-dependent PSMs We are
in the process of developing a PSM library
containing interaction-specific PSMs for supporting
the development of negotiation agents in a variety of
real-world domains. Future work will also identify
and integrate software process steps required within
an agent-oriented methodology.
ACKNOWLEDGMENTS
This work is supported by the Australian Research
Council.
REFERENCES
Beydoun G., Krishna A.K., Ghose A. & Low G.C. 2007.
Towards Ontology-Based MAS Methodologies:
Ontology Based Early Requirements Information
Systems Development Conference. Galway.
Beydoun G., Tran N., Low G. & Henderson-Sellers B.
2006. Foundations of Ontology-Based Methodologies
for Multi-agent Systems: Springer LNCS, pp. 111-
123.
Bresciani P., Giorgini P., Giunchiglia F., Mylopoulos J. &
Perini A. 2004. TROPOS:An Agent Oriented Software
Development Methodology. Journal of Autonomous
Agents and Multi-Agent Systems, 8, 203-236.
Breuker J. 1999. Indexing Problem Solving Methods for
Reuse. Knowledge Acquisition, Modeling and
Management, 1621.
Bruaux S., Kassel G. & Morel G. 2005. An ontological
approach to the construction of problem-solving
models. K-CAP '05: Proceedings of the 3rd
international conference on Knowledge capture.
Cattoni R., Potrich A., Charlton P. & Mamdani E. 1999.
Evaluating the FIPA standards on the field: an audio
video entertainment applications. First Asia-Pacific
Conference on Intelligent Agent Technology.
Cuní G., Esteva M., Garcia P., Puertas E., Sierra C., et al.
2004. MASFIT: Multi-Agent System for Fish Trading.
Proceedings of the 16th European Conference on
Artificial Intelligence.
Decker S., Erdmann M., Fensel D. & Studer R. 1999.
Ontobroker: Ontology Based Access to Distributed
and Semi-Structured Information. DS, 8, 351-369.
Deloach S.A., Oyenan W.H. & Matson E.T. 2008. A
capabilities-based model for adaptive organizations.
Autonomous Agents and Multi-Agent Systems, 16, 13-
56.
Fatima S.S., Wooldridge M. & Jennings N. 2004. Optimal
negotiation of multiple issues in incomplete
information settings AAMAS04. New York: ACM.
Fensel D., Motta E., Benjamins V.R., Crubezy M., Decker
S., et al. 2002. The Unified Problem-solving Method
Development Language UPML. Knowledge and
Information Systems, 5, 83-131.
Fensel D., Motta E., Decker S. & Zdrahal Z. 1997. Using
Ontologies for Defining Tasks, Problem-Solving
Methods and Their Mappings. In: E. Plaza & R.
Benjamin (eds.) 10th European Knowledge
Acquisition Workshop (EKAW97). Spain: Springer, pp.
113-128.
Gómez M. & Plaza E. 2007. The ORCAS e-Institution: a
Platform to Develop Open, Reusable and Configurable
Multi-Agent Systems. International Journal of
Intelligent Control and Systems, 12, 130-141.
Iglesias C.A. & Garijo M. 2005. The Agent-Oriented
Methodology MAS-CommonKADS. In: B.
Henderson-Sellers & P. Giorgini (eds.) Agent-
Oriented Methodologies: IDEA Group Publishing, p.
46.78.
Jennings N., Faratin P., Lomuscio A., Parsons S., Sierra
C., et al. 2001. Automated negotiation: prospects,
methods and challenges. International Journal of
Group Decision and Negotiation
, 10, 199-215.
Lomuscio A., Wooldridge M. & Jennings N. 2001. A
Classification Scheme for Negotiation in Electronic
Commerce. Agent Mediated Electronic Commerce,
The European AgentLink Perspective, 19-33.
Müller J.-P. 2002. Emergence of Collective Behaviour and
Problem Solving. Engineering Societies in the Agents
World, 1-21.
Pavon J., Gomez-Sanz J. & Fuentest R. 2005. The
INGENIAS Methodology and Tools. In: B.
Henderson-Sellers & P. Giorgini (eds.) Agent-
Oriented Methodologies: IDEA Group Publishing, pp.
236-276.
Sandholm T. 1999. Distributed rational decision making.
Multiagent systems: a modern approach to distributed
artificial intelligence, 201 - 258.
Schreiber G., Akkermans H., Anjewierden A., Hoog R.d.,
Shadbolt N., et al. 2001. Knowledge Engineering And
Management: The CommonKADS Methodology.
London: The MIT Press.
Studer R., Benjamins V.R. & Fensel D. 1998. Knowledge
engineering: principles and methods. Data and
Knowledge Engineering, 25, 161-197.
Tran Q.-N.N. & Low G. 2008. MOBMAS: A
Methodology for Ontology-Based Multi-Agent
Systems Development. Information and Software
Technology, 50, 697-722.
TOWARDS PROBLEM SOLVING METHODS IN MULTI-AGENT SYSTEMS
313