Developing distributed software systems that operate in open, evolving and heterogeneous environments is becoming a critical issue in IT application domains such as web services and autonomic computing. Multi-Agent System (MAS) are offering both technological solutions as well as an effective paradigm for the development of this type of complex systems. Agents behavior at run time is influenced by the environment within which they operate [6], so designing agents able to behave in an effective way should benefit from a deep knowledge of that environment, including the stakeholders who have specific needs and expectations from the agents being designed. The specification of agents behavior is typically addressed during detail design, a step which precedes implementation. Analysis of the environment, on the other hand, is performed earlier, during domain and system requirements modelling, which rest on environment-and problem-oriented abstractions.
Our goal is to address the problem of supporting traceability between requirements, design and code artifacts, borrowing ideas and standards proposed by the Model Driven Architecture (MDA) initiative [2]. In particular, we conceives system development in terms of a chain of model transformations, namely, from a domain and stakeholder requirements model (Computationally Independent Model CIM, in MDA terminology) to a detailed design models of agents behavior, which provides a Platform Independent Model (PIM), and from a PIM to a Platform Specific Model (PSM), from which code and other development artifacts can then be straightforwardly derived.
To this aim, we extended the Tropos [1] agent-oriented methodology by introducing agent capability modeling since requirements analysis [4], and provided a tool-supported process. The Tropos methodology rests on a conceptual framework which includes the notions of agent, goal, plan and social dependency between agents for goal achievement. These knowledge level concepts are uniformly used throughout the software development process which is organized along five main phases: Early Requirements, Late Requirements, Architectural Design, Detailed Design and Implementation. We adopted a revised definition of the agent capability notion which distinguishes the concept of ability from the concept of opportunity [4]. The ability part represents a way to achieve a given goal. The opportunity part represents user preferences and environmental conditions, which may enable or disable the execution of the ability part, at run time.
An high level view of the development process we are proposing can be summarized as follows.
Step 1. Capability modeling starts during requirements analysis by identifying agent capabilities as a way to accomplish stakeholder needs (CIM model). Ability part of capabilities are specified in Tropos by a means-end relationship between the goal and the plan. Goal/plan ORdecomposition allows to describe alternative capabilities that can achieve a goal. The opportunity part is described via plan/softgoal contributions (e.g. see Fig. 1, (a)), plan, sof tgoal, metric Step 2. Further refinements of capability models are provided during architectural and detailed design where capability dynamic properties are specified by a set of UML 2.0 activity and sequence diagrams. These requirements analysis and design phases are supported by a Tropos modeler, based on the Eclipse platform, which uses a MOF [2] compliant implementation of the Tropos language metamodel and of the UML metamodel.
Step 3. Agent code generation results from a transformation of these PIM specifications to the target platform specific model (PSM). PIM to PSM transformation is implemented using automatic transformation techniques: for the ability part [4] while, for the opportunity part, a two steps process is adopted. The first step consists in a mapping from Tropos specification to Jadex, through the Jadex/JADE adapter [5], the second step goes from UML specification of capabilities to JADE. The output of this process is a skeleton of the code of JADE agents which execute plans, correspondingly to the ability specification. The choice of which plan to execute is performed by the agent according to BDI mechanisms which implement the opportunity specification.
Basic features of the resulting MAS implementation, are: a) each agent can play different roles according to special request-messages, related to target goals. That is, an agent can sense the environment and consequently switch to a specific role, execute the most appropriate capability, chosen by using the knowledge coded into the opportunity part; b) the ability part of a capability is implemented as a specialization of the class jade.core.behaviours.FSMBehaviour, namely it represents a final states machine (automaton). This allows to exploit monitoring mechanisms during capability execution with the aim of making the agent aware of failures and able to react to them.