Software agents in network management

The global purpose of this thesis is to investigate Software Agent technology in the Network Management (NM) field. The thesis starts by surveying software agent litterature, thus providing a synthetic view of agent trends and techniques. The survey shows the richness of software agents in terms of potential applications, techniques, mechanisms and languages. In order to determine the best way software agents are applied to NM, the thesis presents a state-of-the-art of agent-based NM applications so far developed. These applications are analyzed according to the different kinds of agent architectures and approaches. This allowed to independently assess each agent technique when applied to NM purposes. The result of the analysis shows that different kinds of software agent approaches are suitable for tackling different NM issues, and that in order for NM to get the maximum benefit of software agents, an agent approach should not be restricted to a unique kind of agents, or a limited set of agent techniques. According to this outcome, the thesis describes a generic skill-based agent architecture targeted to NM applications. This agent architecture is derived from horizontally-layered agent architectures, and therefore inherits the same properties of modularity and equitability between agent layers. Horizontal layers are represented by skill modules that encapsulate agent capabilities and management functionaliy. Skills can be dynamically plugged into the agent without interrupting its execution. The agent brain, which encodes the basic functionality required for the agent operation, is capable of discovering the capabilities brought by loaded skills and dynamically integrating these capabilities to improve the agent behavior with new competences. The skill-based architecture defines a transparent inter-agent communication mechanism that allows the easy implementation of different agent coordination patterns. In addition to these features, the skill-based agent architecture is designed with the properties of flexibility and support for dynamism of NM functionality. The ensemble of these features is validated in the context of two case studies. The first case study considered the implementation of hierarchically distributed NM system in which, management domains could be dynamically reaffected according to changes in the network topology, or to the availability of domain management agents. In addition, a reliability layer is added to insure that the management tasks are reliably performed on the whole network even in the case that a subset of the domain agents become unreliable. The obtained agent system exhibits the properties of dynamism, reliability, fault tolerance, and graceful degradation. The second case study considered the mundane task of the automatic provision of Permanent Virtual Connection services in heterogeneous ATM networks. The skill-based agent architecture is used to implement distributed agents that collaborate together to achieve end-to-end connectivity services in an efficient and cost-effective way. In order to assess the skill-based architecture with possible other agent approaches, the two case studies are reconsidered using two of the most hyped types of agents, namely: BDI (Belief-Desire-Intention) agents, and mobile agents (MA). For BDI agents, the thesis proposes an abstract BDI agent model specially developed for NM applications. A BDI-oriented design process is also proposed and used in order to provide a different implementation of the first case study. The comparison with the skill-based agent implementation allows to show that the proposed BDI model presents powerful abstractions. The skill-based architecture provides howerver a higher degree of flexibility and ease of design. For MAs, the thesis compares an MA approach to the second case study with the skill-based approach. The comparison considers different qualitative and performance parameters. The results of the comparison allow to pinpoint the actual potential benefits of MAs: Simplifying the design of self-contained distributed management tasks and providing a high degree of flexibility and ease of deployment. Skill-based agents still provide better performance, a sufficient degree of flexibility, in addition to being suitable for general-purpose management applications and insensitive to the complexity of management functions. Therefore, the adopted skill-based architecture provides the best compromise to deal with the requirements of modern NM systems and approaches. The thesis concludes with the identification of major pitfalls in the application of software agents in NM, namely the disillusion of obtaining highly intelligent management software with smart properties of learning, proactiveness and self-adaptability. The real benefit of software agents in NM relies in the ability to adapt the large set of agent techniques to tackle NM issues.