A Multi-Agent System for Telephone Network Management1

This paper first briefly describes an approach to constructing real-time, high-performance multi-agent systems called agent-oriented programming (AOP). The main part of the paper illustrates the application of this approach by showing how it can be used in telephone network management. Finally, a prototype telephone traffic manager agent is described. This prototype agent has been implemented at the Open University as a demonstration of the concepts, using a combination of the LISP, REX and GAPPS programming languages. Agent-oriented programming Agent-oriented programming can be throught of as an extension of object-oriented programming (Kiss,1990a; Shoham, 1991). While in OOP the objects passively preserve state information between invocations (receipt of messages), in AOP the basic elements of the software are active agents. By taking action, agents actively attempt to satisfy a value system that describes what is desirable. In order to satisfy their values, agents derive goals from them and then form intentions to take actions to reach these goals. The use of value systems produces systems with capability for autonomous (unsupervised) action because the inclusion of a value hierarchy ensures appropriate action over a wide range of situations. Agents know facts about their environment and about themselves. They communicate with each other by sending messages containing declarative facts, requests for action, declarations of values or goals, etc., which are all instances of communicative actions. The problemsolving capabilities inherent in agents enables them to communicate requests very concisely and declaratively in terms of goals rather than procedural action sequences. The internal architecture of each agent in AOP is a layered structure. The higher layers are dedicated to achieving generality (robustness), while the lower layers are dedicated to achieving real-time performance. This architecture is based on a principle of power-versus-generality tradeoff. The degree of causal coupling to the environment is direct at the lower levels and progressively indirect at the higher levels in terms of space and time (for more details, see Kiss, 1991). Each agent is controlled by an agent program. Execution of agent programs can be concurrent on multiprocessor hardware or multiprogrammed on single-processor hardware. Execution can be based on interpretation or compilation. Interpretive execution is based on the concept of an action cycle. The action cycle determines what action the agent is to take at each time increment in the currrent situation. Interpretive execution is well suited to the higher levels of the agent architecture (Kiss, 1991). Compilation results in low-level machines that are best described as various forms of automata. The simplest form is a logic circuit. This is the form used in the REX system of Rosenschein and Kaelbling (Kaelbling and Wilson, 1988). More complex forms are sequential machines or cellular automata. Many advantages accrue from compilation: fast performance (speedups of 103-105 compared to LISP); embeddable run-time modules; possibility of hardware implementation, etc. Agents in Network Management We advocate that the introduction of agents into network management systems should follow the hierarchical architecture that is already widely adopted in the OSI philosophy. For example, British Telecom's Open Network Architecture (ONA-M) includes five layers: Business Management (part of the overall company management), Service management (manage the information flow in the network from the functional point of view), Network Management (these supervise a number of element management systems), Element Management (hardware or software built into network elements), Network Elements (modems, lines, etc). Within each layer, several functions are distinguished (but the tasks carried out at each layer differ widely): Financial management, Configuration management Resource management, Event management, Performance management, Planning and design management, Access and security management. It is easy to see that both layers (horizontal modularisation) and functions (vertical modularisation) are good candidates for the identification of agents.