1. Application of Agent- Oriented Techniques to Network Supervision Babak Esfandiari, Mitel Corporation

2. Different opportunities for agents in networks n Routing n Network Management u Network Supervision u GDMO/CMIS u TMN

3. Network Supervision Problematic n Fault detection n Alarm filtering and qualification n Multiple and cascading faults n...

4. Existing attempts n Mostly use of expert systems for diagnosis ([Gaiti] [Garijo]…) u Use of agent-oriented architectures (?) u Revealed the importance of explicit representation of Time u No high-level communication between network management platforms u Acquisition of expertise is still a bottleneck

5. Chronicles Chronicle RobotLoadMachine { event (Robot: (outRoom, inRoom), e1); event (Robot: (inRoom, outRoom), e4); event (MachineInput: (unLoaded, loaded), e2); event (Machine: (Stopped, Running), e1); e1 < e2; 1’ <= e3 - e2 <= 6’; 3’ <= e4 - e2 <= 5’; hold (Machine: Running, (e2, e2)); hold (SafetyConditions: True, (e1, e4)); when recognized {report “Successful load”;} }

6. Some theoretical speculations: Agents and OSI layers n Using Newell’s Knowledge Level as the highest communication layer? u Expressing applications behaviors in “rational” terms (Beliefs, Desires, Intentions, …) u Communicating such terms using high-level interaction languages (KIF/KQML?) and protocols

7. Interface Agents “A program that […] provides assistance to a user dealing with a particular application. Such agents learn by watching over the shoulder of the user and detecting patterns and regularities…” (Maes)

8. The Assistant’s structure

9. Use of BDI to specify the agent’s behavior n Trust as a modal operator n B(a,f) Λ Trusts(a,b,f) -> K(a,f) n Trusts(a,a,f) ? n Trusts(a,human operator,*) n Trusts(a,b,f) with b := other agent ?

10. Learnability of chronicles: a set of Oracles n PASSIVE: supplies events and actions n PASSIVE S : PASSIVE + no overlapping n ACTIVE MQ : {events}+action -> yes/no n ACTIVE EQ :chronicle ->yes/(no+example)

11. Learnability of chronicles: Results n With one chronicle per action: u positive with PASSIVE S u positive with ACTIVE MQ + PASSIVE n If more than one chronicle per action: u negative with any oracle Ô Difficulties: Ô overlapping Ô x chronicles/action Ô where find such oracles ?

12. The Learning System 3 steps: n Chronicle creation n Chronicle evaluation n Chronicle confirmation

15. An example (1) Evaluation of: a b c ->  Unconfirmed chronicle base: 1: a b c d ->  Trust: 1 2: a b c e ->  Trust: 1 3: a b c f ->  Trust: 2 Confirmed chronicle base: 1: a b c g ->  Trust: 3

16. An example (2) Unconfirmed chronicle base: 1: d ->  Trust: 1 2: e ->  Trust: 1 3: a b c f ->  Trust: 2 Confirmed chronicle base: 1: a b c g ->  Trust: 3 2: a b c ->  Trust: 3

17. MAGENTA: MAnaGEmeNT Application or Multi-AGENT Assistant ? ObjectManager: processes the query CommunicationManager: sends and receives messages EventManager: triggers event notifications Management Application: processes the events and publishes queries

18. Experimentation n The local network n Transpac data n Robot behavior pattern detection n Help to a Smalltalk programmer n Overlapping management n Collaborative learning

19. Results

20. Finding other oracles: Collaborative assistance Presentation protocol Matchmaking protocol Query protocol

21. Conclusion and perspectives n Summary: u Use of Interface Agents in Network Supervision F Theoretical results on chronicle learning F Appropriate algorithms u Use of Network Management standards to build an Agent Development platform n Next: u Improve the algorithms (first order, partial order) u Big scale experimentation u Other applications of MAGENTA: remote programming, distributed debugging,...