利用行動代理人的互助式學習之網路管理系統

Abstract

網路管理系統之主要任務是協助網路管理者監控整個網路以提供快速且有效率的服務以確保資訊的通訊品質。從網路專家的角度，可用下列四個準則來評估一個網路管理系統的良莠：「自動化」、「即時性」、「耐衝擊性」與「系統複雜度」。「自動化」即一個網路管理系統可以有多少比例的網路管理相關工作自動處理，而不需管理者協助。「即時性」即網路管理系統自動處理網路的管理相關工作，尤其是網路障礙的排除，可以多快處理完成。「耐衝擊性」即網路任何節點或連線，尤其是網路管理系統所在節點，發生故障時，網路管理之運作是否正常如昔。節點上網路管理系統之「系統複雜度」指網路管理系統所處之各節點耗用多少系統資源。 網路管理系統的架構隨著網路複雜度的提升由中央集權式的管理架構演變成階層式的網路管理架構，但仍然無法完全滿足上述之管理系統的需求，因此有人提出應用軟體代理人於網路管理系統之構想。行動代理人是軟體代理人中技術最為成熟的一種，目前的研究均肯定應用行動代理人於網路管理確可提升管理效率，但現今之研究成果大都將重心放在行動力上而忽略了代理人的自動化、學習能力與合作能力等特點，造成網路管理系統的改善不如預期。 本研究針對提升網路管理系統之自動化、即時性、耐衝擊性與降低端點系統複雜度等四個網路管理系統需求，提出提升行動代理人在網路管理系統之應用效率的策略，設計出行動代理人之間能夠以互助的方式主動學習彼此的經驗以加速障礙排除之網路管理系統：為提升自動化，行動代理人具備學習知識的能力，能將知識主動地分享給其他人，以豐富自己的知識庫；為提升即時性，行動代理人間互相支援以分擔彼此的工作量並且共同合作將複雜的警報作關連性的處理，以便立即找出障礙發生的地點；為提升耐衝擊性，系統中重要的節點有備份的機制；為降低端點的管理系統複雜度，行動代理人之知識庫將精簡化分散於節點間。 本文詳述此管理系統之系統架構及四個與組態管理和障礙管理相關的模組之設計原理及網管知識建立與分佈規則，並實作其中的「拓樸圖建立」模組以驗證所提行動代理人之實用性。

The main functionality of Network Management System (NMS) is to help network managers monitor and control the whole network for providing fast and efficient services to ensure the quality of information communication. From the perspective of network managers, we propose following four metrics to evaluate the performance of a NMS: Automation, Real-Time, Robustness and Management System Complexity. ‘Automation’ means the percentage of a NMS that can deal with the network-related tasks automatically without intervention from managers. ‘Real-Time’ means NMS can handle the tasks as soon as possible, especially for fault correction. ‘Robustness’ means any component crashes in the network, NMS can still work as usual. ‘Management System Complexity’ means the quantity of management system resource consumed by each node in NMS. Although the fast development of network has pushed NMS to evolve from centralized paradigm to hierarchy paradigm, it’s still with high system complexity and lack of real-time and robustness. Therefore, the most mature software agent, mobile agent, was proposed for NMS. Most researches think that mobile agents can promote the efficiency of management. But researches so far put the emphasis on the mobility and omit the other characteristics of mobile agents, such as automation, learning skill and cooperation. So the efficiency of using mobile agent in NMS is not so good as expectation. To improve the ‘Automation’ capability of NMS, mobile agents should learn and share knowledge automatically to enrich their own knowledge repository. To improve the ‘Real-Time’ capability of NMS, mobile agents should help each other to balance the workload and correlate the complex fault alarms by cooperation. To improve the ‘Robustness’ capability of NMS, there should be a backup mechanism with each important node in the system. To improve the ‘Management System Complexity’ capability of NMS, the knowledge repository should be simplified and distributed among all nodes. Based on these capability requirements, a collaboratively learning NMS is proposed that all the mobile agents in this system can learn experience mutually by help each other. The system architecture, detail design of the configuration/fault management-related modules and the principles of distribution and building of management knowledge of the NMS are described in this thesis.