Managing Large-Scale Computer Communication Networks Using Adaptive Learning Systems

Georgia Tech Research Corporation, Atlanta GA

Investigators

Abstract

9908578 Ji Managing large computer communication networks through distributed agents is an emerg-ing area of research towards automated and scaleable network management. In the context of this research, an agent is a simple algorithm which can reside at a network node, and perform local decisions on the state of a network using local information. As many such agents can be used in a distributed fashion at a large network, little is known how local decisions result in collective effects at network level, and what decentralized algorithms can be used by agents at network nodes so that a good global performance can be achieved network-wide. To provide answers to these questions is crucial to advancing network management from device-centric to network-centric which is required by the next generation networks. The goal of this project is to investigate adaptive learning approaches as potential solutions to these open problems. In particular, dynamic probabilistic graph models will be investigated and shown to provide a systematic approach for accomplishing the following tasks: (1) developing (local) models to cooperate and aggregate information form a neighborhood of agents, (2) using the (local) models developed to perform distributed decisions at network nodes adaptively, (3) developing (global) model by aggregating local statistics to characterize resulting global effects of a network of agents. A rich set of methods in the area of adaptive learning systems, especially the dynamic graph models, will be shown to provide both a theoretical foundation, and practical approaches in managing large networks with an irregular and a changing topology, and inaccessible network components. ***