| Today's network technology environment is varied and diverse. Disparate technologies, such as Optical Fiber, DSL, 802.11 Wireless LANs and Wi-Max, are deployed in varying localities, sometimes simultaneously, and expected to work seamlessly through the use of the TCP/IP protocol stack. While TCP/IP has allowed for substantial interoperability among such technologies, there remain limitations. TCP's incorrect inference of link congestion as the sole source of packet losses in wireless networks is one such example. Network management of diverse network deployment scenarios, be they multi-hop wireless, wire-lined or heterogeneous environments, point to the need for fully automated network management to comprehend the underlying technologies used, their inherent limitations, so as to fully maximize the optimizations performed in terms of policy-based management decisions. Such a feat is by no means trivial to engineer. However, this dissertation endeavors to illustrate a conceptual framework, coupled with experimental design and analysis, in order to convey the feasibility and utility of this approach, namely the approach of taking the human out of the decision-making loop in network management systems.;In this dissertation, we study cognitive agent-based optimization of network processes in three scenarios: a) choosing ideal cluster-head nodes in purely multi-hop wireless network deployment based upon stability metrics, b) dynamically tuning the update interval in the AODV routing protocol to lessen the overhead and increase overall packet delivery rates, and c) introducing multi-domain intelligent resource brokering, leveraging machine learning techniques to improve the decision making abilities of the agents. And in all three cases we illustrate the ability of these network agents to observe network state, leverage machine learning to extract understanding from those state observations, and determine configuration decisions in an optimized and autonomous fashion. |
