Distributed policy learning for the Cognitive Network Management System

The Cognitive Network Management System (CNMS) is a joint research initiative between Rockwell Collins and the Institute for Human Machine Cognition that aims to provide automated, policy-based real time network management for complex MANET networks. CNMS is a lightweight and efficient policy management framework designed to mitigate the need for centralized network management, reduce operator hands-on time, and increase network reliability. CNMS provides the necessary reasoning and enforcement mechanisms for the on-demand management of network topology and communication resources. Furthermore, it supports fully distributed policy learning mechanisms that enable networked devices to adapt at run-time to unanticipated network conditions and application requirements by creating and distributing learned policies. In this paper we describe the CNMS architecture and functionality. We focus our discussions on the CNMS architecture and policy learning mechanisms. Policy learning in CNMS is intrinsically distributed, and based on network performance observations for the refinement of contexts, and actions. In this paper we describe two examples for policy adaptation, one based on link capacity monitoring, and one based on adaptive frequency hoping strategies for interference mitigation. We then present some results from NS-3 evaluations of these two examples. CNMS has been implemented and tested in a two channel wireless testbed using the Universal Software Radio Peripheral (USRP) and 802.11 wireless networking communication. We briefly discuss the testbed and describe some of our experimental results followed by a brief discussion of our findings, and recommendations for future work.