We are developing a cognitive networking control and management architecture accepts multiple generations of networking components for high capacity, fast setup, and global reach. The overall goal of cognitive networking project is to architect, simulate, prototype, and experiment on a network with
- improved robustness and adaptability,
- improved usability and comprehensibility,
- improved security and stability, and
- reduced human intervention for operation and configuration
The areas of research that cognitive networking draws upon are cognitive networking, network control and management for heterogeneous networking, mobile ad-hoc networking, all optical networking, and photonic switching technologies
We use the OPNET simulation tool as part of the OPNET University Program.
1. Overview of Cognitive Networking in Cross-layer approaches
Current network control systems have limited ability to adapt to changing network conditions. By adding autonomous intelligence, based on machine learning, to the network management agents, it is possible to improve the Quality of Service (QoS) by reconfiguring the network management strategy around areas of interest such as user context, network state, inter-working scope and application demands. The self-configuration of network systems will have cross-layer ramifications for the protocol stack, from the physical (PHY), Medium Access Control (MAC), network, and transport layers to the middleware, presentation and application layers. Therefore, cross-layer design approaches are critical for the efficient utilization of limited resources with QoS guarantees in future wireless networks. The recent cross-layer architectures showed that a cross-layer approach is advantageous for wireless networks. Better network system performance can be obtained from context exchanges across protocol layers. We have developed self-configuration techniques with machine learning agents in a cross-layer approach as shown in Fig. 1, which could overcome the potential scope limitations of network management in heterogeneous wireless networks by allowing networks to observe, act, and learn in order to optimize their performance
Initial proposals on the implementation of cross-layer management interaction are discussed in the current literature. These approaches can be divided into three categories: direct communication between layers, a shared database across the layers, and novel contextual abstractions. We have exploited cross-layer interactions among layers by a shared network status module, supporting vertical communications among the layers, by acting as a repository for information collected by network protocols. Each layer-specific module can access the shared network status module to exchange its data and interact.
2. Machine learning for Cognitive Networking
Rather than devising an entirely new protocol based solely upon link-layer measurement, as others have done, our work employs reinforcement learning to make existing routing and management protocol more adaptive to network dynamics. We chose the AODV routing protocol for augmentation based upon the body of work available regarding the protocol’s performance, as well as its suitability for dynamic mesh environments. The overall research aim of our work is oriented towards cognitively intelligent network management. As such, the issue of creating an intelligent network-layer framework is the most compelling area in this nascent field. The first step in this overall goal is the dynamic tuning of routing protocol. Specifically, we employ the Q-Learning technique to reinforce the success or failure of route management processes, and modify our management policy accordingly. Fig 2 shows the learning and management flow by machine learning agents. Each time an action is executed, the agent receives an immediate reward from the environment. It then uses this reward and the expected long-term reward to update the Q-values, which in turn influences future action selection to enhance performance.