Cognitive Networks: Towards Self-Aware Networks

5.4. Open Issues and Research Challenges

Most research in the field of machine learning has been motivated by problems in pattern recognition, robotics, medical diagnosis, marketing and related commercial areas. This accounts for the predominance of supervised classification and reinforcement learning in current research. The networking domain requires several shifts in focus and raises several exciting new research challenges, which we discuss in this section.

5.4.1. From Supervised to Autonomous Learning

As we have seen above, the dominant problem formulation in machine learning is supervised learning, where a 'teacher' labels the training data to indicate the desired response. While there are some potential applications of supervised learning in Knowledge Plane applications (e.g., for recognizing known networking misconfigurations and intrusions), there are many more applications for autonomous learning methods that do not require a teacher. In particular, many of the networking applications involve looking for anomalies in real-time data streams, which can be formulated as a combination of unsupervised learning and learning for interpretation.

Anomaly detection has been studied in machine learning, but usually it has considered only a fixed level of abstraction. For networking, there can be anomalies at the level of individual packets, but also at the level of connections, protocols, traffic flows, and network-wide disturbances. A very interesting challenge for machine learning ...