Heterogeneous networks, wireline and wireless, are becoming increasingly complex in their architecture. Heterogeneity in networking components and architectures leads to multi-dimensional variable representations. Examples of heterogeneous variables include those that describe topologies, channel conditions, network-flows, and physical failures. These variables exhibit complex dependencies at different spatial-temporal scales. Meanwhile, as the network-size increases, network state representations can involve thousands of heterogeneous variables. This results in scalability challenges.
In this talk, we discuss when and how machine learning approaches can help meet challenges of heterogeneity and scalability in effective network modeling and management. In particular, we appeal to multi-layer probabilistic graphical models to study overall network performance and derive cross-layer network control and management features. We first provide a general multi-layer graphical model and mathematical representation for network overlays. We then present two examples involving hierarchical modeling, management and salability analysis. We identify several directions for generalizing the hierarchical modeling and management, and discuss open issues and possible future directions.