Research On Fast Network Resource Allocation With Machine Learning Method

Software-Defined Networks is an emerging architecture that separates the control plane and data plane. A logical centralized controller is responsible for all the control decisions and switches in data plane only simply forward packets. This paradigm enables flexible network dataflow management. Confronting with increasing network lord pressure and complex application requirements, it has already become a very important research topic to accomplish reasonable resource allocation and efficiently enhance network performance under such kind of centralized control architecture.Through the mathematical formulation for this complex traffic engineering problem, we have demonstrated that it is a NP-complete problem. Although currently many advanced heuristic algorithms have been proposed for network resource allocation, especially dynamic routing decision problem, they lead to relatively high computational time cost. So this will bring about a crucial problem whether such a heuristic approach to this NP-complete problem is of any use in practice.This master thesis proposes a novel idea for accomplishing dynamic routing decision in real-time. For a reliable network environment, in the controller we construct a routing decision meta-layer with multiple supervised machine learning models. In order to make sure our trained machine learning models have same capability as heuristic algorithm that can make routing decision according to current network state, we consider heuristic algorithm’s input and its corresponding output as training sample’s feature and label respectively. After completing all models’ training process, the meta-layer can substitute for the time-consuming heuristic algorithm. When our controller receives a new communication request, it can independently give heuristic-like optimal result. What’s more, it will cost little computational time, which is quite feasible for practical network.On top of that, for the sake of more effectively extracting useful network features, and making our machine learning model take all constraint conditions and optimization objectives into its decision consideration, we adopt the deep learning principle and propose a Gaussian-binary conditional classification restricted Boltzmann machine as our base model in meta-layer. Furthermore, we present its sufficient mathematical inference and demonstration, including specific training algorithm. Then combining with simulation experiments, we prove that Gaussian-binary conditional classification restricted Boltzmann machine can successfully apply in this dynamic routing decision framework and achieve fast network resource allocation.