| dc.description.abstract | The rapid uptake of mobile devices and applications are posing unprecedented traffic burdens on the existing networking infrastructures. In order to maximize both user experience and investment return, the networking and communications systems are evolving to the next gen- eration – 5G, which is expected to support more flexibility, agility, and intelligence towards provisioned services and infrastructure management. Fulfilling these tasks is challenging, as nowadays networks are increasingly heterogeneous, dynamic and expanded with large sizes. Network softwarization is one of the critical enabling technologies to implement these requirements in 5G. In addition to these problems investigated in preliminary researches about this technology, many new emerging application requirements and advanced opti- mization & learning technologies are introducing more challenges & opportunities for its fully application in practical production environment. This motivates this thesis to develop a new learning augmented optimization technology, which merges both the advanced opti- mization and learning techniques to meet the distinct characteristics of the new application environment. To be more specific, the abstracts of the key contents in this thesis are listed as follows: • We first develop a stochastic solution to augment the optimization of the Network Function Virtualization (NFV) services in dynamical networks. In contrast to the dominant NFV solutions applied for the deterministic networking environments, the inherent network dynamics and uncertainties from 5G infrastructure are impeding the rollout of NFV in many emerging networking applications. Therefore, Chapter 3 investigates the issues of network utility degradation when implementing NFV in dynamical networks, and proposes a robust NFV solution with full respect to the underlying stochastic features. By exploiting the hierarchical decision structures in this problem, a distributed computing framework with two-level decomposition is designed to facilitate a distributed implementation of the proposed model in large-scale networks. • Next, Chapter 4 aims to intertwin the traditional optimization and learning technologies. In order to reap the merits of both optimization and learning technologies but avoid their limitations, promissing integrative approaches are investigated to combine the traditional optimization theories with advanced learning methods. Subsequently, an online optimization process is designed to learn the system dynamics for the network slicing problem, another critical challenge for network softwarization. Specifically, we first present a two-stage slicing optimization model with time-averaged constraints and objective to safeguard the network slicing operations in time-varying networks. Directly solving an off-line solution to this problem is intractable since the future system realizations are unknown before decisions. To address this, we combine the historical learning and Lyapunov stability theories, and develop a learning augmented online optimization approach. This facilitates the system to learn a safe slicing solution from both historical records and real-time observations. We prove that the proposed solution is always feasible and nearly optimal, up to a constant additive factor. Finally, simulation experiments are also provided to demonstrate the considerable improvement of the proposals. • The success of traditional solutions to optimizing the stochastic systems often requires solving a base optimization program repeatedly until convergence. For each iteration, the base program exhibits the same model structure, but only differing in their input data. Such properties of the stochastic optimization systems encourage the work of Chapter 5, in which we apply the latest deep learning technologies to abstract the core structures of an optimization model and then use the learned deep learning model to directly generate the solutions to the equivalent optimization model. In this respect, an encoder-decoder based learning model is developed in Chapter 5 to improve the optimization of network slices. In order to facilitate the solving of the constrained combinatorial optimization program in a deep learning manner, we design a problem-specific decoding process by integrating program constraints and problem context information into the training process. The deep learning model, once trained, can be used to directly generate the solution to any specific problem instance. This avoids the extensive computation in traditional approaches, which re-solve the whole combinatorial optimization problem for every instance from the scratch. With the help of the REINFORCE gradient estimator, the obtained deep learning model in the experiments achieves significantly reduced computation time and optimality loss. | en_GB |
