In today's telecommunication industry, the explosively increasing traffic is pushing forward the evolution of network management strategies. To achieve a more scalable and flexible network architecture, network slicing has been proposed with key enabling technologies such as software-defined networks (SDN), network function virtualization (NFV), which has attracted significant attention from academia and industry. In network slicing design and management, the grand challenges faced by many network service providers (NSPs) and network operators is how to cost-effectively allocate resources to agilely accommodate a variety of applications and services with diverse quality-of-service (QoS) requirements, how to assign a service request to the right slice with performance guarantee and low resource cost, and how to manage the routing strategy to balance the bandwidth bottleneck and improve the network utility. Moreover, future network paradigm also tends to be more intelligent to address the challenge of dynamic nature of network traffic.

This dissertation presents a set of machine learning based tools to help NSPs and network operators make smart slice resource allocation at the planning stage and optimally utilize and scale slices at the operation stage. The dissertation aims to fill the gap of service requirement and resource quantity from three perspective: resource allocation of network slices with performance provisioning, network slice requests admission leveraging multiplexing gain, and reinforcement learning based routing evaluation and boosting.

First, we study the right amount of resources for a network slice to ensure the quality-of-service requirement with Stochastic Network Calculus theory. Given the traffic demand and quality-of-service requirement, our proposed tool is able to determine the size of a network slice. Consider the dynamic of network traffic arrivals, we study the traffic range supported by a network slice with allocated resources and propose an adaptive network slice resizing strategy to decide when and how much to resize with the change of traffic demand.

Second, we design a reinforcement learning based network slice assignment strategy that leverages the multiplexing gain among temporary diverse network slicing requests. Due to the high cost of deploying numerous network slices for the diversified QoS requirements from various applications, we propose the concept of network slice bundles with allocated resources and specified QoS where multiple slice requests can be assigned. By studying the time series feature of arrivals, our proposed solution is able to maximize the multiplexing gain through smoothing peak arrivals and increase resource utilization.

Third, we explore the intelligent traffic routing approaches from the packet level and epoch level (fixed time period with multiple packets) with different reinforcement learning algorithms. The proposed solutions promote intelligent algorithms and methods for the management of fundamental blocks in networks, which facilitate practical and sustainable tools for future network architecture.

The proposed tools in this dissertation will greatly facilitate dialogue between network service customers and network service providers and speed up the broad adoption of network slicing technology. The applicability of our algorithms in resource allocation management, traffic admission management, and traffic routing management can help design and optimize future generations of network systems.