Towards machine learning enabled future-generation wireless network optimization
Master of Science
SubjectRadio-frequency identification networks
Mobile edge computing
Wireless ad-hoc networks
MetadataShow full item record
We anticipate that there will be an enormous amount of wireless devices connected to the Internet through the future-generation wireless networks. Those wireless devices vary from self-driving vehicles to smart wearable devices and intelligent house- hold electrical appliances. Under such circumstances, the network resource optimization faces the challenge of the requirement of both flexibility and performance. Current wireless communication still relies on one-size-fits-all optimization algorithms, which require meticulous design and elaborate maintenance, thus not flexible and cannot meet the growing requirements well. The future-generation wireless networks should be “smarter”, which means that the artificial intelligence-driven software-level design will play a more significant role in network optimization. In this thesis, we present three different ways of leveraging artificial intelligence (AI) and machine learning (ML) to design network optimization algorithms for three wireless Internet of things network optimization problems. Our ML-based approaches cover the use of multi-layer feed-forward artificial neural network and the graph convolutional network as the core of our AI decision-makers. The learning methods are supervised learning (for static decision-making) and reinforcement learning (for dynamic decision-making). We demonstrate the viability of applying ML in future- generation wireless network optimizations through extensive simulations. We summarize our discovery on the advantage of using ML in wireless network optimizations as the following three aspects: 1. Enabling the distributed decision-making to achieve the performance that near a centralized solution, without the requirement of multi-hop information; 2. Tackling with dynamic optimization through distributed self-learning decision- making agents, instead of designing a sophisticated optimization algorithm; 3. Reducing the time used in optimizing the solution of a combinatorial optimization problem. We envision that in the foreseeable future, AI and ML could help network service designers and operators to improve the network quality of experience swiftly and less expensively.