Electrical and Computer Engineering ETDs

Publication Date



The leading role of the HetNet (Heterogeneous Networks) strategy as the key Radio Access Network (RAN) architecture for future 5G networks poses serious challenges to the current cell selection mechanisms used in cellular networks. The max-SINR algorithm, although effective historically for performing the most essential networking function of wireless networks, is inefficient at best and obsolete at worst in 5G HetNets. The foreseen embarrassment of riches and diversified propagation characteristics of network attachment points spanning multiple Radio Access Technologies (RAT) requires novel and creative context-aware system designs. The association and routing decisions, in the context of single-RAT or multi-RAT connections, need to be optimized to efficiently exploit the benefits of the architecture. However, the high computational complexity required for multi-parametric optimization of utility functions, the difficulty of modeling and solving Markov Decision Processes, the lack of guarantees of stability of Game Theory algorithms, and the rigidness of simpler methods like Cell Range Expansion and operator policies managed by the Access Network Discovery and Selection Function (ANDSF), makes neither of these state-of-the-art approaches a favorite. This Thesis proposes a framework that relies on Machine Learning techniques at the terminal device-level for Cognitive RAT Selection. The use of cognition allows the terminal device to learn both a multi-parametric state model and effective decision policies, based on the experience of the device itself. This implies that a terminal, after observing its environment during a learning period, may formulate a system characterization and optimize its own association decisions without any external intervention. In our proposal, this is achieved through clustering of appropriately defined feature vectors for building a system state model, supervised classification to obtain the current system state, and reinforcement learning for learning good policies. This Thesis describes the above framework in detail and recommends adaptations based on the experimentation with the X-means, k-Nearest Neighbors, and Q-learning algorithms, the building blocks of the solution. The network performance of the proposed framework is evaluated in a multi-agent environment implemented in MATLAB where it is compared with alternative RAT selection mechanisms.


machine learning, 5g, user association, load balancing, clustering, reinforcement learning

Document Type




Degree Name

Electrical Engineering

Level of Degree


Department Name

Electrical and Computer Engineering

First Committee Member (Chair)

Martinez-Ramon, Manel

Second Committee Member

Christodoulou, Christos