Análisis de algoritmos de aprendizaje máquina aplicados a la asignación de recursos en la red de acceso para redes 5G con segmentación

Abstract

This degree project focuses on the study and improvement of the random access method in 5G networks, applying two approaches for the dynamic management of preambles. In this context, users request uplink resources by transmitting preambles, which are characterized by being orthogonal signatures. However, one problem that

arises is the possible congestion and network collapse when multiple identical signa- tures are transmitted in a short period of time. The first method consists of the use of

supervised learning by training a neural network using a set of data that contains in- formation about the preambles that have previously benefited the system during tests

carried out. The neural network processes this information and generates outputs that indicate which preambles are best suited to avoid crashes and improve the probability of successful access to uplink resources. The second proposed method involves the use of reinforcement learning, which implements a policy for preamble management

based on Q-learning. In this approach, a function called Q is used that evaluates ac- tions taken in specific states of the system. The system continually feeds itself back,

reinforcing those choices that benefit the system’s performance and punishing those that harm it. In this way, the random access system can learn to allocate preambles more intelligently and effectively, adapting to network conditions and optimizing the utilization of available resources.