Browsing by Subject "5G NR"

Machine learning (ML) is becoming increasingly important in the telecommunications industry. The purpose of machine learning models in telecommunications is to outperform a classical receiver’s performance by fine-tuning parameters. Since ML models have the advantage of being more concise, their performance is easier to evaluate, contrary to a classical receiver’s multiple blocks each with their own small errors. Evaluating the said models, however, is challenging. To identify the correct parameters is also not trivial. To address this issue, a coherent and reliant hyperparameter optimization method needs to be introduced. This thesis investigates how a hyperparameter optimization method can be implemented, and which one is best suited for the problem. It looks into the value it provides, the metrics displayed for each hyperparameter set during training and inference, and the challenges of realising such a system, in addition to various other qualities needed for an efficient training stage. The framework aims to provide valuable insight into model accuracy, validation loss, computing cost, signal-to-noise ratio improvement, and available resources when using hyperparameter tuning. The framework uses grid search optimization, Bayesian optimization as well as genetic algorithm optimization to determine which performs better, and compare the results between them. Grid search will act as a reference baseline for the performance of the two algorithms. The thesis is split into two parts: Phase One, which implements the system in a sandbox-like manner, essentially acting as a testing platform to assess the implementation compatibility. Phase Two inspects a more real-case scenario more suitable for a 5G physical layer environment. The proposed framework uses modern, widely used orchestration and development tools. These include ResNet, Pytorch, and sklearn.