Adaptive Deep Learning Approaches for Power Control and Modulation Recognition in Dynamic Wireless Networks

Abstract

With a high and diverse demand for wireless communications, the optimization of communication systems became more complicated, so that the convex optimization-based traditional communication techniques are hard to satisfy the performance requirements of next-generation communication networks. To mitigate such limitations, the introduction of machine learning algorithms has received great attention. In other words, it leverages vast amounts of data to learn and adapt to optimization problems, meeting high requirements for data rate, latency, and energy consumption. However, deep learning-based approaches in modern systems have challenges with adaptability, scalability, potential inaccuracies, and suboptimal performance. To improve real-time applications, researchers and engineers must address these limitations and develop robust wireless communication systems. To this end, in this thesis, we propose a new deep learning-based power control method for maximizing the sum rate subject to rate requirements in the interference-limited device-to-device (D2D) communications and Bayesian learning approach for the modulation classification problem. For the first problem, based on the dynamic nature of D2D communications, we consider the environment where system parameters, such as the number of devices, rate requirements, and deployment area, unpredictably change over time. To deal with the low adaptability and scalability problems of the conventional deep learning-based approaches in dynamic environments, we develop an environment-adaptive power control method by leveraging graph neural network (GNN) architecture and meta-learning approach. In the developed method, we design the node feature and message update rule for GNN by taking into account the characteristics of power optimization problem and meta-train model by treating some past environments as meta-tasks. Simulation results verify that the developed method outperforms the conventional GNN-based power control methods in terms of the average sum-throughput achievement ratio by facilitating the model adaptation to new unseen environments. For the second problem, we aim to quantify the uncertainty of predictions of the deep learning model while maintain a high classification performance. Specifically, we propose a framework based on Long Short-Term Memory (LSTM) modules and Laplace approximation. The introduction of LSTM architecture into Bayesian learning not only improves the classification accuracy by effectively exploiting the temporal features but also enables to perform an uncertainty-aware incremental modulation classification. In our approach, we first pre-train the LSTM model and then use Laplace approximation on its last layer to create a lightweight model that is aware of uncertainty. Simulation results demonstrate that our method is superior to frequentist methods in terms of predicting uncertainty and classification accuracy.