Power Allocation Strategies of Non-Orthogonal HARQ: From Deep Learning Perspective

This paper introduces a power-constrained Non-Orthogonal Hybrid Automatic Repeat Request (N-HARQ) scheme to enhance communications in edge networks, aiming to improve the transmission reliability and reduce the latency between edge servers and devices. To address the challenge of deriving a closed-form expression for the outage probability of N-HARQ, we propose a twin evaluation network which leverages the concept of the digital twin (DT) to provide an accurate outage approximation. We then formulate a long-term average throughput (LTAT) maximization problem to ensure constrained power consumption and high reliability over fading channels. Considering the non-convex and intractable nature of the LTAT maximization problem, we propose three distinct deep learning approaches to design the optimal power allocation strategies, including Graph Convolutional Network (GCN), Long Short-Term Memory (LSTM), and Multi-Layer Perceptron (MLP). The trainable weights of these neural networks are adjusted by using the primal-dual learning method. Numerical results demonstrate the effectiveness of the proposed twin evaluation network and the three deep learning approaches in enhancing throughput and reliability. The GCN shows superior performance in both metrics, then followed by the LSTM-based method, whereas the MLP-based method exhibits the worst performance.