Data-Driven Joint Modulator and Demodulator Based on Meta-Learning for HARQ Schemes

In this paper, we propose a data-driven joint modulator and demodulator based on meta-learning for hybrid automatic repeat request (HARQ) schemes, which include Type-I HARQ, HARQ with chase combining (HARQ-CC), and HARQ with incremental redundancy (HARQ-IR). Data-driven deep learning (DL) techniques have proven particularly effective for enabling wireless systems to adapt to complex scenarios with unknown or imperfect channel conditions that are difficult to characterize analytically. Most existing DL approaches to realize data-driven communication frameworks are based on joint source-channel coding (JSCC). However, the difficulty in detecting transmission errors makes JSCC unsuitable for HARQ systems. To address this limitation, we propose a novel architecture that combines classical channel coding with DLbased modulation and demodulation. Specifically, we employ an autoencoder to implement both the modulator and demodulator. To quickly adapt to time-varying fading channels, the ModelAgnostic Meta-Learning (MAML) algorithm is leveraged to accelerate the convergence of the model as well as enhance the reliability of HARQ. The MAML framework is divided into two phases, including the pre-training phase to acquire shared initialization parameters across various channel conditions and the fine-tuning phase to rapidly adapt to specific channel states. Simulation results demonstrate that the proposed architecture offers dramatic improvements in convergence speed and generalization capabilities, and meanwhile boosts the reliability of HARQ schemes.