Learning Reference Signal (LRS): Learning Intelligent Radio Access With Meta-Learned Reference Signal

Online adaptation to the dynamic channel conditions requires intelligent receivers with online learning ability as well as efficient training samples. In this letter, we propose a new type of reference signal termed learning reference signal (LRS), which serves as the online training samples for the fast adaptation of the deep neural network (DNN)-aided intelligent receiver. Specifically, we propose a model-agnostic meta-learning (MAML)-based LRS design framework, where the LRS sequence is regarded as the meta parameter and is meta-learned during the offline training. The maximum loss reduction criteria for LRS design is proposed such that the online meta-update based on LRS can maximize the reduction of the symbol error rate (SER). Furthermore, a Matthew effect in gradient-based training of LRS, which causes imbalanced update on different LRS symbols, is identified and then tackled by a novel symbol bundling and multi-stage updating method to ensure convergence. From experiments, we observe that the learned LRS contains both constellation points and non-constellation points, and achieves more than 4 dB SER gain compared to using arbitrary constellation points as training samples.