SNN-Based Early HARQ Predictor Design For 5G Networks

This paper studies the early hybrid automatic repeat request (E-HARQ) in the 5G new radio (NR). The earliest ARQ feedback to the transmitter within 0.2 msec, suitable for ultra-reliable-low-latency (URLLC) services, happens when the feedback indicates retransmission or a new data request is sent before the decoding process. In this case, the feedback is based on predicting the decoding outcome of the codeword bits sent as symbols with a specific modulation and coding scheme (MCS). Existing state-of-the-art neural network-based E-HARQ predictor exploits log-likelihood-ratio (LLR), calculated by the symbol detector, to predict ARQ feedback. They also did not include MCS as an input, so individual predictors are needed for different MCSs. This paper proposes a single NN-based E-HARQ predictor for different MCSs. The predictor has a single hidden layer, and it uses the channel estimates, the MCS information, the redundancy versions, and the approximate probability distribution function of LLRs at the receiver as inputs to predict decoding outcomes. Simulation results show that the proposed predictor reduces the latency of existing NN-based E-HARQ predictors and traditional HARQ by 46% and 60%, respectively. Its complexity is shown to be, on average, 99% lower than other predictors.