DRL-Based Energy Efficient Power Adaptation for Fast HARQ in the Finite Blocklength Regime

In this paper, a point-to-point communication system with low latency and high reliability is studied. A fast hybrid automatic repeat request (HARQ) protocol is applied, where some HARQ feedback is omitted and the associated channel uses are incorporated for data transmission in fast HARQ. Based on relevant results on the decoding error probability over finite blocklength (FBL) codes, a long-term bit energy minimization problem is formulated in the presence of feedback delay and reliability constraints. Considering the non-convexity of the optimization problem and small decoding error probabilities, a finite-episode Markov Decision Process (MDP) with a double-layer penalty reward is formulated. An actor-critic based deep reinforcement learning (DRL) algorithm is subsequently designed. Through numerical evaluations, it is shown that compared with the conventional HARQ and the existing fast HARQ protocol, the proposed scheme is more energy efficient especially when the packet size is large.