DDPG-Driven Deep-Unfolding With Adaptive Depth for Channel Estimation With Sparse Bayesian Learning

Deep-unfolding neural networks (NNs) have received great attention since they achieve satisfactory performance with relatively low complexity. Typically, these deep-unfolding NNs are restricted to a fixed-depth for all inputs. However, the optimal number of layers required convergence changes different In this paper, we first develop framework deep deterministic policy gradient (DDPG)-driven adaptive depth inputs, where trainable parameters NN learned by DDPG, rather than updated stochastic descent algorithm directly. Specifically, optimization variables, parameters, and architecture designed as state, action, state transition respectively. Then, is employed deal channel estimation problem in massive multiple-input multiple-output systems. formulate an off-grid basis sparse Bayesian learning (SBL)-based solve it. Secondly, SBL-based unfolded into layer-wise structure set introduced parameters. Thirdly, proposed DDPG-driven based on algorithm. To realize depth, design halting score indicate when stop, which function reconstruction error. Furthermore, extended general (DNNs). Simulation results show that outperforms conventional algorithms DNNs fixed much reduced layers.