Deep Learning for Blocklength Optimization in Fully Connected RIS-Aided Short-Packet NOMA

In this paper, we propose a deep learning (DL)based framework to optimize fully connected reconfigurable intelligent surface assisted downlink short packet communication. The system comprises of a multi-antenna base station and multiple user equipment, with an objective to jointly optimize beamforming, RIS phase shifts, and block lengths to maximize the finite blocklength (FBL) rate while meeting ultra-reliable low-latency communication requirements. A gradient-based DL framework is proposed to optimize this high-dimensional and non-convex problem. The proposed approach efficiently addresses the complexity of joint optimization over Rician fading channels. Simulations show significant improvements in sum rate and resource allocation compared to traditional techniques, such as traditional machine learning and gradient descent.