Beyond-Diagonal RIS for Wideband NOMA: OFE-REDQ with SIC-Aware Power Allocation

Orthogonal multiple access assigns disjoint time-frequency resources and limits reuse in dense 6G settings. Non orthogonal multiple access mitigates this by superposing users and separating streams with successive interference cancellation, but performance is sensitive to SIC errors and channel uncertainty. Many RIS-NOMA studies assume perfect channels and ideal SIC, ignore feedback delay and noise, and restrict RIS to diagonal control, while the wideband joint design of base-station precoding, RIS configuration, and inter/intra-cluster power is nonconvex. In this research we address these gaps with a wideband OFDM downlink that employs beyond diagonal RIS with two-bit eigen-phase quantization, imperfect feedback (delay, noise, random phase flips), MMSE precoding from estimated cluster representatives, and an explicit per-subcarrier SIC feasibility rule. We develop OFE-REDQ, a reinforcement-learning agent that combines an optimized feature encoder with a randomized ensemble of critics and a deterministic actor to jointly control RIS settings, inter-cluster power shares, and intra-cluster splits. Stability is ensured by Polyak target updates, subset-minimum bootstrapping, and target-value clipping. In 3GPP TR 38.901 CDL-D UMi simulations, OFE-REDQ achieves higher sum rate, faster and more stable convergence, and higher SIC success than DDPG and vanilla REDQ, indicating suitability for practical RIS-aided NOMA under realistic hardware and feedback constraints.