Generalized Delay-Adaptive Codebook Design for Asynchronous Grant-Free Multiple Access via Dual-Fusion DDPG

Grant-free (GF) access reduces scheduling over-head while introducing new challenges due to asynchronous arrivals from geographically dispersed users. In this paper, we investigate the codebook design for asynchronous multi-user uplink communication. By explicitly modeling the inconsistency of user sampling under asynchronous transmission, we analytically characterize the inter-user interference resulting from multiple symbols of other users with arbitrary transmission delays. Then, we formulate the optimal codebook design as a max-min problem, and propose a dual-fusion deep deterministic policy gradient (DF-DDPG) algorithm to resolve it across diverse channel delay realizations. The proposed dual-fusion architecture learns the intrinsic mapping between delay realizations and codebook structures. This capability enables the trained model to generalize to arbitrary delay realizations, facilitating low-complexity adaptation without extensive retraining. Simulations demonstrate that the proposed codebook achieves 1.7dB and 1.5dB block error rate (BLER) performance gains over state-of-the-art in additive white Gaussian noise (AWGN) and Rayleigh fading channels, respectively.