Abstract
This paper investigates SCMA-enabled uplink transmission with hybrid automatic repeat request (HARQ) and adaptive modulation and coding (AMC), referred to as SCMA-HARQ, for Low Earth Orbit (LEO) satellite communications. We first formulate the expected throughput of the proposed SCMA-HARQ by decoupling the multiuser interference among SCMA users. The resulting problem is formulated to maximize the expected system throughput, leading to a mixed discrete-continuous optimization with per-user throughput and reliability constraints. To address the inherent nonconvexity and the large interference-coupled decision space, we propose a federated hybrid soft actor-critic (FH-SAC) framework, in which each TU learns a decentralized policy for joint modulation and coding scheme (MCS) selection and transmit power control, while periodic parameter aggregation enhances learning stability and scalability. By incorporating HARQ-aware throughput modeling into the soft actor-critic learning process, the proposed approach explicitly accounts for retransmission effects in the control loop. Simulation results demonstrate the superiority of the proposed FH-SAC scheme in SCMA-HARQ systems in terms of convergence speed, robustness with respect to the number of supported users, and overall throughput improvement.