Rate-Splitting MIMO Multiple Access — An Amortized Optimization Approach

Next-generation wireless technologies such as rate-splitting multiple access (RSMA) and massive MIMO are char-acterized by optimization problems too complex to solve in real-time, hence suboptimal heuristics are adopted in practice. As we explore in this paper, machine learning techniques have the potential to upend this paradigm, offering new algorithms for obtaining near-optimal solutions to a desired class of optimization problems. We consider downlink beamforming optimization (sum rate maximization and max-min fairness) for a MISO interference channel employing RSMA and a large antenna array (up to 64 elements). At such a scale, the linear-MMSE beamformers are the state-of-the-art since the optimal beamformers are too complex to compute. We apply the framework of amortized opti-mization to obtain an iterative operator that designs beamformers with much less computation than existing optimization algorithms (weighted-MMSE) and whose complexity scales linearly with the size of the antenna array. Our simulation results show that the learned optimizer is competitive with state-of-the-art numerical optimization in terms of sum rate, but requires a fraction of the computational cost. For the max-min fairness problem, the learned optimizer outperforms all baseline methods.