Abstract
Reconfigurable intelligent surfaces (RISs) enable programmable wireless propagation, but optimizing discrete RIS phases remains computationally demanding for large multiple-input multiple-output (MIMO) systems. This paper proposes the Surrogate-Assisted Rank-1 Genetic Algorithm (SARGA), a training-free framework that combines Frobenius-norm-based candidate screening, selective exact spectral-efficiency (SE) evaluations, and cached rank-1 channel updates. The method optimizes directly over the discrete phase alphabet without offline training or continuous-phase quantization. SARGA is evaluated using indoor and outdoor SimRIS channels, direct-link-blocked conditions, and comparisons with classical coordinate-ascent optimization (AO), an adaptive genetic algorithm (AGA), the Walrus Optimization Algorithm (WaOA), and adapted deep reinforcement learning (DRL) and graph neural network (GNN) baselines. For a 256-element, two-bit RIS, SARGA achieves mean SE values of 123.93 ± 48.52 and 110.99 ± 68.80 bps/Hz indoors and outdoors, respectively. Under 440 exact evaluations, its indoor performance is statistically indistinguishable from AGA (p = 0.77) and exceeds the tested WaOA adaptation (p < 0.001). Relative to locally optimized continuous-phase AO, SARGA achieves mean channel-wise SE ratios of 96.13% indoors and 95.81% outdoors. The method also exhibits limited SE degradation under the evaluated CSI perturbations and phase resolutions. The results support SARGA as a competitive training-free approach that reduces exact-evaluation requirements while retaining strong SE under the tested conditions. Its computational benefits depend on the comparison budget and implementation.