Practical Perspectives on Stand-Alone Passive RIS Operation From GA-Aided Adversarial Bandit Approach

Practical real-time experiments are an essential step in demonstrating the effectiveness of optimizing Reconfigurable Intelligent Surface (RIS)-aided communication, a pivotal technology for future 5G and 6G networks. In this paper, we present a comprehensive approach to RIS optimization using multiple algorithms, including Genetic Algorithms (GA), Deep Reinforcement Learning (DRL), Adversarial Bandit (AB), GA-aided DRL, and a novel GA-aided AB framework. The experiments are conducted on a custom-built 96-element RIS operating at 3.75 GHz, specifically designed to enhance signal strength and adaptability in realistic propagation conditions. Across diverse receiver placements, the proposed GA-aided AB consistently accelerates convergence and optimizes RIS configurations more effectively than the learning and heuristic baselines, while maintaining real-time operation. GA–aided AB achieves a peak gain of 7.1 dB and remains the top performer even at the most challenging receiver positions like wider angles and longer paths in NLoS settings. Compared to DRL and GA benchmarks, the method delivers higher signal improvements without introducing meaningful computational overhead, underscoring its suitability for embedded deployment. These results demonstrate that a training-free hybrid search combining evolutionary proposals with adversarial bandit selection scales to large RIS configuration spaces and improves practical deployability. This study serves as a foundation for future RIS research, emphasizing the practical benefits of integrating optimization algorithms in real-world deployments.