Adaptive Learning-Enhanced Cross-Entropy Algorithm for RIS-Aided Communications: Blind Beamforming Design and Experimental Validation

Widely adopted reconfigurable intelligent surface (RIS) beamforming schemes generally rely on estimating accurate channel state information (CSI) exploiting convex optimization algorithms, leading to high estimation overhead, computational complexity, estimation errors, and poor compatibility with existing protocols. Fortunately, blind beamforming schemes eliminate the need of CSI estimation and convex optimization. However, conventional blind beamforming schemes (e.g., cross-entropy (CE) scheme) search for a satisfactory solution in extremely large RIS configuration space, which may suffer from slow convergence and falling into the local optimum. To address these issues, we propose an adaptive learning-enhanced cross-entropy (ALECE) blind beamforming algorithm. Specifically, ALECE adaptively determines the number of elite samples for updating the probability transition matrix, thereby refining the search direction, reducing the solution space, and accelerating convergence. Besides, to avoid local optimum and loss of promising solutions, we incorporate a learning factor conditional and sample mean (CSM) to facilitate exploration and preserve high-quality samples. Furthermore, we provide convergence proof and formal complexity analysis. Finally, both simulation and experiment results demonstrate that, compared with the conventional CE method, the proposed ALECE algorithm can effectively trade off convergence speed while incurring only a marginal performance loss through an appropriate choice of learning factor.