Deep Learning-Based LIS Reflection Matrix Design: Impact of Optimizer Selection

Large Intelligent Surfaces (LIS) are emerging as a key technology in future wireless communication systems. This work investigates a practical LIS architecture where a small subset of elements is activated and linked to the baseband for sensing purposes. In contrast, most elements remain passive. A multilayer perceptron (MLP), a deep learning-based framework trained on the DeepMIMO dataset, is employed to estimate the optimal reflection beamforming vector from partial channel state information (CSI) obtained through the active elements. To analyze the impact of different optimization strategies, the performance of six optimizers—Lion, Adam, AdamW, NAdam, SGDM, and RMSprop—is examined. The overall system performance is evaluated based on the achievable rate, outage probability, and training time. The findings show that while the Lion optimizer delivers performance that nearly matches an ideal genie-aided system, it requires a longer training period. Conversely, RMSprop requires the least training time but yields the poorest results. Consequently, selecting an appropriate optimizer is crucial for enhancing system performance in wireless networks supported by LIS.