Emerging 6G wireless systems demand ever higher spectral efficiency. Reconfigurable Intelligent Surfaces (RIS) large arrays of passive reflecting elements can reconfigure the radio environment via programmable phase shifts. This work presents a deep reinforcement learning (DRL) approach to jointly optimize the multi-antenna base station beamforming and the RIS phase shifts in a down- link 5G MISO network. Using a DDPG-based agent, we maximize the sum rate (bps/Hz) by treating it as the DRL reward. Our simulations (based on Saglam’s RIS-MISO code demonstrate sum rates exceeding 28 bps/Hz in large configura- tions for example, over 30 bps/Hz at 30 dB transmit power with 32 BS antennas and 32 RIS elements. We analyze how key factors (BS/RIS size, power, learn- ing hyperparameters) affect performance through visualization. The DRL agent learns effective continuous beamforming and phase control without explicit chan- nel models. Results confirm that more antennas or RIS elements significantly boost throughput, and that careful tuning of DRL parameters (learning rate, exploration decay) is crucial. We include the Saglam repository as a reference to aid reproducibility.
Hole et al. (Sun,) studied this question.