Reinforcement learning-based beamforming for massive multiple-input multiple-output systems: A robust and energy-efficient approach