We investigate the joint transmit beamforming and reconfigurable intelligent surface (RIS) configuration problem to maximize the sum downlink rate of a RIS-aided cellular multiuser multiple input single output (MU-MISO) system under imperfect channel state information (CSI) and hardware impairments by considering a practical phase-dependent RIS amplitude model. To this end, we present a novel deep reinforcement learning (DRL) framework and compare its performance against a vanilla DRL agent under two scenarios: the golden standard where the base station (BS) knows the channel and the phasedependentRIS amplitude model perfectly, and the mismatch scenario where the BS has imperfect CSI and assumes idealRIS reflections. Our numerical results show that the introduced framework substantially outperforms the vanilla DRL agent under mismatch and approaches the golden standard.