1: \begin{abstract}

2: % In this article, we present a novel \ac{DRL} framework for the problem of joint downlink transmit beamforming and passive \ac{RIS} configuration to maximize the sum rate under imperfect \ac{CSI} and phase-dependent \ac{RIS} reflection amplitude. We consider two settings for learning agents: the golden standard with perfect \ac{CSI} and full knowledge of the phase-dependent amplitude model and a DRL agent under imperfect \ac{CSI} and ideal reflection assumption. Adhering to imperfect \ac{CSI} and unknown phase-dependent amplitude model, the introduced method is compared against the considered cases. Our empirical studies indicate that the introduced approach significantly outperforms the \ac{DRL} agent that assumes ideal reflection and closely follows the golden standard by a negligible margin in terms of the convergence rate and achieved sum rates. 

3: We investigate the joint transmit beamforming and \ac{RIS} configuration problem to maximize the sum downlink rate of a \ac{RIS}-aided cellular \ac{MU-MISO} system under imperfect \ac{CSI} and hardware impairments by considering a practical phase-dependent \ac{RIS} amplitude model. To this end, we present a novel \ac{DRL} framework and compare its performance against a vanilla \ac{DRL} agent under two scenarios: the golden standard where the \ac{BS} knows the channel and the phase-dependent \ac{RIS} amplitude model perfectly, and the mismatch scenario where the \ac{BS} has imperfect \ac{CSI} and assumes ideal \ac{RIS} reflections. Our numerical results show that the introduced framework substantially outperforms the vanilla \ac{DRL} agent under mismatch and approaches the golden standard.

4: \end{abstract}

5: