1: \begin{abstract}

2: The \gls{ris} is a promising technology for next-generation wireless communication.

3: It comprises many passive antennas,

4: which reflect signals from the transmitter to the receiver with adjusted phases without changing the amplitude.

5: The large number of the antennas enables a huge potential of signal processing despite the simple functionality of a single antenna.

6: However, it also makes the \gls{ris} configuration a high dimensional problem,

7: which might not have a closed-form solution and has a high complexity and, 

8: as a result, 

9: severe difficulty in online real-time application if we apply iterative numerical solutions.

10: In this paper, we introduce a machine learning approach to maximize the \gls{wsr}.

11: We propose a dedicated neural network architecture called \emph{RISNet}.

12: The \gls{ris} optimization is designed according to the \gls{ris} property of product and direct channel and homogeneous \gls{ris} antennas.

13: The architecture is scalable due to the fact that the number of trainable parameters is independent from the number of \gls{ris} antennas

14: (because all antennas share the same parameters).

15: The \gls{wmmse} precoding is applied and an \gls{ao} training procedure is designed.

16: Testing results show that the proposed approach outperforms the state-of-the-art \gls{bcd} algorithm.

17: Moreover, although the training takes several hours,

18: online testing with trained model (application) is almost instant,

19: which makes it feasible for real-time application.

20: Compared to it, the \gls{bcd} algorithm requires much more convergence time.

21: Therefore, the proposed method outperforms the state-of-the-art algorithm in both performance and complexity.

22: \end{abstract}

23: