Improving Physical Layer Computation with Reconfigurable Intelligent Surfaces

In this paper, we maximize the computational rate in the physical layer of a multiuser uplink setup using the compute-and-forward method in an environment equipped with Reconfigurable Intelligent Surfaces (RIS). This is achieved by interacting the elements of the RIS with the environment to find the optimal phase angles which best align the effective uplink channel gains with the required linear combination coefficients. For this purpose, the Deep Deterministic Policy Gradient (DDPG) algorithm, a member of Deep Reinforcement Learning (DRL) cat-egory, is used. Our proposed method improves the performance of the required linear computation by carefully adjusting the phases of RIS elements. To compare the results of the DDPG-based method with previous work in a similar scenario, the Alternating Optimization (AO) approach is also examined. In a typical uplink scenario, it is shown that the performance is improved by 33, 20 and 10 percent (on average) for a setup with 2, 5 and 10 users, respectively. In addition to the above-mentioned performance improvement, another advantage of the proposed DDPG-based algorithm is that it does not require Channel State Information (CSI) being fed back to RIS controller, whereas the alternating optimization method requires.