Joint RIS Optimization and Channel Estimation With Decision Tree-Based Adaptive Reconfiguration

Reconfigurable intelligent surfaces (RISs) are seen as a promising technology to improve cellular network coverage, due to their ability to steer the impinging signals in desired directions. The design of the RIS can be easily addressed by assuming full channel knowledge. Nevertheless, estimating the channels to and from the RIS is a challenging problem, as it requires a huge training overhead. This paper proposes an efficient configuration optimization jointly with channel estimation by exploiting deep learning tools. In particular, we propose an algorithm that works in two steps. The first step is based on a decision tree that requires few end-to-end channel estimates with different RIS configurations. The configurations are iteratively selected based on an estimate of the mutual information between the obtained rates and the optimal configuration. The second step instead provides the minimum mean-square-error estimate of the optimal RIS configuration based on the data rates estimated on the channels obtained in the first step through a neural network (NN) trained with a supervised approach. Numerical results confirm that the proposed solution provides a configuration close to the optimal, with achievable rates approaching the upper bound obtained with perfect channel knowledge.