Reconfigurable intelligent surface (RIS) is considered as one of the key enabling technologies for 6G. RIS is a meta-surface consisting of a large number of low-cost passive elements with configurable phases to "reflect" wireless signals to empower smart and reconfigurable radio environments. We will first discuss RIS-assisted wireless communication and its various research aspects. Then, we will introduce our work on joint base station, relay, and RIS beamforming in hybrid relay-RIS multiuser systems. We will introduce both model-based and learning-based solutions. The learning-based solution, especially a graph neural network (GNN)-based method, demonstrates superior sum-rate performance, robustness against channel imperfections and variations, and three-order-of-magnitude complexity reduction, compared to model-based approaches.