This work introduces a movable‑element (ME) multi‑functional reconfigurable intelligent surface (MF‑RIS) architecture, denoted as MM‑RIS. Beyond conventional phase‑shifting, MM‑RIS endows each element with mobility and supports reflection, transmission, amplification, and energy harvesting (EH), thereby improving communication flexibility and energy self‑sustainability.
On top of MM‑RIS, a hybrid rate‑splitting multiple access (RSMA) and non‑orthogonal multiple access (NOMA) transmission scheme is designed to boost spectral efficiency. An energy efficiency (EE) maximization problem is formulated, jointly optimizing transmit beamforming, power and rate allocation, MM‑RIS configurations, panel deployment, and ME positions.
The resulting problem is a non‑convex mixed continuous‑discrete optimization. To tackle it, we propose the constrained hybrid deep reinforcement learning framework CRAFTS (Constrained Hybrid deep reinforcement learning with federated transformer and parametrized sharing). CRAFTS integrates multi‑agent proximal policy optimization (MAPPO) for continuous actions and multi‑agent deep Q‑networks (MADQN) for discrete actions via bi‑directional parameter sharing, enabling coordinated continuous‑discrete decision making.
Inside the framework, a transformer‑based encoder enriches state representation, while federated learning (FL) facilitates distributed knowledge sharing across agents. Constraint satisfaction is enforced by a Lagrangian‑based adaptive penalty mechanism that dynamically adjusts penalty weights during training.
Simulation results demonstrate that the MM‑RIS architecture combined with CRAFTS markedly outperforms existing benchmarks in both EE and constraint satisfaction, confirming the synergistic benefits of spatially reconfigurable MM‑RIS, hybrid RSMA‑NOMA transmission, and intelligent collaborative learning.
Review