Reconfigurable Intelligent Surface

Unlike active relays, RIS elements are passive or semi-passive, requiring no power amplifiers. Each unit cell contains a tunable impedance—typically a PIN diode or varactor—controlled by a DC bias voltage. A neural network computes the optimal phase shift matrix Φ to maximize the signal-to-noise ratio (SNR) at the receiver by constructively combining reflected paths.

• No RF chains: Eliminates expensive and power-hungry analog front-ends
• Phase resolution: Typically 1-bit to 3-bit quantization per element
• Control plane: FPGA or microcontroller applies neural network outputs via a serial bus

The core computational challenge is solving the non-convex optimization of the reflection coefficient matrix. Deep reinforcement learning (DRL) agents, such as Deep Deterministic Policy Gradient (DDPG), learn to map channel state information (CSI) directly to optimal phase configurations without explicit channel estimation.

• Input: Pilot signals or compressed CSI feedback
• Output: Vector of phase shifts for N elements
• Training: Reward function based on achievable rate or received signal strength
• Alternative: Graph Neural Networks (GNNs) model the RIS as a graph for scalable, distributed control

Estimating the cascaded transmitter-RIS-receiver channel is challenging because RIS elements lack active RF chains. Neural networks address this through compressed sensing and deep unfolding architectures. The Learned AMP (LAMP) network unrolls the Approximate Message Passing algorithm to estimate the high-dimensional cascaded channel from limited pilot measurements.

• Cascaded channel: Product of Tx-RIS and RIS-Rx channels
• Pilot overhead: Reduced by exploiting channel sparsity in the angular domain
• Twin RIS architectures: One RIS for sensing, one for reflection, enabling parallel channel acquisition

RIS enables holographic MIMO—a spatially continuous aperture approaching the theoretical limits of electromagnetic information theory. By densely packing sub-wavelength elements, the RIS approximates a continuous surface current distribution. Neural networks optimize the holographic beamforming weights to generate highly focused, diffraction-limited beams.

• Element spacing: λ/5 or less for holographic operation
• Fourier optics: Neural networks learn the mapping from desired far-field pattern to aperture field distribution
• Application: Terahertz (THz) communications where path loss is extreme and pencil beams are essential

In dense urban deployments, multiple RIS panels must be jointly optimized to avoid destructive interference and maximize coverage. Multi-agent reinforcement learning (MARL) frameworks, such as QMIX or MADDPG, allow distributed RIS controllers to learn cooperative policies through a centralized training with decentralized execution (CTDE) paradigm.

• Challenge: Non-stationary environment as each RIS adapts
• Communication: Agents share latent representations, not raw CSI
• Handover: Neural networks predict optimal RIS association as users move through the environment

RIS can be configured to perform over-the-air computation (AirComp) by exploiting wave superposition. The reflection coefficients are set so that the received signal directly represents a mathematical function—such as a weighted sum—of distributed sensor transmissions. Neural networks learn the optimal phase profile to minimize computation error while suppressing noise.

• Federated learning: RIS accelerates model aggregation by computing the weighted average of local gradients in the analog domain
• Function: Nomographic functions like arithmetic mean, geometric mean, or maximum
• Error metric: Mean squared error (MSE) between desired and received function value