Abstract
This paper investigates artificial intelligence (AI) empowered schemes for reconfigurable intelligent surface (RIS) assisted networks from the perspective of fast implementation. We formulate a weighted sum-rate maximization problem for a multi-RIS-assisted network. To avoid huge channel estimation overhead due to activating all RISs, we propose a computer vision (CV) enabled RIS selection scheme based on a single shot multi-box detector. To realize real-time resource allocation, a deep neural network (DNN) enabled transmit design is developed to learn the optimal mapping from channel information to transmit beamformers and phase shift matrix. Numerical results illustrate that the CV module is able to select the RIS with the best propagation condition. The well-trained DNN achieves similar sum-rate performance to the existing alternative optimization method but with much smaller inference time.
Introduction
Reconfigurable Intelligent Surfaces (RIS) are a cornerstone of B5G/6G wireless technology, utilizing passive elements to steer radio signals and bypass obstacles. However, two major bottlenecks prevent their large-scale deployment: massive channel estimation overhead and high computational complexity. Traditional iterative algorithms, such as Alternative Optimization (AO), are far too slow to keep up with mobile users in real-time. This paper proposes a two-stage AI framework that uses Computer Vision (CV) to "see" the environment and Deep Neural Networks (DNN) to "learn" the optimization mapping, enabling near-optimal data rates with microsecond latency.
CV-Enabled RIS Selection
To minimize channel estimation overhead, the system employs a CV module to identify the optimal RIS for a given environment. Instead of activating all RISs and measuring their channels, a top-view camera captures the scene to identify users and obstacles.
As shown in the system model, the CV module utilizes a Single Shot Multi-box Detector (SSD) with a VGG16 backbone to detect entities. Following detection, the Canny edge detection algorithm calculates precise coordinates to determine if a line-of-sight (LoS) exists or if a signal must penetrate an obstacle (NLoS). By calculating the effective pathloss for each potential RIS, the system activates only the one with the best propagation conditions, drastically cutting down on channel estimation needs.
DNN-Enabled Transmission Design
Once an RIS is selected, the system must configure its phase shifts and the transmitter's beamforming. The authors propose a fully connected DNN that takes channel state information (CSI) as input and predicts power allocation and phase shift matrices.
Crucially, the DNN is trained via unsupervised learning, using the negative weighted sum-rate as the loss function. This allows the network to maximize throughput directly without requiring a slow "teacher" algorithm to provide labels for the 100,000 training samples. This approach trades a small sliver of mathematical optimality for a massive increase in execution speed.
Performance Evaluation
The proposed framework was tested in a scenario with 6 RISs, 8 transmit antennas, and 4 mobile users. The CV module successfully identified RIS units with the least penetration loss and operated at 11 Frames Per Second (FPS), sufficient for tracking users with moderate mobility.
As summarized in the performance table, the DNN achieved a sum-rate of 34.44 bit/s/Hz, which is within 10% of the performance of the heavy-duty AO algorithm. Most importantly, the DNN inference time was approximately 0.02 milliseconds, representing a million-fold increase in speed compared to the AO algorithm's 70-119 seconds. This enables true real-time resource allocation in dynamic wireless environments.
Conclusion
This AI-empowered framework demonstrates that computer vision and deep learning can effectively overcome the implementation barriers of RIS technology. By using vision as a "zero-overhead" sensor and DNNs for instantaneous optimization, the system achieves a balance between high data rates and real-time responsiveness. Future work will explore the generalization of these models to more complex urban environments and higher mobility scenarios.
Summary updated