Abstract: This article presents a smart approach for optimized Access Point (AP) placement in 3D buildings through the use of Deep Learning (DL), specifically a U-Net architecture. The continuous growth of the Internet of Things (IoT) has increased the demand for dense AP deployments in both outdoor and indoor environments. The conventional approach, which combines ray tracing with standard optimization algorithms, is computationally expensive and impractical for large-scale environments. To address this issue, a modified U-Net is designed and trained to rapidly and effectively place APs within a 3D building map while maintaining a minimal coverage error. In detail, the Neural Network (NN) is trained to process building scenarios with up to five floors and a maximum of five APs. The considered performance metrics were the percentage coverage difference ( ΔCov (%)) and the computational time ratio (Rt) compared to a generic Simulated Annealing (SA) algorithm. For the evaluated scenarios, ΔCov (%) ranges from -14.6% to -1.7%, while Rt highlights a significant computational gain, with the NN achieving up to a 700-fold speed-up over the SA optimizer. Thus, the proposed U-Net approach ensures only a limited degradation in coverage performance while providing a significantly faster response compared to the SA optimizer.
