Edge-Native AI-Enabled Spatial Decision Region Formation for Optical Attocell-Based Indoor VLC

This research study investigates the impact of machine learning (ML) hyperparameters and the training dataset on optical attocell formation in indoor 6G visible light communication (VLC) networks that leverage edge-native artificial intelligence (AI). The work employs a probabilistic neural network (PNN) to form the optical attocell given the user's location, i.e., classifying the VLC transmitter based on its proximity to the user. The numerical simulations show that the distribution and density of the training dataset directly influence the spatial resolution and convergence of classification boundaries, while the variance parameter of the PNN pattern layer, σ², governs the smoothness, stability, and generalization of the resulting decision regions. In the presence of a spatially ambiguous training sample, the higher value of σ² destabilizes the decision boundaries, resulting in the fusion of attocell regions; conversely, in the case of the limited dataset density with sparse spatial samples, the increase in σ² results in stabilized and smooth decision boundaries for attocell formation. The results indicate that the ML-driven optical attocell formation is not fixed by the VLC system's geometrical model, but dynamically adapts to the learning parameters and data availability, hence providing a novel design dimension in practical edge-native AI-based VLC deployments.