A machine learning approach to predict the k-coverage probability of wireless multihop networks considering boundary and shadowing effects.

Network coverage is a pivotal performance metric of wireless multihop networks (WMNs) determining the quality of service rendered by the network. Earlier, a few studies have analysed the network coverage by incorporating shadowing effects (SEs) and ignoring the influence of boundary effects (BEs). Besides, there is a void in the literature considering BEs plus SEs together. These approaches not only provide overestimation in network coverage, but also requires intensive simulation for their validation before the actual network installation; thus, increasing the computational time significantly. Furthermore, simulation time shoots up with an increase in network parameters like the number of sensor nodes (SNs) and their sensing range. In this study, we tackle this high simulation time problem by proposing a generalised regression neural network (GRNN) based machine learning (ML) approach to predict the k-coverage performance of a WMN placed in a rectangular-shaped region (RSR). To train the GRNN algorithm for two different set-ups, i.e., without and with BEs, we extract six potential features, namely length of RSR, breadth of RSR, sensing range of SNs, number of SNs, standard deviation of SEs σ, and the value required k through simulations. We also evaluate the importance of individual feature utilising regression tree ensemble technique and simultaneously analysed the sensitivity of each feature to predict the k-coverage probability of the network. The proposed approach has a better prediction accuracy of the k-coverage metric for both with and without BEs scenarios (having R = 0.78 and Root Mean Square Error (RMSE) = 0.14 for with BEs scenario, and R = 0.78 and RMSE = 0.15 for without BEs scenario). It can also be observed that the proposed approach achieves a higher accuracy with minimum computational time complexity as compared to other existing benchmark algorithms.