Learning the Node Distribution of Biased Waypoint Mobility using Mixture Density Networks

In mobile ad hoc networks, the knowledge of the node distribution can facilitate improvements in applications like online facility deployment and offline simulation acceleration. Considering that exact formulation of such spatial density functions is mostly intractable, we resort to mixture density networks to learn the node distribution in a closed form, assuming the node follows a mobility model that allows biased waypoints. Unlike most mixture models whose kernels are the Gaussian distribution, we propose to adopt and mix two new types of distribution, namely the multivariate generalized beta distribution and the bivariate GB1 (generalized beta of the first kind) distribution, to suit our application for modeling hotspots over a square area. The results demonstrate that our proposed mixture models using the new kernels can well approximate the target node distribution in shape, surpassing the traditional Gaussian mixture model in terms of approximation accuracy and also the cost of parameters to learn.