Learning structural properties of wireless ad-hoc networks non-parametrically from spectral activity samples

In this paper we apply a Bayesian non-parametric model to segmenting time series of observed wireless node transmission activity in order to learn routing patterns in an unknown ad-hoc network, as well as its topology. This emulates cognition of a spectrum sensing radio network capable of geolocating transmitting nodes of another network and detecting starting and stopping times of their packet transmissions with various degree of accuracy. Each wireless node of the monitored network is described by a Hidden Semi-Markov model (HSMM), where the states, state durations and emissions, and transition probabilities between states are uncovered based solely on RF spectrum observations. State durations are random variables that correspond to node's activity segments. The learning of the HSMM model is non-parametric and derived from the Hierarchical Dirichlet Process (HDP) prior. We demonstrate the effectiveness of this approach using an NS-3 simulated 802.11 wireless network whose nodes are placed on a grid such that one node's grid neighbors are the only nodes in its range. In this network relaying with routing is necessary to reach other network nodes. The quality of learning for the proposed approach is analyzed based on simulated and synthetic data.