A Bayesian Inference Model for Dynamic Neighbor Discovery in Tactical Networks

This paper introduces a Bayesian Inference model for dynamically changing the neighbor discovery parameters in order to minimize the time to recover from long radio link disconnections in tactical networks. We start with the hypothesis that given a multi-layer tactical system, we can use the in/out chains to learn the distribution of changes in the radio data link with an increasing confidence as time goes by. Moreover, we propose that a static neighbor discovery configuration (e.g., hello time, aggregation time, and so on) in dynamic networks can never be optimal and that it is possible to calculate the close to optimal neighbor discovery configuration dynamically. Thus, we introduce a Bayesian Inference model composed of a Markov-Chain-Monte-Carlo (MCMC) and a Long Short-Term Memory (LSTM) to compute a dynamic configuration by learning the probability distribution of the link conditions (computed using the in/out model) and adapting the configuration parameters dynamically according to this distribution. Our hypothesis was verified with numerical calculations using different patterns of link data rate change. The quantitative analysis suggests that at the tactical edge dynamic control signaling can improve connectivity in ever-changing network scenarios.