BARD: Bayesian-assisted resource discovery in sensor networks

Abstract

Data dissemination in sensor networks requires four components: resource discovery, route establishment, packet forwarding, and route maintenance. Resource discovery can be the most costly aspect if meta-data does not exist to guide the search. Geographic routing can minimize search cost when resources are defined by location, and hash-based techniques like data-centric storage can make searching more efficient, subject to increased storage cost. In general, however, flooding is required to locate all resources matching a specification. In this paper, we propose BARD, Bayesian-assisted resource discovery, an approach that optimizes resource discovery in sensor networks by modeling search and routing as a stochastic process. BARD exploits the attribute structure of diffusion and prior routing history to avoid flooding for similar queries. BARD models attributes as random variables and finds routes to arbitrary value sets via Bayesian estimation. Results of occasional flooded queries establish a baseline probability distribution, which is used to focus additional queries. Since this process is probabilistic and approximate, even partial matches from prior searches can still reduce the scope of search. We evaluate the benefits of BARD by extending directed diffusion and examining control overhead with and without our Bayesian filter. These simulations demonstrate a 28% to 73% reduction in control traffic, depending on the number and locations of sources and sinks.