Information-driven cooperative sampling strategies for spatial estimation by robotic sensor networks.

摘要

Networks of environmental sensors play an increasingly important role in scientific studies of the ocean, rivers, and the atmosphere. Robotic sensors can improve the efficiency of data collection, adapt to changes in the environment, and provide a robust response to individual failures. Ideally, online path planning algorithms should be statistically aware, driving the sensors towards those sampling locations which will provide the most information. At the same time, such algorithms need to be distributed and scalable to make robotic networks capable of operating in an autonomous and robust fashion. The combination of complex statistical modeling and distributed coordination presents difficult technical challenges: traditional statistical modeling and inference assume full availability of all measurements and central computation. While collecting sample values at a central location is certainly a desirable property, the paradigm for distributed motion coordination builds on partial, fragmented data. We present two alternative approaches to the problem of distributed optimal sampling design.;First, under a restricted class of spatio-temporal model, we consider the asymptotic regime of near independence between distinct sample locations. This study formally justifies the intuitive notion of space filling designs, thus transforming the statistical design problem into a geometric one. We provide distributed algorithms for optimal sampling under these conditions.;Second, for a more general family of Bayesian spatio-temporal models, we consider a gradient approach to sequential optimal design. We consider two well known optimality criteria: maximizing predictive entropy over potential sample locations, and minimizing average posterior variance over a predictive region. We introduce a hybrid network of static computation and control nodes and dynamic sensing agents, and we develop approximations of these two objective functions which may be calculated in a cooperative way by this network. We detail a distributed gradient-based algorithm for obtaining local optima of the approximate objective function in a sequential setting.