We consider the in-network computation of approximate “big picture” summaries in bandwidth-constrained sensor networks. First we review early work on computing the Haar wavelet decomposition as a User-Defined Aggregate in a sensor query engine. We argue that this technique can be significantly improved by choosing a function-specific network topology. We generalize this discussion to a loose definition of a 2-level optimization problem that maps from a function to what we call a support graph for the function, and from there to an aggregation tree that is chosen from possible subgraphs of the physical network connectivity. This work is frankly quite preliminary: we raise a number of questions but provide relatively few answers. The intent of the paper is to lay groundwork for discussion and further research.
Joseph M. Hellerstein, Wei Wang