How challenging is the data set? Farinaz Koushanfar ECE Department Rice University Abstract
In sensor networks and other
large-scale networking scenarios, the problems are often complex. Aside
from the combinatorial hardness of the problems, the underlying
variables and parameters are intrinsically uncertain. To address the
same problem, a multitude of statistical models and algorithms have
been designed that work well on one data set, but often cannot be
readily incorporated nor compared with others. Even when a data set is
publicly available, the degree of challenge in addressing the data is
often not clear. There is a need for known sets of benchmark instances,
datasets and comparison metrics to be used for evaluations.
In this talk, I describe our ongoing work in generation of challenging benchmark instances for a popular complex sensor network problem: ad-hoc location discovery. The problem addresses determining the spatial coordinates of the distributed nodes, given noisy and maybe inconsistent (outlier) measurements of the inter-node distances and, a small number of nodes with known locations. Our goal is to generate benchmark instances that contain a spectrum of computationally challenging location discovery input data with controlled parameters such as size, uncertainty, topology, and combinatorial hardness. The benchmark generation approach utilizes a combination of real-world data and its distribution, experiment organization, resampling, instance complexity, feasibility, and sensitivity of the location discovery to the uncertain variables. |