A layered abductive inference framework for diagramming group motions


Many events in different domains are characterized by a large number of individual moving elements, either in pursuit of a goal in groups (as in military operations), or subject to underlying physical forces that group elements with similar motion (as in weather phenomena). Visualizing and reasoning about happenings in such domains are often facilitated by abstracting the mass of spatiotemporal data into spatial diagrams of group motions, and overlaying them on abstractions of static features, like maps. The standard approach has been to use a clustering algorithm to group the entities and extract their motions, which often produces unsatisfactory results when the data is noisy and incomplete. In this paper, we present for the task a multilayered abductive inference framework where hypotheses largely flow upwards from raw data to a diagram, but there is also a top-down control that asks lower levels to supply alternatives if the higher level hypotheses are not deemed sufficiently coherent. This top-down-bottom-up interplay to combine numerical clustering algorithms with symbolic knowledge about consistency models the flexibility of human reasoning in such tasks that enables the framework to successfully deal with extremely noisy and incomplete data. We present experimental results as obtained by deploying the proposed framework and discuss several issues related to construction of such diagrams, from data sets over a few days of a large number of military units engaged in exercises at the National Technical Center, California, USA. © 2006 Oxford University Press.

Publication Title

Logic Journal of the IGPL