Abstract
We investigate some problems related to implementation of uncertainty management, in particular the handling of computational and conceptual difficulties that easily appear in complex problems. The uncertainty polytope resulting from a set of inequality judgments on probabilities and means in a problem has very high dimension, but can be represented by a projection on a low-dimensional space if the judgments are structured into a graph with low tree-width. With this representation many judgments of independence become vacuous. The uncertainty polytope is high-dimensional and thus difficult to grasp or visualize. We propose a method to sample uniformly and efficiently from the polytope, as a means to obtain various summaries not obtainable by linear programming, such as volume, center of gravity, principal axes, etc.
Keywords. Learning, uncertainty, decomposition, uncertainty polytope
The paper is available in the following formats:
Authors addresses:KTH
E-mail addresses:
| Stefan Arnborg | stefan@nada.kth.se |
Related Web Sites