Information Gathering in Bayesian Networks with an Application to Petroleum Prospecting

• Activity Number: 642
• Type: Contributed
• Date/Time: Thursday, August 7, 2014 : 10:30 AM to 12:20 PM
• Sponsor: IMS
• Author(s): Marie Lilleborge*+ and Ragnar Hauge and Jo Eidsvik
• Companies: Norwegian Computing Center and Norwegian Computing Center and Norwegian University of Science and Technology
• Keywords: Bayesian Networks ; information ; design of experiments ; petroleum ; entropy

Abstract

We compare information gathering schemes in order to answer what is a natural information measure for Bayesian Networks. The optimal data acquisition design is not obvious for BNs, as the dependency structure may vary dramatically. Our application is prospect selection for petroleum exploration, where a company has a portfolio of possible drilling sites and the data gathering often is carried out during seasonal campaigns. We use information measures to compare possible observation sets, and to learn which variables are most informative about all observable variables combined.

Three measures for binary random variables will be discussed: Shannon Entropy, sum of Variances and overall Prediction Error. The Shannon Entropy is often considered the standard measure of information, the Variance measure links uncertainty and variance, and the Prediction Error measure is tied to decision making rules. For general BNs, we observe that Shannon Entropy has limitations because it does not value probability updates outside the observation set. Theoretical properties of each measure are presented through examples. We study a case for petroleum exploration in the North Sea.