AND/OR Importance Sampling

The paper introduces an AND/OR importance sampling scheme for probabilistic graphical models. In contrast to conventional importance sampling, AND/OR importance sampling caches samples in the AND/OR space and then extracts a new sample mean from the stored samples. We prove that the AND/OR sample mean may have lower variance than conventional importance sampling; thereby providing a theoretical justification for preferring it over conventional importance sampling. Our empirical evaluation confirms that AND/OR importance sampling is far more accurate than conventional importance sampling in many cases.

• Type: Conference
• Year: 2008
• Where: UAI
• Authors: Vibhav Gogate, Rina Dechter
• Keywords: AND/OR Importance, Artificial Intelligence, Conventional Importance, Importance Sampling, UAI 2008