"Sensitivity Analysis in Bayesian Networks: Formal Foundations and Practical Applications."

4/11/2003: 10:30am - 12:00pm
11th Floor Large Conference Room

Abstract: I will present in this talk a new theory for sensitivity analysis in Bayesian networks and discuss a number of its applications. The theory is based on an algerbraic understanding of Bayesian networks in terms of mutlivariate polynomials, and provides some key insights into the (in)significance of probabilities used in quantifying a Bayesian network. Specifically, the theory allows one to answer, efficiently, a large set of practical questions including the following: 1. What network parameters (probabilities) do I need to change, and by how much, to ensure that the answers returned by my reasoning system satisfy some constraints? 2. What guarantees can be offered on the robustness of my reasoning system if I am about to change a parameter from its current value of, say, .3 to .4? 3. How do we formally measure the global change in a Bayesian network model which is induced by a local change to one of its parameters? I will disucss many applications of the presented theory, including model debugging, sensor placement, and planning. I will also discuss a recent application to Bayesian Network Classifiers, in which we characterize the impact of network parameters on the classifier induced by a Bayesian network. The talk will involve a demonstration of the SamIam system, a general purpose tool for inference in Bayesian networks, which incorporates the above results. This is joint work with Hei Chan.

About Adnan Darwiche: Dr. Adnan Darwiche is an Associate Professor of Computer Science at UCLA. He received his Ph.D. and M.S. degrees in Computer Science from Stanford University in 1993 and 1989, respectively. Dr. Darwiche's main research interests are in probabilistic and symbolic automated reasoning and their applications, especially to system analysis and diagnostics. Dr. Darwiche is the general chair of the 19th Conference on Uncertainty in Artificial Intelligence (UAI) to take place in August 2003. He was also a program co-chair of the 18th Conference on Uncertainty in Artificial Intelligence (UAI) and a program co-chair of the 11th International Workshop on Principles of Diagnosis. Dr. Darwiche has published more than 60 papers in leading journals and conferences relating to automated reasoning, diagnosis and AI. Prior to joining UCLA, Dr.Darwiche was a senior scientist and manager of the department of Diagnostics and Modeling at Rockwell Science Center, where he developed diagnostic systems for space, avionics and industrial automation systems.

Last updated: Mon Jun 19 17:44:06 2006