Neural Network Methods for Bayesian Networks

SPONSORSHIP: 

NSF/IRA

BRIEF DESCRIPTION

Using Bayesian networks (AKA belief networks) to solve problems involving causal and probabilistic inferences has been an active research area in AI, but existing methods are limited because of their complexities and the difficulty in obtaining the needed causal knowledge. This project, sponsored by NSF, is aimed at overcoming these problems by developing neural network methods which can be directly applied to belief networks. Such integration would retain the representational power of belief networks and at the same time take advantage of computational and learning capabilities of neural networks. Specifically, we are concentrating on two research topics here. The first one is to develop a neural network learning method for constructing and dynamically updating belief networks (both the network structures and the probability distributions) from case data. The second one is to adopt neural network optimization techniques to certain inference tasks in belief networks.

RELEVANT PAPERS

FOR MORE INFORMATION

Contact Yun Peng, ypeng@umbc.edu .