Linear-space best-first search
Artificial Intelligence
Data Mining and Knowledge Discovery
A Guide to the Literature on Learning Probabilistic Networks from Data
IEEE Transactions on Knowledge and Data Engineering
A study of causal discovery with weak links and small samples
IJCAI'97 Proceedings of the Fifteenth international joint conference on Artifical intelligence - Volume 2
The Bayesian structural EM algorithm
UAI'98 Proceedings of the Fourteenth conference on Uncertainty in artificial intelligence
Graphical models and exponential families
UAI'98 Proceedings of the Fourteenth conference on Uncertainty in artificial intelligence
Learning mixtures of DAG models
UAI'98 Proceedings of the Fourteenth conference on Uncertainty in artificial intelligence
Learning equivalence classes of Bayesian network structures
UAI'96 Proceedings of the Twelfth international conference on Uncertainty in artificial intelligence
Hi-index | 0.00 |
This article provides an overview of how to handle uncertainty about which Bayesian network to use for calculating the effect of an ideal manipulation or a classification. The Bayesian approach to handling uncertainty is to put a prior distribution over all of the Bayesian networks and their parameters, and then use this to calculate a posterior distribution over the quantity of interest. This is in general computationally infeasible, due to the huge number of different Bayesian networks over a given set of variables. Other approaches approximate the Bayesian answer using Monte Carlo Markov chain algorithms, or Bayesian model averaging, where all Bayesian networks except for a few good Bayesian networks are ignored in order to simplify the calculations. The latter approach requires searching among the vast space of Bayesian networks for the good Bayesian networks. Several methods for scoring Bayesian networks, and several search algorithms are described. It is shown how the problems of equivalent models and latent variables complicate both searching and scoring. Finally, it is shown how searching over equivalence classes of Bayesian networks, instead of searching over Bayesian networks can simplify both scoring and search.