Abstract
The problem of learning the structure of
Bayesian networks from complete discrete
data with a limit on parent set size is consid-
ered. Learning is cast explicitly as an optimi-
sation problem where the goal is to find a BN
structure which maximises log marginal like-
lihood (BDe score). Integer programming,
specifically the SCIP framework, is used to
solve this optimisation problem. Acyclic-
ity constraints are added to the integer pro-
gram (IP) during solving in the form of cut-
ting planes. Finding good cutting planes is
the key to the success of the approach—the
search for such cutting planes is effected using
a sub-IP. Results show that this is a particu-
larly fast method for exact BN learning.
Bayesian networks from complete discrete
data with a limit on parent set size is consid-
ered. Learning is cast explicitly as an optimi-
sation problem where the goal is to find a BN
structure which maximises log marginal like-
lihood (BDe score). Integer programming,
specifically the SCIP framework, is used to
solve this optimisation problem. Acyclic-
ity constraints are added to the integer pro-
gram (IP) during solving in the form of cut-
ting planes. Finding good cutting planes is
the key to the success of the approach—the
search for such cutting planes is effected using
a sub-IP. Results show that this is a particu-
larly fast method for exact BN learning.
| Original language | English |
|---|---|
| Title of host publication | Proceedings of the 27th Conference on Uncertainty in Artificial Intelligence (UAI 2011) |
| Pages | 153-160 |
| Publication status | Published - 14 Jul 2011 |