A Genetic Algorithm for Tuning Variable Orderings in Bayesian Network Structure Learning

Learning a Bayesian network from data is NP-hard even without considering unobserved or irrelevant variables. Many previous Bayesian network learning algorithms require that a node ordering is available before learning. Unfortunately, this is usually not the case in many real-world applications. To make greedy search usable when node orderings are unknown, we have developed a permutation genetic algorithm (GA) wrapper to tune the variable ordering given as input to K2, a score-based BN learning algorithm. We have used a probabilistic inference criterion as the GA’s fitness function and we are also trying some other criterion to evaluate the learning result such as the learning fixed-point property.

Authors

• Haipeng Guo Harbin Institute of Technology
• Julie A. Stilson and William H. Hsu

Keywords

No keywords are indexed for this paper.