Bayesian Network (BN), is a probabilistic graphical model that represents a set of random variables, and their conditional dependencies via a directed acyclic graph (DAG). To exploit the power of BN in knowledge representation and inference, the BN structures (BNs) have to be constructed for the given domain. This is done by learning both the structure and joint probability distribution parameters from domain data.
Learning BNs is a NP hard problem with important practical applications. Difficult challenges remain in overcoming the computational complexity and reducing computation time on domain datasets of medium to large size. In this research, we have developed two novel approaches (ChainACO and K2ACO) based on chain structure model and K2 greedy search, for fast BNs learning use Ant Colony Optimization (ACO). Both approaches belong to the class of search and score algorithms, operating on the space of orderings of data variables, using Chain score and K2 score as scoring functions.
We have demonstrated that using the chain score can sometimes reduce computational complexity of evaluation while maintaining reasonable solution quality. Experimental results on a range of benchmarks show problem-specific trade-offs between solution quality and computational effort.  The main focus of our research is to relate differences in performance on standard benchmarks to the interplay between algorithm design, scoring function and known structural properties of the underlying Bayesian Network. Work to date has examined how the distribution of search varies with algorithm design and scoring method and we are currently conducting landscape analysis on a range of benchmarks