Abstract
The search of optimal Bayesian network from a database of observations is NP-hard. Nevertheless, several heuristic search strategies have been found to be effective. We present a new population-based algorithm to learn the structure of Bayesian networks without assuming any ordering of nodes and allowing for the presence of both discrete and continuous random variables. Numerical performances of our mixed-genetic algorithm, (M-GA), are investigated on a case study taken from the literature and compared with greedy search
